Your attention please!

The following text was published as a book by Prentice Hall in 1994.

It has since gone out of print, and the copyright was formally reversed
back to the authors Kim Waldén and Jean-Marc Nerson in 2001.

We have decided to make the text publicly available with the following provision.

You may download and copy the text at will, but as soon as some part is extracted
and presented out of context of the complete book, there must always be a clearly
visible reference to the book and its authors, and also a pointer to the website
www.bon-method.com (since the printed book is no longer available) so that any
reader of the partial content can quickly find the original full text. The text of this
provision must also be included in any such presentation.
 Seamless
Object-Oriented
Software
Architecture
Analysis and Design of Reliable Systems

Kim Waldén
Jean-Marc Nerson

Printed version
September 6, 1994
Contents

Series editor’s preface xi

Preface xiii

Part I Introduction 1

1 Object-oriented software development 3

1.1 Introduction 3
1.2 Seamlessness 7
1.3 Reversibility 7
1.4 Software contracting 9

2 The BON approach 11

2.1 Introduction 11
2.2 What is not in BON 12
2.3 Other methods 15
2.4 The BON position 16
2.5 Characteristics of the notation 17
2.6 Static and dynamic models 24

Part II The model 27

3 The static model—classes and clusters 29

3.1 Introduction 29
3.2 BON modeling charts 30
3.3 System state information 35
3.4 Typed class interfaces 36
3.5 Class header 37
3.6 Indexing clause 37

v
vi CONTENTS

3.7 Inheritance clause 39

3.8 Class features 39
3.9 Class invariant 47
3.10 Graphical and textual specification 48
3.11 The BON assertion language 49
3.12 The BON predicate logic 52
3.13 Assertions and partial functions 57
3.14 Scalable system description 59
3.15 Class headers 59
3.16 Clusters 60

4 Static relations 65
4.1 Inheritance relations 65
4.2 Client relations 69
4.3 Class vs. object dependencies 86
4.4 Indirect client dependencies 88
4.5 Semantic links 89

5 The dynamic model 90

5.1 Introduction 90
5.2 Dynamic charts 93
5.3 Dynamic diagrams 99

Part III The method 117

6 Issues of a general method 119

6.1 Introduction 119
6.2 Simulating an ideal process 120
6.3 Enterprise modeling 121
6.4 What is analysis and what is design? 122
6.5 Reuse 125
6.6 User centered design 129
6.7 Roles of analysis objects 135
6.8 Modeling a system interface 136
6.9 System usage 140
6.10 A paradigm shift 142
 CONTENTS vii

7 The BON process 143

7.1 The BON deliverables 143
7.2 Characteristics of the process 148
7.3 Task 1: Delineate system borderline 152
7.4 Task 2: List candidate classes 160
7.5 Task 3: Select classes and group into clusters 162
7.6 Task 4: Define classes 166
7.7 Task 5: Sketch system behaviors 168
7.8 Task 6: Define public features 171
7.9 Task 7: Refine system 172
7.10 Task 8: Generalize 174
7.11 Task 9: Complete and review system 175

8 BON standard activities 178

8.1 Finding classes 178
8.2 Classifying 189
8.3 Clustering 194
8.4 Defining class features 198
8.5 Selecting and describing object scenarios 203
8.6 Working out contracting conditions 205
8.7 Assessing reuse 212
8.8 Indexing and documenting 215
8.9 Evolving the system architecture 220

Part IV Case studies 229

9 A conference management system 231

9.1 System borderline 232
9.2 Candidate classes in the problem domain 235
9.3 Class selection and clustering 236
9.4 Class definition 242
9.5 System behavior 252
9.6 Formal class description 260

10 A video recorder 270

10.1 System borderline 270
viii CONTENTS

10.2 Candidate classes 274

10.3 Class selection and clustering 275
10.4 Class definition 279
10.5 System behavior 283
10.6 Formal class description 286
10.7 Final static architecture 288

11 Relational and object-oriented coexistence 289

11.1 From data storage to object persistence 290
11.2 Object models and relational models 293
11.3 A relational database wrapper 295
11.4 Interfacing an existing relational schema 303
11.5 Querying a persistent object model 312
11.6 Persistent object management 315
11.7 Automatic generation of SQL statements 324
11.8 Full static architecture 331

12 Exercises 332
12.1 Clustering a problem domain 332
12.2 Defining class relationships 334
12.3 Assertions and classification 335
12.4 Dynamic behavior 336
12.5 Prescription and description 337
12.6 Traffic-control system 338
12.7 Dice game 339
12.8 Car rental company 340
12.9 Truck freight 341
12.10 Real-time process control 345

Part V Appendices 347

Appendix A BON textual grammar 349

A.1 Introduction 349
A.2 The syntax notation 350
A.3 BON specification 352
A.4 Informal charts 352
 CONTENTS ix

A.5 Static diagrams 354

A.6 Class interface description 355
A.7 Formal assertions 357
A.8 Dynamic diagrams 358
A.9 Notational tuning 359
A.10 Lexical components 359

Appendix B BON textual examples 364

B.1 BON specification 364
B.2 Informal charts 365
B.3 Static diagrams 367
B.4 Dynamic diagrams 374

Appendix C BON quick reference 381

Appendix D Other approaches 386

Appendix E Glossary of terms 388

References 411

Index 425
x CONTENTS
Series editor’s preface

A rumor has been spreading for some time among people that follow progress in
object-oriented analysis and design: “Wait for BON!” Those not in the know
would ask what in the world BON could be. Indeed, the publicity around the
Business Object Notation has been modest—an article in the Communications of
the ACM, presentations at a workshop or two, public seminars in Europe and
North America, tutorials at TOOLS and other conferences—but it was enough to
attract the attention of many O-O enthusiasts who were dissatisfied with the
limitations of first-generation analysis methods. In the meantime, BON was
being taught to many practitioners, applied in numerous industrial projects, and
repeatedly polished as a result.
As this book finally reaches publication it is certain to cause a major advance
in the field of object-oriented methods. Its most remarkable feature is the
thoroughness with which it applies object-oriented principles, unencumbered by
leftovers from earlier methods. Going O-O all the way is not a matter of
dogmatism, but the secret for obtaining the real benefits of the method, following
in particular from two principles developed at length in the book: seamless
development, the removal of artificial gaps and mismatches between successive
software development activities; and reversibility, the recognition that at any step
of the development process, including implementation and maintenance, it must
be possible to update the results of earlier phases such as analysis and design,
and still maintain full consistency between the analysis, design, implementation
and maintenance views. By ensuring seamlessness and reversibility it is possible
to obtain a continuous software development process, essential to the quality of
the resulting products.
This book is also one of a select few in the OOAD literature that pays serious
attention to the question of software reliability, by using some elements of
formal reasoning, in particular assertions, as a way to specify semantic properties
of a system at the earliest possible stage.
Following the presentation of the model and method in parts I, II, and III, a
large section of the book (part IV) is devoted to a set of in-depth case studies and
to exercises, drawn for the most part from projects in which the authors acted as

xi
xii SERIES EDITOR’S PREFACE

consultants. This abundant practical material will help readers apply the ideas of
BON to their own application areas.
From now on, no one will be able to claim knowledge of object-oriented
analysis and design who has not read Kim Waldén and Jean-Marc Nerson.

Bertrand Meyer
Preface

In the past few years, object-oriented techniques have finally made the passage
from the programming-in-the-small island to the mainland of programming-in-
the-large. Accompanying this transition has been a change in the role and
perception of software methods: in addition to their well-established use in the
earliest stages of a project—requirements analysis and system specification—
they are increasingly viewed as providing the intellectual support needed across
the entire software construction process, through design and implementation to
maintenance and reengineering. The object-oriented approach is best suited to
achieve this seamlessness of the software development process, without which it
would not be possible to meet the quality and productivity challenges that
confront the software industry.
This book shows how a consistent set of object-oriented abstractions can be
applied throughout the process, based on three major ideas: seamlessness,
reversibility, and contracting.
Seamlessness, as in the first word of the title, follows from the observation
that the similarities between the tasks to be carried out at the various steps of a
project far outweigh their inevitable differences, making it possible to obtain a
continuous process that facilitates communication between the various actors
involved, ensures a direct mapping between a problem and its software solution,
and results in a high level of quality for the final product.
Reversibility means that the seamless procedure must work in both directions:
if one modifies a system that has already reached the implementation phase—a
frequent case in practice—it must be possible to reflect the modification back to
the higher levels of design, specification, and analysis. Without such
reversibility the products of these earlier stages would soon become obsolete,
raising disturbing questions about their very role in the software process. Since
current object-oriented methods are still dominated by hybrid approaches—that
is to say, encumber the application of object-oriented principles with techniques
drawn from non-object-oriented analysis methods and with constructs drawn
from non-object-oriented languages—reversibility has so far been almost absent
from the concerns of the object-oriented literature.

xiii
xiv PREFACE

The contract model was introduced to a wider audience as early as 1988 by

Bertrand Meyer in his magnificent introductory book Object-Oriented Software
Construction (OOSC), which quickly became, and still is, the standard reference
on basic object-oriented concepts. In a sense, the present book is a continuation
of OOSC, carrying some of its software engineering ideas to their logical
conclusion in the area of analysis and design. The result is a method called BON
(Business Object Notation) which contains a set of concepts and corresponding
notations to support object-oriented modeling centered around the three
principles of seamlessness, reversibility, and software contracting.
In the rapidly growing field of object-oriented software development, many
subfields have now accumulated enough experience and techniques to warrant
books of their own. When presenting our ideas, we therefore had to make a
choice: either to cover most of the interesting areas and remain shallow, or to
limit the scope and leave room for more substance. We chose the latter, and
BON concentrates on the basic application-independent ideas of general analysis
and design of software systems.
We have also refrained from including yet another explanation of the basic
object-oriented concepts. There are two main reasons for this. First, the
concepts may be simple enough to define, but understanding their implications in
a deeper sense takes much longer. Therefore, a short overview will not be
enough for those who do not already understand the concepts, while a more
substantial discussion will add significant volume and be utterly boring to
experienced readers. Second, in the general spirit of this book, we believe good
texts should be reused rather than rewritten each time they are needed.
So we will assume that the meaning of classes, instances (objects),
polymorphism, dynamic binding, etc., is already familiar to the reader. If not,
we recommend the OOSC book cited above (a significantly revised second
edition is to appear during 1994). As a small compensation for a basic overview,
we have included a reasonably extensive glossary of terms with brief
explanations of the most important BON concepts and many of the familiar
object-oriented terms. It can be used by the novice as a starting point for further
reading, and by the knowledgeable reader to find out more precisely what flavor
we have chosen for widely used terms whose meaning is not fixed in the object-
oriented literature.
BON was initiated in 1989 by Jean-Marc Nerson, then chief developer of the
ISE Eiffel 2.2 environment, who presented early ideas in a tutorial at the second
TOOLS conference held in Paris in 1990. The ideas were picked up at Enea
Data in Sweden by Kim Waldén, then technically responsible for the company’s
Eiffel distribution in Scandinavia, who started to teach courses on BON in
Sweden shortly after. This was the beginning of a collaboration which gradually
led to a joint development of the notation and method. The BON technique was
 PREFACE xv

applied in several industrial developments, and Jean-Marc published three

articles in 1991−92. However, we soon realized that more substantial
documentation on the subject would be necessary to make the method available
to a wider audience, and in 1993 we made the bold decision to jointly publish a
book on BON (bold, because the only time available for this work was evenings
and weekends).
As is often the case with visions whose fulfillment requires far more resources
than are available, it is best not to understand the full extent of the work
beforehand—then you can just go ahead and do it anyway, which is what
happened with this book although at the expense of our friends and families.
However, we found that writing about something you believe in wholeheartedly
has the mysterious capacity of extending the number of daily hours well beyond
the standard 24 (in itself a possible subject of another book). If we succeed in
communicating to our readers just a small fraction of the joy involved in writing
this book, the effort will have been worthwhile.

Scope of the book

The book is intended for software professionals as well as for students at the
graduate and undergraduate levels. We believe it can be read by anyone who has
acquired a general understanding of the problems of software engineering, and
who has some inclination for abstract thinking.
The knowledgeable software engineer used to dealing with practical solutions
may discover that it is not all that easy to keep analysis and design models free
from premature implementation decisions. On the other hand, to achieve a deep
understanding of the technology, it is probably even more important for the high-
level analyst to occasionally take an object-oriented design (not necessarily
large) all the way through implementation. Never to do this is somewhat like
trying to become a mountaineer without ever climbing: there is little replacement
for watching polymorphism and dynamic binding in live action.

Book structure
The book consists of an introduction, three main parts, and five appendices. The
main parts treat in order: the concepts and notations of BON; the BON process
for producing analysis and design models; and a practical part with three case
studies and exercises.
The introduction (chapters 1−2) discusses the general principles which have
guided the development of BON and positions the method relative to other
approaches.
The model part (chapters 3−5) explains the static and dynamic models of BON
and the corresponding notation. Untyped modeling charts are used for the very
xvi PREFACE

early phases, and these are later refined into fully typed descriptions with
semantic specifications added. This part is the core of the book around which
everything else is built, so it should be read carefully. (Sections 3.11−3.13 on
the BON assertion language may be skipped on first reading by those less
interested in formal specification.)
The method part (chapters 6−8) describes how work is carried out with BON.
It starts with a discussion of a number of general modeling issues (chapter 6).
This serves as background for a detailed description of the BON process tasks,
presented in chapter 7. These tasks concentrate on what should be produced (the
deliverables). Finally, chapter 8 discusses the standard modeling activities
needed to produce the desired results, and is focused on how to attack the various
subproblems.
The practical part (chapters 9−12) then presents three case studies and a
number of exercises (collected in chapter 12). The three case studies model in
turn: a conference management system; the control system of a video recorder;
and a mapping between a relational database and an object model.
The concluding five appendices contain in order: a complete grammar for the
BON textual language; a number of examples in the form of textual versions for
several of the graphical diagrams presented earlier in the book; a quick reference
to the BON notation; a list of references to other analysis and design approaches;
and a glossary of terms.

Acknowledgments
BON started as an attempt to extend the concepts of the Eiffel language into the
realm of analysis and design, so indirectly we owe our greatest debt to its
designer Bertrand Meyer. His systematic effort to introduce the powerful idea of
software contracting to the systems development industry, and make it part of
everyday software engineering, has served as our main source of inspiration. We
also thank him for valuable comments and discussions on draft versions of the
book, and for his general moral support.
We gratefully acknowledge the reviewers of Prentice Hall for their insightful
critique and supportive attitude, and Brian Henderson-Sellers for taking the time
to read a full draft and provide helpful comments. We thank Hans Marmolin for
sharing his views on user-centered design and for letting us present his reference
model in chapter 6.
We thank the members of the European consortium “Business Class”, derived
from the ESPRIT II Research and Development Program, which partly
sponsored early work on BON through the development of a Design Workbench
under the project leadership of Jean-Pierre Sarkis.
 PREFACE xvii

We would also like to thank the staff at Enea Object Technology, many of
whom have served as guinea pigs when new ideas were tried out. Special thanks
go to Per Grape, who followed the text closely as it developed, read a number of
successive book drafts, and contributed several improvements to the overall
method and notation. Nils Undén, Björn Strihagen, and Michael Öberg read full
drafts and helped weed out early mistakes. Magnus Lövkvist raised several
issues during the application of BON in a commercial project, which lead to
improvements of the text, and Roland Persson provided interesting views from
his experience with Smalltalk applications. Daniel Rodríguez, Björn Strihagen,
Lennart Gustafsson, Jan Erik Ekelöf, Niklas Odenteg, Mike Henry, and Michael
Öberg used BON in several contexts and provided valuable feedback. We thank
them all for their contributions.
We are grateful to Anders Johansson from Cap Programator for helpful
comments on successive drafts, and for his strong support of the BON approach.
Thanks also to Roy Clarke from LM Ericsson Data and Keith Gunn from SHL
Systemhouse for their detailed comments on the English language and for
detecting errors in some of the examples through careful reading. Special thanks
to Michael Öberg and Juha Juslin, present and former managers of Enea Object
Technology, for their positive attitude regarding the production of this book.
The initial impulse on object-oriented analysis and design leading to the basic
ideas that later became BON was given by David J. Hopkins. The BON
acronym was first coined by Christine Mingins, and later reinterpreted as
“Business Object Notation”. In all, the BON method and notation has been
influenced and recast by more than four years of continuous industrial practice
and experience.
Finally, we express our gratitude to all friends and family members for bearing
with us during a full year of almost total anti-social behavior.

Stockholm and Paris K. W.

August 1994 J.-M. N.
Trademark notice

All trade names, service names, trademarks, or registered trademarks

mentioned in this book are property of their respective owners.

Authors’ addresses

Kim Waldén, Enea Data AB

Box 232, S-183 23 Täby, Sweden
Electronic mail: [email protected]

Jean-Marc Nerson, Société des Outils du Logiciel,

104 rue Castagnary, 75015 Paris, France
Electronic mail: [email protected]
 Part I
Introduction
2 PART I
1 Object-oriented software
development

1.1 INTRODUCTION
What is the potential of the object-oriented paradigm? How much improvement
of the software development process can we reasonably expect from using this
technology, which 25 years after its initial invention finally seems to be
conquering the software industry?
Fred Brooks, in his well-known article “No Silver Bullet: Essence and
Accidents in Software Engineering” [Brooks 1987], divides the difficulties of
building software into essence and accidents. The essence of a piece of software
is a construct of interlocking concepts: data sets, relationships among data items,
algorithms, and function invocations. This construct is the general architecture
of the software—that part of its logical structure which is independent of any
particular machine representation, but still detailed enough to allow
unambiguous translation to executable code. The accidents, by contrast, are
everything else—all the gory details and contortions necessary for representing
the essence in a given computing environment.
Brooks believes the hard part of building software is the specification, design,
and testing of the essential conceptual constructs, as opposed to representing
them and testing the fidelity of the representations (the accidental part). If this is
true, he concludes, building software will always be hard. Languages and tools,
no matter how powerful, can only take us that far when the real problem is to
decide what exactly we want to express.
At first sight, Brook’s conclusion may seem to invalidate all claims that
object-oriented abstraction has the potential to increase software productivity by
a significant factor. In fact, if object-oriented techniques are mainly taught and
used to build new systems from scratch, as often seems to be the case in industry
today, only marginal productivity improvements can probably be expected. If,
on the other hand, the emphasis is shifted from individual systems to the

3
4 OBJECT-ORIENTED SOFTWARE DEVELOPMENT

production and use of tailorable software components, a profound change

becomes possible.

Benefits of a reusability approach

There are two reasons for optimism. First, the cost of software can still be
reduced by an order of magnitude by removing most of the accidental difficulties
from industrial software engineering—maybe not for a single system version, but
surely over a product’s life cycle. Methods and implementation languages are
not enough, however, to achieve this cost reduction, no matter how conceptually
powerful and highly automated. We also need access to a large base of reusable
components which encapsulate the basic concepts that are being reinvented over
and over in today’s industrial software projects.
Second, reusable abstractions are not limited to hiding accidental difficulties,
but can also be used to attack the essence of software design. The complexity
involved in solving a problem depends not only on the problem, but just as much
on the primitive concepts available for reasoning about the problem. So if we
can increase the expressive power and understandability of these primitives in
various problem areas, the complexity of corresponding abstract designs can also
be reduced.
As a side effect, investing in reuse brings another crucial advantage. Software
components that are used and reused many times in many different contexts
stand the chance of acquiring much higher quality through successive
improvement than is ever economically feasible for components that are just
used within a single project. This enables new abstractions to gradually evolve
until they become conceptually strong enough to become part of the system
developer’s standard vocabulary. This may, in turn, lead to the discovery of new
useful abstractions at yet higher levels that would otherwise not have been found
owing to the initial effort required.

Initial difficulties
There has been significant effort invested over the past two decades to build and
use repositories of software components for industrial systems development.
Although certain application areas have seen some successes, achieving a high
degree of reuse in the general case has turned out to be much more difficult in
practice than first expected. Much of the failure has been attributed to
organizational shortcomings, such as lack of clear responsibility roles (reuse
managers), no consistent management policy, lack of automated tools support,
and conflicts with short-term project budgets. Other problems are commercial in
nature, such as how to protect reusable designs enough to make the effort
invested worthwhile for the originators. These problems do not go away just
 INTRODUCTION 5

because we switch technology, and must still be solved.

But the key to it all is object-oriented abstraction—the only technique flexible
enough for building the general components needed. Since reuse efforts have
mainly relied on traditional techniques, it is no surprise that they have largely
failed. As long as we lack the basic means to produce the right components,
formal organization and automated browsing tools can do little to help. On the
other hand, object-orientation does not automatically solve all problems, and
many object-oriented projects have also reported difficulties attaining their reuse
goals.
This, however, must not be taken as a sign that large-scale reuse is
unattainable in practice, or that the object-oriented approach does not work. On
the contrary, there are a number of reasons why these initial difficulties are only
to be expected. First, most industrial object-oriented projects are still using
hybrid languages or hybrid methods, or both. The resulting mix of partly
contradictory concepts creates confusion and delays the mental shift necessary to
take full advantage of the new approach. The requirement of backward
compatibility for hybrid languages also makes it impossible to support cleanly all
of the central object-oriented concepts, which in turn makes the construction of
high-quality component libraries difficult.
Second, even if the technical means are a prerequisite and must come first, the
organizational aspects are also crucial. Many projects have failed because of
inadequate training, lack of management support or reuse coordination. These
are problems that must be addressed in parallel, particularly for large
organizations.
Third, the size and quality of commercially available class libraries is highly
variable, and even the best object-oriented environments only cover a small part
of what one would wish for. Since good abstractions need to be developed
incrementally with many alternative approaches tried, it will naturally take some
time before we can expect anything close to a complete encapsulation of the
most commonly reinvented software components.

The road to reuse of knowledge

If we compare the current trend towards object-oriented languages with the
transition to high-level languages in the 1970s and early 1980s, the situation,
although it has many similarities, is also quite different. The control structures
of languages like Pascal and C embody abstractions that the assembly
programmers were already using mentally (often occurring as comments in some
pseudo-Algol notation), so the big payoffs were immediate. When the tedious
and error-prone translations of these constructs into sequences of machine
instructions were no longer needed, work could proceed as before, only much
faster.
6 OBJECT-ORIENTED SOFTWARE DEVELOPMENT

In some important areas, the same is true when moving from what can be
considered a traditional language today (such as Pascal or C above) to a good
object-oriented environment. The ability to use off-the-shelf components
representing the basic data structures so fundamental for almost any computing
algorithm (lists, hash tables, queues, stacks), without the need to know anything
about their implementation, is a direct parallel. Another such area is graphical
interfaces. But object-oriented abstraction means much more, since it can also
be used to create new concepts in almost every conceivable area. This means its
greatest potential (in the long run) lies not in representing the concepts with
which we are already familiar, but rather in serving as a vehicle for inventing
new ones.
This is the main reason why object-oriented technology is a technology of
investment more than of short-term profit (even if the latter is by no means
precluded). The really big payoffs will come from reuse at more domain-
specific levels. It is possible to capture whole application types in so-called
frameworks, and only tailor the small portions that need to be different from one
situation to another. Successful frameworks are hardly ever conceived as such
from the beginning. Rather they evolve by gradual adaptation of a group of
components solving a particular problem into also solving other, similar
problems that occur in practice. The usefulness of the resulting structures is thus
empirically proven, which guarantees low cost/benefit ratios.
So we must not despair if things appear to go slowly—after all, we are
reaching for the stars. The future potential is enormous, and even though
extensive training and organizational support is necessary and not free, we need
not go very far down the road to reuse before our investment starts to show
returns. And from there, things will only get better.
In this book, we will present a view of object-oriented analysis and design
derived from the basic premise that extensive software reuse is indeed essential,
and that it can be attained in practice provided we take advantage of the object-
oriented concepts in a way that is compatible with this goal. This view
emphasizes certain aspects of object-oriented technology which we think have
not been sufficiently addressed.
What exactly, then, are the object-oriented qualities that have the capacity to
turn software reuse into standard practice and finally give the term software
engineering its intended meaning? In addition to the extreme flexibility provided
by the class concept—allowing us to build open components that can be
combined and tailored through inheritance—three crucial aspects of object-
orientation already mentioned in the preface, seamlessness, reversibility, and
software contracting, deserve much more attention than they have had so far in
the literature on analysis and design. We will take a look at them in order.
 SEAMLESSNESS 7

1.2 SEAMLESSNESS
The object-oriented approach is the only method known to date that has the
potential to turn analysis, design, and implementation of general software
systems into a truly seamless process. A smooth transition from user
requirements over analysis and design into running systems has been the goal of
software engineering for over 20 years, but traditional methods (although often
claiming to have the solution) have generally failed in practice. This is not
surprising, since the designers of concepts and notations for such methods are
forced to choose between Scylla and Charybdis. Either you provide an easy
translation to some traditional programming language, which forces the notation
to become just another procedural language (often introducing more complexity
than it solves), or you invent a completely different high-level notation and keep
the barrier between specification and code.
What makes object-orientation so attractive is that the same abstraction
mechanism (the class) can be used in all development phases. The basic
concepts needed to model objects representing such external notions as hospitals,
airplanes, and wide area networks are not essentially different from what is
needed for objects representing quadruple precision floating point numbers,
street addresses, or process dispatchers. The semantic interpretation of the
abstractions encapsulated by the classes may vary, but the general problem
remains the same: to specify class consistency, relations with other classes, and
behavior through applicable operations.
Being able to keep the same paradigm from initial feasibility study all the way
through production and maintenance of a working system brings enormous
advantages. Communication between project members with different roles is
greatly improved when the basic concepts are the same for everybody.
Education is facilitated and the artificial barriers between specifiers and
implementors vanish, making room for a holistic view of the system life cycle.
Seamlessness also facilitates requirements traceability. Since the classes
introduced in the analysis phase will still be present in the final system, tracing
the propagation of initial requirements through design and implementation
becomes much easier.

1.3 REVERSIBILITY
True seamlessness means more than just easy transition from specification to
implementation. Far too many object-oriented methods rely on the unspoken
assumption that the analysis and design notation will only be used in the early
development phases, and then translated once into program code—object
oriented or not. But at some point (in fact, very soon) the initial system will be
8 OBJECT-ORIENTED SOFTWARE DEVELOPMENT

modified to meet new requirements. Ideally, this would mean changing first the
topmost descriptions, and then successively propagating all changes downwards
until the code is reached. However, this is not the way it works in practice for
most systems.
Since high-level specification can only represent a crude sketch of a system,
lots of details and problems ignored at that point will have to be taken care of
before the specifications can be made executable. This means that a whole new
world of abstractions in terms of implementation language concepts will be
created, and the main interest and creative effort will gradually shift to this
environment. Successive refinements and corrections will tend to be applied
directly to the program code, since only there do we have enough expressive
power to resolve all obscurities and detailed decisions that could not be
addressed by the specifications. And some of these details will nearly always
turn out to have a significant impact on the system structure. (If the program
code could be automatically generated from the specifications, the latter would
simply become our new programming language and we would not need to talk
about the lower levels at all.)
However, if abstract system description is to keep its value beyond the first
translation into program code, changes to the code must be reflected back into
the specifications at regular intervals. Here is where all traditional methods
break down. If the conceptual primitives used by the specification and
implementation languages, respectively, cannot be directly mapped to each other
(which is always the case in non-object-oriented approaches) this will lead to a
creeping divergence between specification and implementation. It simply
becomes too expensive to keep the two worlds consistent as the system evolves,
since this would mean repeated non-trivial translations between more or less
incompatible conceptual structures.
In fact, even if you try hard to keep all specifications up to date, there is no
way of knowing if they really are (because of the conceptual mismatch) so
people will usually not trust them anyway. After all, only the executable
specifications, that is the program code, ever get to talk to the hardware which
carries out the system actions. It is the complete program code that decides
whether the airplane will take off and land safely, not the blueprints drawn by the
analyst / designer. A correct system can run without problems even if its
specification is wrong, but not the reverse. Therefore, when we need to choose
in practice which description to favor, the choice is easy.
The value of the specifications is therefore directly related to the ease by
which they can be seamlessly translated to and from program code. Those
claiming that only the very high-level requirements and analysis models matter,
without giving any hint as to how the mapping to and from the executable code
can be done, do not seem to have fully understood what it means to manage the
 REVERSIBILITY 9

multi-billion dollar investment represented by today’s software. It is probably

not a coincidence that the high-level modeling concepts of object-oriented
technology were discovered by people who were struggling to master the
complexity of programming.
Unlike any other approach, the object-oriented method is inherently reversible.
Since the classes of analysis and design can be made part of the final system, any
changes to the implementation affecting the structure and interface of these
classes then become immediately visible, but only if we refrain from including
elements from other fields, such as entity−relationship diagrams, state transition
diagrams, or data flow diagrams as standard parts of our approach. Mixing
paradigms breaks the reversibility and introduces new complexity, which will in
most cases outweigh the expected benefits.
This does not mean that such techniques can never be used in object-oriented
systems. Some applications may benefit from an occasional entity−relationship
diagram, and modeling certain abstractions using state transition diagrams can be
extremely powerful, but basing a general method on them misses the point.
Instead, we should take advantage of the special qualities of object-orientation:
its simplicity, coherence, and extreme generality. It provides the same support
for abstraction at all levels without forcing them to be viewed in any particular
way. This makes the approach unique among development methods, and its
basic concepts have proved sufficient to specify and implement most of the
software we need, almost regardless of application area.

1.4 SOFTWARE CONTRACTING

Software designed for reuse needs to be of extra high quality, since its potential
to increase productivity also brings the risk of causing much more harm than
before. Writing most of the software from scratch in traditional style at least has
the advantage of limiting the effects of mistakes to the particular system being
developed. However, if an inadequate software component is used in thousands
of applications, the accumulated damage can be very high indeed. To a certain
extent this is countered by the extensive testing a heavily reused piece of
software is subjected to, but testing can only reveal a small percentage of
potential errors and whenever the usage pattern changes, previously concealed
problems are likely to manifest themselves.
The whole idea of the object-oriented approach is to design software to mirror
the high-level concepts in various application domains by building successively
more powerful components in layers of abstractions, each standing on the
shoulders of the previous ones. We know of no better way to master complexity,
but it also means that the resulting structure becomes totally dependent on the
correct functioning of its constituent parts; if some of the central abstractions
10 OBJECT-ORIENTED SOFTWARE DEVELOPMENT

fail, the whole building may fall apart. It is therefore even more important than
before to find ways to guarantee software correctness.
Fortunately, in recent years a very promising method has been proposed to
bring elements from the research fields of abstract data types and formal
specification into standard use in software engineering. This is the theory of
software contracting [Meyer 1992c]. The idea is to use assertions to define the
semantics of each class. The prerequisites and resulting behavior of each
operation are specified through pre- and postconditions, and the overall class
consistency through the class invariant. These semantic specifications then
form the basis for a contract between each class, the supplier, and all classes
using its operations, the clients. A software system is viewed as a network of
cooperating clients and suppliers whose exchange of requests and services are
precisely defined through decentralized contracts.
Based on the contracts, a consistent error handling mechanism is possible. If
the assertions are monitored at run-time, contract violations can be made to cause
system exceptions. Decentralized handlers may then be defined and
implemented as part of a general exception management facility to take care of
error recovery.
Software contracting represents a significant step towards the routine
production of correct software and should be included in any object-oriented
analysis and design method aimed at building reliable, high-quality professional
products.
2 The BON approach

2.1 INTRODUCTION
The method described in this book is called BON, which stands for “Business
Object Notation”. It presents a set of concepts for modeling object-oriented
software, a supporting notation in two versions—one graphical and one textual—
and a set of rules and guidelines to be used in producing the models. BON
focuses on the fundamental elements of analysis and design, and the method is
meant to be integrated with and adapted to the various development frameworks
and standards that may apply in different organizations.
BON supports general software development with no special application types
in mind, and is particularly aimed at products with high demands on quality and
reliability. The concepts and notations are designed to encourage a reusability
approach by emphasizing the points raised in the previous chapter: seamlessness,
reversibility, and software contracting. Contrary to the somewhat resigned
attitude found also in many object-oriented camps about the attainability of
massive reuse, we claim that this is indeed the major goal of the technique.
BON does not introduce any fundamentally new concepts; the basic object-
oriented ideas combined with elements from software specification are sufficient
as primitives. Rather it is the detailed definition and arrangement of the concepts
expressed by a scalable notation that can make a qualitative difference. (To a
certain extent, BON is defined by what it does not include, since seamlessness
and simplicity are the guiding stars.)
The reader may of course wonder whether this goal could not have been
achieved without introducing a new notation when so many have already been
published. Could we not just adapt one of the more widely used existing
notations to fulfill our purpose? Unfortunately not. Although the concepts used
in many proposed methods may seem to be more or less the same (classes,
operations, relations), there are subtle differences below the surface that prevent
a one-to-one mapping between them. Since a notation is just a way of presenting
the underlying concepts in a comprehensive and readable form, reuse does not
work in this case.

11
12 THE BON APPROACH

The field of object-oriented analysis and design is still young and immature,
and it is only natural that many competing approaches and accompanying
notations will continue to emerge until it is time for a general shake-out. Even
though this may create some confusion for potential users, it is really in their best
interest. Standardizing too early is extremely harmful, since it narrows the
modeling perspective and stands in the way of real understanding. (We are
pleased to see that this view is shared by many well-known specialists in the
field through a recent open letter “Premature Methods Standardization
Considered Harmful” [Mellor 1993].)

2.2 WHAT IS NOT IN BON

Many existing methods for analysis and design, probably the majority, include
either data modeling using some variant of the entity−relationship approach (ER
modeling [Chen 1976]) or finite state machines (FSM modeling), or both, as a
significant part of their high-level system description. The idea is to combine the
strengths of these techniques (which are well understood and have been in use
for a long time in traditional environments) with those of the object-oriented
concepts, in order to benefit from both worlds.
However, such approaches seriously impede the seamlessness and reversibility
advocated in the previous chapter. In our view, this disadvantage far outweighs
any benefits gained. With FSM modeling, the impedance mismatch is obvious
since there is no easy mapping between the state transition graphs and an
eventual implementation (unless we actually model every object as a state
machine, thereby giving up all abstractional power of the class concept). With
ER modeling, the situation is a bit more complicated.

Why not ER modeling?

Proponents of analysis-level ER modeling claim that binary associations are
more general than references between classes, since in the latter case explicit
design choices have been made that are still kept open in the former. This is of
course true, since a particular association between two classes can be represented
in many ways. However, the alternatives kept open by restricting all class
dependencies to binary associations are usually not the interesting ones.
In fact, in order to arrive at a good system model we cannot keep everything
open, because this would effectively prevent us from understanding the problem.
So the question is not whether design decisions should be made, but rather which
ones to make.
ER modeling is of course no exception to this. On the contrary, just as many
arbitrary or questionable decisions must be taken in order to arrive at an ER
 WHAT IS NOT IN BON 13

model, as with a pure object-oriented model. In the latter case, we must choose
which concepts to model as classes and which ones should become operations on
the classes. With ER modeling, we must instead choose which concepts to
represent as entities, which ones will become attributes of the entities, and which
ones will become associations between them. This is by no means easier or less
susceptible to change than choosing classes and operations.
For example, an attribute in ER modeling is viewed as a “property” of an
entity, and is represented by a value. But what is a property? Consider the entity
EMPLOYEE and the concept of being the most popular person in a service
department. Clearly, being most popular is a personal property, but modeling it
as an attribute of EMPLOYEE may be quite wrong from the system’s point of
view. The employee abstraction may not know about its condition, and the only
way this knowledge will manifest itself may instead be as a combination of
attributes of other entities, perhaps STATISTICS and CUSTOMER_POLL.
So we have a problem here: either an attribute is thought of as corresponding
directly to a data value stored in the entity, in which case it is too low level, or
else it is just a vague “property”, which is too high level, since it does not tell us
enough about the system. The object-oriented middle way is a class operation
returning a value, which may be of any type. This avoids premature decisions
about where various values will be stored, but still tells enough about the system
behavior to allow for seamless transition into implementation.
Another trouble spot is what level of “normalization” to choose for the
attributes and relations between entities. A binary association between two
entities A and B is generally not supposed to tell anything about how A and B
are connected (other than through semantic labels, such as “works for” and the
reverse role “employs”). By this reasoning, the transitive law must also apply: if
B is in turn associated to C through the association “has child”, then A is
associated to C through “works for parent of”, and C to A through “is child to
employer of” (see figure 2.1). Since ER models are usually connected graphs,
applying the law recursively yields a diagram where every entity has a binary
association to every other entity.

works for
A B
employs

works for parent of has child

is child to employer of
has parent
C

Figure 2.1 ER modeling: transitive law

14 THE BON APPROACH

Therefore, only the relations considered most important are included in the
graph, while the rest remain implicit. Some authors recommend separating an
orthogonal base of independent attributes and relations, and mark the others as
derived if they are shown at all. However, it is far from evident which ones to
choose as the orthogonal base, and it is also not clear what “derived” means.
For example, consider the entities depicted in figure 2.2. We may pick the
two associations between MOTHER−SON and MOTHER−DAUGHTER as the
orthogonal base. The brother−sister association between SON−DAUGHTER
then becomes derived, since for any pair of SON and DAUGHTER we can infer
whether they are siblings or not. But we could also have chosen any other two as
the orthogonal base (assuming brother−sister means full siblings, sharing both
parents).

MOTHER
has son has daughter

has mother has mother

has sister
SON DAUGHTER
has brother

Figure 2.2 ER modeling: derived relations

Moreover, properties of derivability are often global in nature. From the

viewpoint of the SON, the mother and sister roles may be equally important, and
the fact that one happens to be derivable from the other should not always be
emphasized. On the contrary, it may later turn out that the mother role in the
system can also be fulfilled by the entity STEPMOTHER, and then the
derivability will no longer hold.
For the same reason, the difference between derived attributes and base
attributes is not always clear either. To determine derivability one must often
make premature assumptions about the underlying information content. These
problems disappear with a pure object-oriented approach, since we are then
concentrating on system behavior rather than on what information needs to be
stored. Since this behavior is what is needed to fulfill system functionality,
placing the emphasis there results in much more resilience to future change.
The bottom line is that we give up the seamlessness inherent in an object-
oriented approach if we mix in ER modeling at the analysis and design levels,
but we do not get much in return. In practice, real systems often have several
hundred potential entities and relations between them. The resulting complete
ER diagram becomes huge and is essentially a flat structure.
 WHAT IS NOT IN BON 15

To achieve some locality, large diagrams are often split up into smaller
overlapping parts, where each part contains some group of entities together with
a subset of mutual relations. This is made possible by the more or less arbitrary
omission of associations. The technique resembles the splitting of program
flowcharts in older days, but instead of leaving lots of dangling arrows with
references to other diagrams, the borderline relations are simply suppressed.
However, this does not change the inherent flatness of the structure, and rather
than the “zooming” capability so important for understanding a large system, we
are stuck with a form of “panning”.
The emphasis on associations as a modeling concept separated from object
behavior favors a global system view. It breaks encapsulation and concentrates
more on short-term detail than on local concepts that may have the potential to
survive longer. ER modeling as part of an analysis and design approach will not
help us find classes which represent interesting concepts with the potential of
being used in other contexts, except solving part of the current problem. The
seamlessness and reversibility inherent in object-oriented development can
therefore not be fully exploited, which in turn works against reuse. Furthermore,
we do not need this type of separate association, since the object-oriented
primitives can be used directly to model any concepts we want.
For these reasons, BON has been designed to follow a different track. Rather
than trying to include concepts from traditional data modeling or the so-called
structured techniques with all their accompanying drawbacks, a more fruitful
alliance is sought: the combination of object-oriented flexibility with the clarity
and expressive power of strong typing and formal contracts between classes.

2.3 OTHER METHODS

Somewhere between 20 and 50 object-oriented methods have been published to
date (depending on which ones are viewed as separate approaches) and the
number is still growing. Therefore, any kind of short summary will necessarily
become oversimplified and unfair to some. Moreover, most of the methods are
continuously evolving and so it is very difficult to know whether you are really
criticizing the latest version or just something that is already considered obsolete
by the authors.
The reader will therefore not find the usual overview of analysis and design
methods in this book, but is instead referred to the many summaries and
comparisons already available, as for example in [Graham 1994, Wilkie 1993].
A list of publications on other approaches can be found in appendix D.
BON concentrates on the seamless, reversible specification of software, using
the contract model. The basic difference compared to most other methods thus
becomes obvious in the light of the previous discussion. The majority of the
16 THE BON APPROACH

methods listed in appendix D are hybrid methods based on ER modeling and/or

FSM modeling. Some are oversimplified and not expressive enough, while
others are trying to include too much and therefore lack the simplicity necessary
to be an effective help in the search for reusable abstractions.
A few methods are purely object oriented, and as such approaches with which
we sympathize. However, none of them employ strong typing, which makes it
practically impossible to deal with software contracting, other than through
natural language.
We will not go into any detail regarding the strong and weak points of other
proposed methods, since this is not the purpose of this book. An extensive
bibliography on object-oriented literature in general can be found in
[Booch 1994].

2.4 THE BON POSITION

Because of its flexibility, it is not difficult to find mappings into other domains
which would seem to indicate that object-oriented modeling is really equivalent
to finite state machine modeling, or data modeling with operations added, or
process modeling with signal passing, or even functional abstraction. However,
these mappings are not very useful (they are applicable in much the same sense
as any procedural program is equivalent to a Turing machine). To achieve full
benefit from the technology one needs to concentrate on the aspects which are
most important.
Object-oriented encapsulation is an extremely general approach. It can be
used to capture abstractions of such clarity that they act as live teachers to
anyone reading the specifications. On the other hand, inappropriate
encapsulation can create a mess of incomprehensible complexity. The class
concept takes no stand about what is represented; it is up to the designer to fill
the corresponding type with meaning. In many other disciplines this attitude of
generality would be considered too tolerant, but in software production it is
exactly the quality we seek.
Since software is developed to meet just about any conceivable need, the
concepts used to mold it must be tailorable in any desired direction. As soon as
we think we know what is best in all situations and start imposing one detailed
model for all needs, we are doomed to fail for applications that need a different
view of the world. Our imagination is inherently limited, and the blacksmith’s
child tends to see the world as a collection of tilt hammers and iron bars.
Engineers brought up in environments where system size is measured in
thousands of modules are often convinced that any development needs at least
ten layers of standardized structure descriptions before detailed design can even
begin, while garage floor developers may regard anything beyond a big
 THE BON POSITION 17

whiteboard and a good language environment as mere bureaucratic obstacles.

Only the basic concepts for expressing object-oriented abstractions are general
enough to reconcile these extremes; the rest has to do with tailoring and should
not be included in a general method. Strong typing and the software contract
model, on the other hand, fit like a glove on this flexible instrument and can
provide the support for correctness and reliability needed by the software
industry without impairing its generality.
With the previous discussion as background, we are now ready to present the
basic principles underlying the design of the BON notation.

2.5 CHARACTERISTICS OF THE NOTATION

Generality
The notation is not restricted to any specific application domains, nor does it try
to cover every special need that may arise for certain applications or for certain
programming languages used. Instead, BON concentrates on what is essential
for object-oriented development in general, and tries to define a consistent
notation to support the corresponding concepts. The user is then free to
complement the notation with whatever more might be needed for a particular
project.

Seamlessness
BON regards the seamless approach as the only possible road to extensive future
reuse. Formalisms from other fields, which are often adapted and used as part of
proposed object-oriented analysis and design methods, such as state transition
diagrams, process diagrams, Petri nets, entity−relationship diagrams, data flow
charts, etc., are therefore not addressed in BON. In case one should want to use
some of them as extra support in some projects or application domains, there are
enough existing notations to select from. In particular, we recommend the
statecharts of David Harel as complementary notation for state transition
diagrams, which can be very helpful for some applications [Harel 1988].

Reversibility
To promote reuse and achieve true seamlessness, the core elements of a notation
for analysis and design should represent concepts that are directly mappable not
only to, but also from, an executable object-oriented language.
Besides making it possible to maintain long-term consistency between
specification and implementation, reversibility is also important for the reuse of
analysis and design elements. We cannot expect the current set of proposed
18 THE BON APPROACH

notations for analysis and design to be standardized in the near future. As

advocated earlier, this is not even desirable, since we first need more experience
from their use in real projects. Therefore, broad reuse of high-level specification
elements will have to rely on something more stable. A small extension of a
widely available standardized object-oriented language can fill the need for a
common reusable and exchangeable representation, whose graphical view may
vary. At present, Eiffel [Meyer 1992a] seems a good candidate as the reference
language for such a notation.
The great advantage with this approach is that the notation can be guaranteed
to map easily to a set of concepts proved to be implementable in practice, and
with adequate efficiency for industrial software products. The notation should
also be applicable to already existing networks of classes and then used as a
high-level documentation aid. The ability to generate high-level views on
collections of classes developed without notational support would mean a
tremendous advantage when trying to understand large, external class libraries.

Scalability
The system examples found in textbooks on analysis and design are nearly
always small, which is natural since important points need to be illustrated
without being obscured by too much irrelevant detail. However, we must make
sure that the notation will scale up, and still be useful for large systems. A
diagram for a toy example might look nice and clean using almost any notation,
but facing real-life systems is quite another story.
The first thing to note is that whenever system size reaches more than 20−30
classes, we need something more than the class concept to describe its structure:
a facility to group classes into higher-level units. We will use the term clustering
for such a facility, and a group of classes chosen according to some criterion will
be called a cluster. The reason clusters are normally not included among the
basic concepts of object-oriented languages is that classes are reusable
abstractions that may be grouped differently at different times and in different
contexts. The cluster therefore represents a much looser structuring than the
class, and the exact configuration of classes is more flexibly handled by system
descriptions outside the programming language proper.
In fact, during implementation we usually need at least two ways of clustering
the classes in a system. First, we need to tell the compiler environment where to
look for classes that are being referenced by other classes. This is often done
through specification of a number of class directories, and a search order.
Different class versions can then be substituted by simple modification of the
search list. Second, we need a different clustering to be used in analysis and
design diagrams (usually with more layers for the high-level classes) whose
 CHARACTERISTICS OF THE NOTATION 19

purpose it is to make the system more understandable to human readers.

The second thing to note when scaling up is that flat partitioning is not enough
to get comprehensive views of large systems. We need views at different levels
of detail and the ability to “zoom” between them. The BON notation uses nested
clustering and element compression to achieve this. Element compression means
representing a set of graphical or textual elements with a simpler element, its
compressed form. A typical example is the use of icons in window-based
graphical user interfaces. Elements containing compressed forms may in turn be
compressed, yielding a powerful method to create views at different levels.
Since the level of compression can be independently selected for each
structural element (recursively), the user may freely choose the amount of detail
shown for each part of a system. New interesting aspects can be shown in
comprehensive form while still keeping the overall structure for easy reference.
This is what is meant by scalability, a major goal in the design of BON.
A few simple changes in the compression levels may lead to a dramatically
different architectural diagram, making the approach particularly well suited for
automatic tool support. Although often used in advanced systems for structured
document manipulation (see for example [Meyer 1988b]), this powerful idea
does not seem to have been exploited in other analysis and design methods.

Typed interface descriptions

A class interface consists of the syntax and semantics of its operations. The
syntactic part of an operation, often called its signature, consists of its name, the
number and types of its arguments (if any) and the type of its return value (if
any). There are two basic policies with regard to types—static typing and
dynamic typing.
Static typing means that the user must specify the types of all operation
signatures (along with the types of all local variables used) in the class text. This
policy is employed by most object-oriented languages which are targeted at
industrial software production, for example C++ [Stroustrup 1992] and Eiffel
[Meyer 1992a], and permits type checking at compile time, as well as generation
of efficient code.
Dynamic typing means that only the names of operations and arguments are
specified in the class text, while the actual object types are decided at run-time.
This policy is employed by languages where extreme flexibility is deemed more
important than safety and efficiency, notably Smalltalk [Goldberg 1983].
What many analysis and design approaches fail to recognize is that static
typing not only permits early type checking and implementation efficiency—it is
also an essential aid for system specification. Assigning types to entities is really
classification, which increases the information content of a set of interface
20 THE BON APPROACH

descriptions considerably. (A type in this context refers either to a basic type

like an integer, a real number, a boolean value, or to a class.)
Instead of having to rely on a combination of uncertain naming conventions
and general comments to guess what types of object will be bound to what
names at run-time, the reader can see the types directly. The precise properties
of each type can then be found by simply inspecting the corresponding class or
basic type, so readers are automatically provided with a data dictionary.
As a simple example we may take a class attribute that will refer to a sorted
list of airplanes (where the order relation could be based on the number of seats
for passenger planes and the load capacity for freighters). In Smalltalk the
untyped attribute name would typically be something like:
fleetOfSortedAircraft
which may be compared with the corresponding typed Eiffel declaration:
fleet: SORTED_LIST [AIRCRAFT]
The amount of precise information conveyed by the two forms is very different.
(In fact, the Smalltalk community is beginning to recognize the lack of static
typing as a problem, and extensions to add typing to the language have been
proposed [Bracha 1993, Wirfs-Brock 1991].)
BON adopts a fully typed notation for class interfaces, but also provides a
very high-level untyped view to be used in communication with non-technical
people.

Support for software contracting

The theory of software contracting [Meyer 1992c] is an important advance in the
quest for correctness and robustness in professional software products. As was
pointed out in the previous chapter, these qualities become even more important
for reusable components, whose behavior will affect many applications. Since
software contracting means adding specification elements to classes, it is ideally
suited as part of a method for object-oriented analysis and design.
Despite this, the contracting idea is at best briefly mentioned in books and
articles on the subject, but seldom emphasized the way it deserves. One reason
may be that few other methods employ a typed notation for specifying class
interfaces, and without typed interface descriptions of each class, it is difficult to
express software contracts formally. So besides its general classification power,
typing also plays an important role in software contracts.
Another crucial point is that seamlessness requires an object-oriented system
and its specification to be based on the same conceptual view of the world. This
important principle needs elaboration.
 CHARACTERISTICS OF THE NOTATION 21

The specification elements or assertions used in software contracts (pre- and

postconditions and class invariants) are statements about the state of the system
before and after applying certain operations. Since an object-oriented system
should be defined exclusively by the external behavior of its objects, the system
state is only available indirectly through public operations, called queries, which
are applicable to the objects.
Including lower-level details, like state variables, in external specifications
would break the encapsulation. Therefore, the primitive elements in assertions
are object queries and basic constants, which may be combined by operators to
form logical expressions. The semantics of each query used in an assertion is in
turn specified by the class of the object on which the query is applied. So an
object-oriented system specification is a set of classes, some or all of whose
operations are defined in terms of the specification of other classes in the set
(recursively).
Most of the objects queried in assertions will also be part of system execution,
and thus the corresponding classes part of the implementation. This is fairly
obvious, since there would not be much point in specifying the behavior of the
objects in a system by mostly reasoning about the behavior of an entirely
different set of objects. Some classes may have been introduced strictly for
specification purposes, such as predicate logic support or advanced set
operations, but these may still be implemented to enable assertion checking at
run-time.
Classes will be added in later phases for two main reasons: incomplete initial
specification and choice of implementation. But the conclusion is that an object-
oriented system and its specification should share the same base of class
abstractions. The idea is that in a well-designed system, most of the classes and
operations needed in specifications to capture the necessary state information are
already available as part of the system. The rest can be added as new application
classes, new operations on existing application classes, or as general
specification support classes.
The result is a system which contains its own decentralized specification as an
integral part. The class operations are partly defined by successively referring to
already specified operations whose semantics are known, thus making them fit to
use as extensions of the specification language. Care must of course be taken to
avoid circular specification and the classes need to be checked and tested in
suitable order to give a well-defined result, but these are issues that must be
addressed by any method.
The great advantage of the approach is that when specification and
implementation share the same set of abstractions, system changes will
automatically be applied to both at the same time. Instead of having to witness
theory and reality gradually drift apart as a system evolves over time, software
22 THE BON APPROACH

managers can now watch the two act as one organic unit turning in concert to
meet the changing requirements of a competitive market.
The specification language of BON uses first-order predicate logic to combine
state functions. This is not enough to fully specify a system, but it takes us a
significant step forward. The recursive contracting model represents a powerful
view of software development that can help produce systems of much greater
clarity and correctness.

Simplicity
Perhaps the most important of all general principles for conceptual models, as
well as for notations, is simplicity. The deep results of the natural sciences seem
to indicate that nature is inherently simple—that complexity is only introduced
by our lack of understanding. The essence is well captured by the French writer
and aviator Antoine de Saint-Exupéry:

It seems that the sole purpose of the work of engineers, designers, and
calculators in drawing offices and research institutes is to polish and
smooth out, lighten this seam, balance that wing until it is no longer
noticed, until it is no longer a wing attached to a fuselage, but a form
fully unfolded, finally freed from the ore, a sort of mysteriously joined
whole, and of the same quality as that of a poem. It seems that
perfection is reached, not when there is nothing more to add, but when
there is no longer anything to remove. [Terre des hommes, 1937]

Of course there is still a long way to go before object-oriented analysis and

design reaches the maturity of airplane construction, but the above may serve as
a main source of inspiration.
The BON notation strives for simplicity and tries to minimize the number of
concepts. For example, there are only two basic relations between classes: the
inheritance relation and the client relation. To obtain multiple views of a system,
we also need relations between classes and clusters and between clusters and
clusters. However, instead of introducing new concepts we use the compression
mechanism discussed earlier to generalize the class relations, and give them
well-defined semantics when clusters are involved.
The classes of a system need to be grouped according to various criteria in
order to make their structural and functional properties visible. Although the
nature of these groups may be very different—subsystems, classes with related
functionality, heirs of a common ancestor, individual parts of a whole—BON
uses only one concept to cover them all: the cluster. Also, there is only one type
of inheritance; differences in purpose and semantics are instead taken care of by
contracting elements.
 CHARACTERISTICS OF THE NOTATION 23

Space economy
The abstract concepts underlying the elements of a graphical notation are
certainly much more important than the particular symbols used. Discussing
details of geometric shapes becomes meaningless unless you know exactly what
you want to illustrate. However, this does not mean that the notation is
unimportant. On the contrary, if it were, we could just forget about fancy
graphics requiring special equipment and software and use only plain text. The
reason we still insist on a graphical presentation is the possibility of
communicating views of the underlying model much faster and more accurately
to a human user.
A graphical notation is a language, and like any language it can give rise to
endless discussions about individual elements—should we use single or double
arrows, ellipses or rectangles, dashed or continuous borderlines?—whose merits
are very much related to personal taste, cultural context, and plain habit. (Again,
this does not mean that such elements are unimportant or basically equivalent,
only that it is difficult to achieve consensus about them.) However, there is one
aspect of any notation designed to give a global overview of a potentially large
and complex structure which is important regardless of the details, and that is
economy of space.
The amount of information that can be conveyed by an overview of some part
of a system—a cluster, a group of clusters, a group of related classes—is very
much dependent on what can be made to fit on one page (where a page is a
terminal screen, a paper sheet, or whatever can be inspected in one glance).
Breaking up an integral context into several pieces that must be viewed
separately is extremely detrimental to the global picture. Since systems often
require that quite a few classes be shown simultaneously for the user to get
comprehensive views of the system structure and the relations between its parts,
it becomes very important to avoid wasting space.
The BON notation pays attention to this problem by providing compressed
forms for all space-consuming graphical layouts. For example, it is too
restrictive (as in many other notations) to have a full class interface with
operations and attributes as the only way to show a class. In BON, the
compressed form of a class is simply its name enclosed in an ellipse (possibly
annotated by small graphical markers). Similarly, BON provides the iconization
of clusters and compression of relationships between classes belonging to
different clusters into relationships between clusters.
It is often possible to pack a considerable amount of information in one high-
level view of a system (or part if it), and still keep the view on one page. The
key is interactive tailoring by successive compression and expansion of various
parts of a diagram until the user’s intent is optimally conveyed.
24 THE BON APPROACH

Basis for intelligent case tools

Well-designed object-oriented systems tend to be small compared to traditionally
developed systems with equivalent functionality. This is because common
patterns of behavior are reused in layers of abstraction rather than repeated over
and over with small variations; flat code is thus transformed into system
structure. However, the advantages gained are not completely free—there is a
price to pay for the increased flexibility and reduction of overall complexity.
The compact object-oriented code is by necessity more intricately structured in
order to achieve more functionality per line of text. Particularly, extensive use of
inheritance tends to fragment the class text. It is often very hard to grasp the full
abstraction represented by classes which have maybe 10−20 ancestors along
multiple lines of heritage if the only means is inspecting the bodies of the classes
involved.
Obviously, we need automatic tool support. For example, as client to a class
we are not necessarily interested in its position in a classification structure.
Therefore, we need a tool that can present the full set of operations of any class,
whether these operations were inherited or not. One of the great advantages with
the increased semantic content in a typed object-oriented notation with software
contracts is that it provides the foundation for much more intelligent object-
oriented case tools than is possible in untyped environments.

2.6 STATIC AND DYNAMIC MODELS

There are many ways to describe a software system, but all descriptions can be
characterized as being either static or dynamic. Static descriptions document the
structure of a system: what the components are and how these components are
related to each other. They do not take time into consideration, but are either
time independent or represent snapshots at certain points in time.
Static descriptions are of two kinds: they either tell you something about the
structure of the system software (that is, how the program modules are organized
in the eventual implementation), or they tell you something about the structure of
the interacting components that come into play when the system is executed.
Both are of course necessary for maintainability of the software being developed,
but the good news is that with a pure object-oriented language the same static
description can be used to capture both aspects. Since a class describes the
behavior of a certain type of object—and nothing else—using the class as the
basic program module makes the static structure of the class text in the software
system map directly to the structure of the objects at execution time.
Dynamic descriptions, by contrast, document how the system will behave over
time. In an object-oriented context, this means describing how objects interact at
 STATIC AND DYNAMIC MODELS 25

execution time; how they invoke operations on each other (passing messages in
Smalltalk terminology) and how the information content of the system changes,
as reflected by the values of the class attributes (state variables) in the system.
These two types of description are very different, and confusion can easily
arise unless they are kept apart. An object-oriented system is best viewed as a
structured collection of classes, each being a (possibly partial) implementation of
an abstract data type. The classes constitute the blueprints specifying the
behavior of each object (class instance) created during a system session. At
system execution time, on the other hand, only communicating objects will exist,
while the classes are left on the engineer’s desk. Each object will behave as
prescribed by its class (serving as its genetic code), but just like a biological
creature it has no access to what is actually governing its pattern of behavior.
The analysis and design of an object-oriented system using the BON method
will result in static and dynamic descriptions of the system being developed.
The static descriptions form the static model of the system. This model contains
formal descriptions of class interfaces, their grouping into clusters as well as
client and inheritance relations between them, showing the system structure. The
dynamic descriptions, on the other hand, make up the system’s dynamic model.
This model specifies system events, what object types are responsible for the
creation of other objects, and system execution scenarios representing selected
types of system usage with diagrams showing object message passing.
Part II (chapters 3−5) will describe these views in detail, the graphical notation
used, as well as the underlying concepts. Part III (chapters 6−8) will then be
devoted to the BON method, containing rules and guidelines to be used in
producing the models.
26 THE BON APPROACH
 Part II
The model
28 PART II
3 The static model—classes
and clusters

3.1 INTRODUCTION
The static model shows the classes making up the system, their interfaces, how
they are related to each other, and how they are grouped in clusters. It
concentrates on the what part and downplays the how. It also fits well when
object-oriented formal specification techniques are used in the early development
stages (for an overview, see [Lano 1994]).
There are two parts to the static model. The first part is a collection of very
high-level untyped modeling charts, which can be used early in the analysis
process to enhance communication with domain experts and end-users, and as
partial documentation aimed at non-technical people. The charts work much like
a set of structured memos to help sort out the initial disorder and overwhelming
amount of requirement details (often contradictory) so common in early
modeling. They were inspired by the index cards used in the CRC method at
Tektronix [Beck 1989].
The second part is a structured description containing fully typed class
interfaces and formal specification of software contracts. This is the main part of
the static model in BON, and its notation has been developed with graphical
manipulation tools in mind. Classes are grouped into clusters, and clusters may
in turn contain both classes and other clusters.
There are two variants of the BON notation: graphical BON and textual BON.
The graphical form is intended for use with automatic tool support as well as for
sketches on paper and whiteboards. It contains a mixture of drawing elements
and text elements, but the text elements are considered graphical as well, since
their location in a two-dimensional figure may be used to convey information.
The textual form is intended for communicating BON descriptions between
various automatic processing tools and for maintaining evolving architectures by
simple file editing in environments that lack dedicated BON case tools. It also

29
30 THE STATIC MODEL—CLASSES AND CLUSTERS

has a formal grammar defined, which may be used as a compact description of

the whole set of BON’s notational concepts, as well as to clarify possible
ambiguities that may occur in the natural language descriptions.
There is a one-to-one structural mapping between the graphical and textual
form of a BON description, but the textual language does not (for obvious
reasons of simplicity) contain any description of spatial layout. It is thus
possible to generate many different diagrams from the same textual description
(all topologically equivalent). The much more complex problem of choosing a
good presentation layout is left to the designers of BON case tools.
Since graphical BON is the most readable and compact form, it will be the
preferred notation in this book. The corresponding textual forms will only be
mentioned occasionally in the general description of the BON method. The
complete formal grammar of textual BON and examples of its usage are
collected in appendices A and B.
This chapter will describe the basic elements of the static model: the classes
and clusters. We will start with the untyped modeling charts and then proceed to
typed class interface descriptions. We will see that textual descriptions can be
made much more concise using graphical representations of the specification
elements. The formal specification language of BON, based on first-order
predicate logic, will be described.
Finally, we look at the scalability of the BON graphical notation through
compression and expansion: how classes may be compressed into headers, how
clusters may be iconized, and so forth. The next chapter will then be devoted to
the static relations between classes and clusters.

3.2 BON MODELING CHARTS

Very early in the process, informal charts may be used to communicate basic
ideas to non-technical people like end-users, customers, and domain experts.
These charts may also later serve as very high-level documentation for the
system, and be stored with the more formal descriptions by a case tool. There
are three types of modeling chart in the static model:
• system chart
• cluster chart
• class chart
The layout of the charts has been chosen so as to give a feeling of physical cards.
Much of the advantages reported from the use of CRC index cards [Beck 1989]
stems from the social communication they give rise to when analysts and people
from the customer side work together to make a first model of the problem, and
 BON MODELING CHARTS 31

actually pass the cards around in order to fill in the information needed.
The modeling charts can be used in a similar way if they are printed in fixed
format as templates on paper and then filled in manually during modeling
sessions. The analyst can later transfer the information to a case tool, which will
store it as complementary documentation. The case tool may also translate the
information in the class charts into a form that can be used as the beginning of
the more formal class interface specification.

System charts
The system chart (exactly one per system) contains a brief description of each
top-level cluster in the system. In BON, a system consists of one or more
clusters, each of which contains a number of classes and/or subclusters. By
convention, BON does not allow classes at the topmost level, so each class
belongs to exactly one (immediately) enclosing cluster. An example of a system
chart for a car rental system is shown in figure 3.1.
The header of a modeling chart is separated from its body by a double line
with the top two rows in standard format. The first row has the chart type to the
left (SYSTEM in this case) followed by the name of the system/cluster / class
described and a chart sequence identification. Sequencing is needed, since all

SYSTEM CAR_RENTAL_SYSTEM Part: 1/1

PURPOSE INDEXING
System keeping track of vehicles and rental author: Jean-Marc Nerson
agreements in a car rental company. keywords: vehicle, rental

Cluster Description

CONTRACT_ELEMENTS Concepts that have to do with rental agreements such as

contracts, clients, means of payment, rentals.

RENTAL_PROPERTIES Properties of individual rentals such as vehicle, rate, extra

options, insurance policy, certified drivers.

VEHICLE_PROPERTIES Properties of rentable vehicles such as availability, location,

models.

OPTIONS Selectable rental options.

OPTION_TYPE Types of options available such as two-door versus four-door

models, automatic versus manual transmission, sunroof and
hifi stereo equipment.

SUPPORT General support classes for handling of time and date,

executable commands and data structures.

Figure 3.1 System chart for a car rental system

32 THE STATIC MODEL—CLASSES AND CLUSTERS

the entries might not fit on a single chart. The second row has a comment clause
to the left describing the purpose of the chart, and an indexing clause to the right.
The indexing clause contains a number of index entries, and each index entry
consists of a keyword with a list of words attached. The number of entries and
what keywords to use is decided per project. The purpose is to record interesting
properties of systems, classes, and clusters, and to facilitate browsing.
Note that the indexing clause in figure 3.1 has two keywords: author and
keywords. The name of the second keyword suggests, of course, that by the
conventions used for recording indexing information in the organization where
this car rental system is developed, vehicle and rental will be considered
keywords. However, this last interpretation occurs inside the target system and
has nothing to do with the general syntactic keyword construct of BON charts.
From BON’s perspective vehicle and rental are just words attached to the
keyword keywords.
The above illustrates an important point in all types of abstract modeling—the
risk of confusing language with metalanguage. The distinction may be obvious
in this example, but there are much more subtle cases where our cultural
background makes it all too easy to automatically assign meaning to words and
symbols also in contexts where we are not supposed to.

Cluster charts
A cluster chart contains a brief description of each class and subcluster in the
cluster. Subcluster names are enclosed in parentheses to separate them from
class names. The recommended procedure is to list all classes first, and then the
subclusters, if any. This is because subclusters often group local services that are
used by the topmost classes in a cluster. A subcluster may, for example, contain
all specialized descendant classes of a given class.
Clusters may be nested to any depth. Two cluster charts are shown in
figure 3.2, where the first cluster contains a subcluster described by the second
chart. Cluster chart headers are similar to system chart headers, but the name
now refers to the cluster instead. The indexing clause contains an entry with
keyword cluster, listing the nearest enclosing cluster of the class.

Class charts
The class charts model individual classes. Classes are viewed as black boxes,
and the information in the class charts is the result of answering the following
questions:
• What information can other classes ask from this class? This translates
into queries applicable to the class.
 BON MODELING CHARTS 33

CLUSTER ORGANIZATION Part: 1/1

PURPOSE INDEXING
Handles all major events occurring during author: Kim Waldén, Jean-Marc Nerson
the organization and completion of a keywords: organization, staff
conference.

Class / (Cluster) Description

CONFERENCE The root class of the conference system.

PROGRAM Information about the final conference program and its

preparation.

TIMETABLE Repository of scheduled events.

(COMMITTEES) The committees engaged in the conference organization to

take care of the technical and administrative parts.

CLUSTER COMMITTEES Part: 1/1

PURPOSE INDEXING
Groups all general and special types of cluster: ORGANIZATION
committees. author: Kim Waldén, Jean-Marc Nerson
keywords: committee, scientific board,
steering board

Class / (Cluster) Description

COMMITTEE General committee abstraction.

STEERING_COMMITTEE Committee in charge of practical arrangements.

PROGRAM_COMMITTEE Committee in charge of selecting technical contributions.

Figure 3.2 Cluster chart

• What services can other classes ask this class to provide? This translates
to commands applicable to the class.
• What rules must be obeyed by the class and its clients? This translates
into constraints on the class.
In this book, when we talk about something (like operations or constraints
above) being applicable to a class, we really mean applicable to the objects
whose behavior is described by the class. Since classes in BON are strictly
viewed as description, not as objects, there should be no risk of confusion.
An example of two class charts is given in figure 3.3. The class chart header
is similar to those of the system and cluster charts. The comment entry
(explicitly labeled “type of object”) contains a short description of the purpose of
34 THE STATIC MODEL—CLASSES AND CLUSTERS

CLASS CITIZEN Part: 1/1

TYPE OF OBJECT INDEXING

Person born or living in a country cluster: CIVIL_STATUS
created: 1993-03-15 jmn
revised: 1993-05-12 kw

Queries Name, Sex, Age, Single, Spouse, Children, Parents,

Impediment to marriage

Commands Marry. Divorce.

Constraints Each citizen has two parents.

At most one spouse allowed.
May not marry children or parents or person of same sex.
Spouse’s spouse must be this person.
All children, if any, must have this person among their parents.

CLASS NOBLEPERSON Part: 1/1

TYPE OF OBJECT INDEXING

Person of noble rank cluster: CIVIL_STATUS
created: 1993-03-15 jmn
revised: 1993-05-12 kw, 1993-12-10 kw

Inherits from CITIZEN

Queries Assets, Butler

Constraints Enough property for independence.

Can only marry other noble person.
Wedding celebrated with style.
Married nobility share their assets and must have a butler.

Figure 3.3 Class charts: types of citizen

the class. Keywords for version control have been added to the indexing clause,
since classes represent the evolving basic components of a system (keeping track
of changes to clusters is usually not very interesting). After the chart header a
number of dynamic entries follow, specified only when non-empty:
• Inherits from
– lists classes that are direct ancestors to this class.
• Queries
– lists applicable queries (value return; may not change system state).
• Commands
– lists applicable commands (no value return; may change system state).
 BON MODELING CHARTS 35

• Constraints
– lists consistency requirement of the class and its operations as well as
general business rules and other information that may affect the design
and implementation of the class.
The constraints may later be translated into formal assertions on the class (pre-
and postconditions and class invariants), but also serve as a kind of formatted
memo for the class. Some constraints at the early analysis level may record
information that will never become formal assertions, but rather serve as a guide
for future design decisions (see the Conference case study for some examples).
This is the reason we have chosen the name constraints for this field in the BON
class charts, rather than the more precise assertions.
This completes the description of the static untyped modeling charts in BON,
and we may proceed to the notation used for more exact specification. But
before we do, we will say a few words to clarify how the operations of a class
relate to the system state.

3.3 SYSTEM STATE INFORMATION

So far, we have deliberately refrained from discussing state variables (class
attributes in this book), since the internal state of an object is not part of its
interface. In some languages, notably Eiffel [Meyer 1992a], attributes may be
exported as read-only variables. However, this is just a convenient trick to
simplify language syntax and should not affect the way we think about the
visible operations of a class. Conceptually, an exported attribute should be
viewed as a function returning the value of some hidden state information that
just happens to have the same name.
The only visible part of an abstract data type is its operations with pre- and
postconditions and the class invariant. Whether it stores necessary history
information as internal variables or just magically remembers the past is none of
the client’s business as long as the data type behaves according to its software
contract.
In this book, we will use the term class feature or just feature to cover both the
operational aspects and the state aspects of class operations. In many
applications—particularly when persistent objects are involved—the designer
often has an early idea about which features should be implemented as state
variables, but there is no reason to make future change harder by letting the
design depend on these ideas. Often the decision whether to store a value as a
data attribute or instead compute it each time it is needed is only a question of
space / time tradeoff which should be postponed as long as possible. Therefore,
the BON notation does not include symbols for state variables.
36 THE STATIC MODEL—CLASSES AND CLUSTERS

3.4 TYPED CLASS INTERFACES

Using modeling charts may be a good way to start analysis, particularly when
non-technical people are involved, but the analyst/designer will soon need
something more expressive. This leads us to the main part of the BON notation:
the structured static diagrams with typing and software contracts.
Class interfaces will be described first; then we will go through the BON
assertion language in more detail. At the end of this chapter class clustering will
be discussed. The static relations are postponed until the next chapter.
A class interface consists of a number of sections, some of which may be
empty. The sections are:
• Class header
• Indexing clause
• Inheritance clause
• Class features
• Class invariant

Compression of interface sections

Class interfaces may be described either textually or graphically. Each graphical
descriptive element in BON has two representations: one expanded form and one
compressed form. With the exception of the class header, each section in a
graphical class interface (empty or not) may be compressed, which means it will
be hidden.
The whole class interface may also be compressed; it is then represented by its
header enclosed in an ellipse. We will see more of this at the end of the chapter,
when clustering and scaling is discussed.

Graphical representation
The graphical form of a class interface with all sections expanded is shown in
figure 3.4. The full interface may look a bit crowded, but in practice only a few
sections will be displayed at a time (we are assuming automated case tool
support here). The indexing section, for example, is usually hidden unless
explicitly requested. Note also that the restricted sections of a class interface (as
explained below) only come in during detailed design; the early phases should
always concentrate on public features.
We will now look at each interface section in turn, and show its specification
in both textual and graphical BON.
 CLASS HEADER 37

CLASS_NAME

Indexing information

Inherits: PARENT
CLASSES

Public features
A, B, C
Features only visible
to classes A, B, C
Invariant
Class invariant

Figure 3.4 Class interface: expanded sections

3.5 CLASS HEADER

The class header consists of the class name, which may be annotated to highlight
certain key properties of the class. The complete set of header annotations is
shown in figure 3.5.

3.6 INDEXING CLAUSE

The indexing clause contains general information about a class to be used for
browsing and configuration management purposes. It consists of a list of index
entries, and each index entry is a keyword followed by a list of words. The
semantic interpretation of keywords and the conventions to be used in creating
indexing information are decided per project. The indexing clause may for
example contain references to arbitrary text, graphics, and perhaps sound, and be
used as the starting point for an advanced multi-media class library
documentation system. A simple indexing clause is shown in figure 3.6.
What keywords to use for the classification of items via indexing clauses
depends strongly on the environment in which they are used. It is probably best
to start out with very terse information and gradually let each new keyword
justify its inclusion as an index entry when experience shows that it is really
useful. It is important that both users and providers of indexing clauses perceive
the contents as relevant information, not as yet another administrative burden.
38 THE STATIC MODEL—CLASSES AND CLUSTERS

CLASS HEADER ANNOTATIONS

Graphical form Textual form Explanation

Class is reused from a library of existing

NAME reused class NAME components.

●
NAME persistent class NAME Class instances are potentially persistent.

NAME
class NAME [G, H] Class is parameterized.
[G, H]

Class will not be fully implemented: it has no

* instances and is only used for classification
NAME deferred class NAME
and interface definition purposes.

Class is implementing the interface of

+
NAME effective class NAME a deferred class, or reimplementing
the interface of an ancestor.

▲
Class is interfaced with the outside world:
NAME interfaced class NAME some class operation encapsulates external
communication (function calls, data, etc.).

Class is a root class: its instances

NAME root class NAME may be created as separate processes.

Figure 3.5 Complete set of class header annotations

synonyms: car, transportation means

application_domains: car rental, car lease
author: John W. Smith
version: 1.1
revision_date: March 30, 1992
spec_refs: srs.1.3.3, srs.3.4.7
keywords: rental, agency, car, vehicle, automobile

Figure 3.6 Indexing clause for a VEHICLE class

 INHERITANCE CLAUSE 39

3.7 INHERITANCE CLAUSE

The inheritance clause lists all parents of the class, if any. It corresponds to the
“Inherits from” entry in the class modeling chart.

3.8 CLASS FEATURES

The features of a class are grouped into feature clauses, which may have
different degrees of visibility:
A. Public features, visible to any client class.
B. Restricted features, available only to a certain group of other classes. Such
features are often implementation oriented and therefore only exported to
classes whose implementations are allowed to depend on them.
Since different features may have different restrictions, there can be any
number of restricted feature clauses, each listing the set of classes
permitted to use them. Listing the single name NONE yields a private
feature clause, not visible to any other class.
Restricted sections are rarely used at the analysis level, but may come
into play when BON is used for detailed design and the implementation
language is known.

Feature names
Within each feature clause, the feature names are listed with optional markers
showing their implementation status, as illustrated by figure 3.7.

FEATURE NAMES
Graphical form Textual form Explanation

name * deferred name Non-implemented feature

name+ effective name Implemented feature
name++ redefined name Reimplemented feature

Figure 3.7 Feature implementation status

A feature may be marked as:

• deferred, which means the feature will not be implemented in this class,
but in some descendant class. This implies that the class is also deferred,
that is cannot have instances of its own, and is used exclusively through
inheritance.
40 THE STATIC MODEL—CLASSES AND CLUSTERS

• effective, which means the feature will be implemented in this class.

Usually this signifies that a deferred feature was inherited and given an
implementation, but the marker may also be used just to emphasize that
the feature will have an implementation.
• redefined, which means an effective feature is inherited from another
class, and the implementation changed.

Feature signatures
Class features are fully typed in BON. This means that the signature (number
and type of possible arguments and return value) is specified for each feature.
Figure 3.8 shows the symbols used.

FEATURE SIGNATURES (partial table)

Graphical form Textual form Explanation

name: TYPE name: TYPE Result type

– arg: TYPE −> arg: TYPE Input argument

Figure 3.8 Typed signatures

If a feature returns a value, its name is followed by the corresponding type

separated by a colon. (The table is partial, since it is also possible to express
variants of client relations, namely aggregations and shared results. These
additional signatures will be explained in the next chapter.) As an example, the
signature of a feature returning the seat number for a given passenger is shown
below:
seat_assignment: NUMERIC
– passenger: PERSON

Feature renaming
After the feature name with its possible return type declaration, there may follow
a rename clause which makes it possible to specify occasional feature renamings
that may occur when a class inherits from another class. The clause consists of a
pair of curly braces enclosing the name of the class from which the renamed
feature was inherited (not always obvious with multiple inheritance) and the old
feature name separated by a dot. The rename clause looks the same in both
graphical and textual form. An example is shown below, where the feature
co_pilot is inherited from class VEHICLE under the name co_driver and then
renamed.
 CLASS FEATURES 41

class AIRCRAFT inherit …

co_pilot++ : PILOT { ^VEHICLE.co_driver }
The feature was also redefined as indicated by the ++ symbol into returning the
type PILOT instead of DRIVER. However, the original return type is not shown
in the rename clause. Since redefinitions may affect the full signature of a
feature (including the types of input arguments), we must look at the interface of
class AIRCRAFT to find out more details. The caret of the rename clause
suggests that the names refer to entities “above this class”.
The next few lines of the feature specification may contain an optional header
comment (each line preceded by a double dash) giving a terse description of the
purpose of the feature. Then the input arguments follow, if any, each preceded
by a small arrow.

Assertions
Finally, after the possible arguments, an optional precondition and an optional
postcondition follow in turn. The notation used for assertions is shown in
figure 3.9.

ASSERTIONS
Graphical form Textual form

? routine preconditions require

routine preconditions

! routine postconditions ensure

routine postconditions
[written in invariant section] invariant
class invariant class invariant

Figure 3.9 Software contracting clauses

The precondition states a predicate that must be true when the feature is called
by a client. It is the client’s responsibility to ensure that the precondition is
indeed fulfilled before calling a supplier.
The postcondition states a predicate that must be true when the feature has
been executed and the supplier object returns control to the client. Given that the
precondition is true on feature entry, it is the supplier’s responsibility to ensure
that the postcondition is true before returning control to the client.
The laws of software contracting [Meyer 1992c] restrict the semantics of
descendant classes so as not to violate the abstraction carefully crafted by the
42 THE STATIC MODEL—CLASSES AND CLUSTERS

ancestors. For example, it may be good practice for a RECTANGLE class to

redefine the implementation of an area feature inherited from a general class
POLYGON to obtain better efficiency, but it would be a disaster if the redefined
feature suddenly started to return the circumference instead. The laws of
contracting reduce this risk by stipulating the following:
1. The class invariants of all ancestors must also be obeyed by all children.
2. Inherited preconditions may only be weakened. This means a new
implementation may never restrict the circumstances under which a client
call is valid beyond what has been specified by the ancestor. It may,
however, relax the rules and also accept calls that would have been
rejected by the ancestor.
3. Inherited postconditions may only be strengthened. This means a new
implementation must fulfill what the ancestors have undertaken to do, but
may deliver more.
Contrary to some other environments, notably Smalltalk, the concepts of subtype
and subclass are viewed as identical in BON. Therefore, specification of types is
directly transferred to specification of the corresponding classes, which is why
the semantically strong contracting laws make sense. Coupled with these laws
BON assumes what is known as the covariant rule for inherited features, which
says that signatures of redefined features may only be of the same corresponding
types as before, or descendant types.1

Example
We will use the two classes described by the class charts in figure 3.3 as an
example, and give the corresponding formal specifications in both textual and
graphical form. The textual description is shown in figure 3.10 and figure 3.11,
and the graphical equivalent in figure 3.12.
The deferred class CITIZEN models a citizen in a country. The first three
features are queries returning the name, sex, and age of the current citizen object,
each of type VALUE, while the fourth feature returns a possible spouse of type
CITIZEN. The next features are also queries, children and parents, both
returning SET [CITIZEN] since there can be more than one of each attached
relative.

1
Some object-oriented notations and languages have instead adopted the contravariant rule, stat-
ing that signatures may only be redefined into ancestor types. This can lead to a simpler mathe-
matical model of the type system, but is in our experience much too restrictive for practical use in
large developments.
 CLASS FEATURES 43

deferred class CITIZEN

feature
name, sex, age: VALUE
spouse: CITIZEN − − Husband or wife
children, parents: SET [CITIZEN] − − Close relatives, if any
single: BOOLEAN − − Is this citizen single?
ensure
Result <−> spouse = Void
end
deferred marry − − Celebrate the wedding.
−> sweetheart: CITIZEN
require
sweetheart /= Void and can_marry (sweetheart)
ensure
spouse = sweetheart
end
can_marry: BOOLEAN − − No legal hindrance?
−> other: CITIZEN
require
other /= Void
ensure
Result −> (single and other.single
and other not member_of children
and other not member_of parents
and sex /= other.sex)
end
divorce − − Admit mistake.
require
not single
ensure
single and (old spouse).single
end
invariant
single or spouse.spouse = Current;
parents.count = 2;
for_all c member_of children it_holds
(exists p member_of c.parents it_holds p = Current)
end − − class CITIZEN

Figure 3.10 Formal specification using textual BON

Then a query follows whose BOOLEAN result tells whether the current citizen
is single or not. The semantics of the feature is specified through a
postcondition. The condition states that the return value of single will be true if
and only if spouse returns Void (no spouse object attached to current citizen).
Result is a predefined variable carrying the return value of a query. The symbols
↔ and ∅ stand for equivalence and void reference respectively (see figure 3.13).
The next public feature is marry, a deferred command (shown by an asterisk
in figure 3.12) that returns no value, but instead alters the object state. It requires
44 THE STATIC MODEL—CLASSES AND CLUSTERS

effective class NOBLEPERSON

inherit
CITIZEN
feature
assets: NUMERIC − − The bare necessities of life
butler: CITIZEN − − Irons the morning paper
redefined spouse: NOBLEPERSON − − Lord or Lady
effective marry − − Celebrate with style.
−> fiancee: NOBLEPERSON
ensure
butler /= Void;
assets <= old assets + fiancee.assets − $50 ,000
end
end − − class NOBLEPERSON

Figure 3.11 Formal specification (continued)

an input argument (marked by an arrow) also of type CITIZEN. The formal

argument is named sweetheart, so that it can be referenced in the semantic
specification of the feature.
The precondition states that marry may only be invoked if there is a
sweetheart available, and there is no impediment to marriage between the two
parties. The postcondition asserts that unless legal reasons forbid, execution of
the command marry will indeed attach sweetheart as the spouse of the current
CITIZEN object.
Legal hindrance is defined by the next query, can_marry, which rules out
bigamy and incest and allows only heterosexual unions. It may only be called on
a non-void citizen. BON assertion expressions use the common syntax o.f (a, b)
to mean invocation of feature f on object o with input arguments a and b.
The feature divorce, which comes next, is also a command. It requires the
citizen to be non-single, so there will be somebody to divorce. The
postcondition then ensures that after the divorce, both parties will indeed be
single again.
To express this, we use the special symbol old, which refers to the value an
expression would have returned, had it been evaluated just before calling the
feature. The parentheses are used to alter operator precedence, so the feature
single is applied to the old spouse object, but with the current system state (after
execution of divorce). Writing old spouse.single would mean applying single to
the old spouse object with the old system state (before execution of divorce),
which would return false since the citizens were then still married.
We save the invariant for the next section, and turn to the second interface—
that of the class NOBLEPERSON. The class is effective (header marked with
plus symbol in the graphical form) and the inheritance clause shows it is a child
 CLASS FEATURES 45

CITIZEN * NOBLEPERSON +

name, sex, age: VALUE Inherits: CITIZEN

spouse: CITIZEN
− − Husband or wife assets: NUMERIC
− − The bare necessities of life
children, parents: SET [CITIZEN]
− − Close relatives, if any butler: CITIZEN
− − Irons the morning paper
single: BOOLEAN
− − Is this citizen single? spouse++ : NOBLEPERSON
−− Lord or Lady
! Result ↔ spouse = ∅ marry+
marry * − − Celebrate with style.
Celebrate the wedding.
−− – fiancee: NOBLEPERSON
– sweetheart: CITIZEN
! butler ≠ ∅ ;
? sweetheart ≠ ∅ and assets ≤ old assets + fiancee.assets
can_marry (sweetheart) − $50 ,000

! spouse = sweetheart
can_marry: BOOLEAN
− − No legal hindrance?
– other: CITIZEN

? other ≠ ∅
! Result → (single and other.single
and other ∉ children
and other ∉ parents
and sex ≠ other.sex)
divorce
−− Admit mistake.
? ¬ single
! single and (old spouse).single

Invariant
single or spouse.spouse = @;
parents.count = 2;
∀ c ∈ children • (∃ p ∈ c.parents • p = @)

Figure 3.12 Equivalent specification using graphical BON

of CITIZEN. (In figure 3.12, this may also be seen from the single arrow which,
as we shall see in the next chapter, represents the inheritance relation.)
The first two features of NOBLEPERSON represent necessary extensions: an
assets feature of type NUMERIC (absolutely essential, considering the ridiculous
prices charged for good hunting grounds these days), and the obligatory
46 THE STATIC MODEL—CLASSES AND CLUSTERS

ASSERTION ELEMENTS
Graphical BON Textual BON Explanation

∆ name delta name Attribute changed

old expr old expr Old return value

Result Result Current query result

@ Current Current object
∅ Void Void reference

+−*/ +−*/ Basic numeric operators

^ ^ Power operator
// // Integer division
\\ \\ Modulo

= = Equal
≠ /= Not equal
< < Less than
≤ <= Less than or equal
> > Greater than
≥ >= Greater than or equal
→ −> Implies (semi-strict)
↔ <−> Equivalent to
¬ not Not
and and And (semi-strict)
or or Or (semi-strict)
xor xor Exclusive or

∃ exists There exists

∀ for_all For all
| such_that Such that
• it_holds It holds
∈ member_of Is in set
∉ not member_of Is not in set
: type : type Is of type
{} {} Enumerated set
.. .. Closed range

Figure 3.13 Elements of assertions

manservant. The third feature redefines the spouse query so it will now return
NOBLEPERSON, thus satisfying both tradition and the covariant rule. Finally,
the marry command is defined to reflect what is expected from a high-class
wedding. The signature is again changed, and the postcondition extended to
ensure that noble couples who link their destinies will not lack domestic support,
and that each party will have access to the accumulated fortune minus the
amount that must be spent to ensure a wedding with style.
 CLASS FEATURES 47

Evaluation order
Logical conjunction (and), disjunction (or), and implication (implies) have so-
called semi-strict interpretation in BON assertions. This means that the
corresponding binary expressions are assumed to be evaluated left to right, so
that if the left operand is true (in the case of or) or false (in the case of and or
implies) the result will be directly inferred without evaluating the right operand.
This makes the precondition of marry above meaningful also when sweetheart
is not attached to any object. With the standard interpretation of the logical
connectives, the precondition would not be fully specified since the expression
can_marry (sweetheart) is not defined for void arguments.
Therefore, some care must be taken when BON assertions are translated to
implementation languages. For example, in Eiffel and then and or else have to
be used, while in C++ the corresponding operators already have semi-strict
semantics. A language-specific case tool could allow individual expressions to
be marked as strict (in case they can be evaluated in any order), to allow for
optimized assertion checking.

3.9 CLASS INVARIANT

Before explaining the details of the invariant example, we would like to reassure
those readers who are likely to break out in a rash when exposed to anything
reminiscent of mathematical formulas: there is nothing in the BON method that
forces you to write formal specifications. The symbols in figure 3.13 are only
there to help you when you want to specify something exactly, and avoid the risk
of misunderstandings always present in natural language.
If you do want precise specifications, all usable notations tend to look a bit
scary at first, but once gotten used to they can provide enormous help and really
put the user in control. A very interesting example showing how difficult it can
be to express even very small everyday problems in natural language with
enough precision to allow correct and unambiguous implementation can be
found in [Meyer 1985]. However, during early analysis, the need for lengthy
formal statements is mostly quite limited.
With this in mind, we proceed with our social example. The class invariant
clause of class CITIZEN, which expresses consistency constraints for the whole
class as an abstraction, contains three statements. They must always be true both
before a visible feature is called by a client and after feature execution, just
before returning to the client.
The first invariant statement says that if you are a citizen, you must be either
single or else married to somebody who is also married to you (the @ symbol in
figure 3.12 stands for current object, cf. figure 3.13). The second statement
48 THE STATIC MODEL—CLASSES AND CLUSTERS

simply says a citizen has always exactly two parents (not necessarily alive, but
that does not alter the fact). The SET class is typically a reused library class with
well-known semantics and a feature count returning the current number of
elements.
The third statement is a trifle more complicated. It uses the familiar
quantification symbols from predicate logic, “for all” and “there exists” (see
figure 3.13), to assert that if you are a citizen with children, each one of them
must have you as one of their parents.
The interpretation is as follows. For each child c which is a member of the list
of children of this citizen, the statement inside parentheses holds. This
statement, in turn, says that there is a parent p among the members of the list of
parents to c, such that p is the current object (this citizen).

3.10 GRAPHICAL AND TEXTUAL SPECIFICATION

The graphical class interfaces are not to be viewed as yet another set of modeling
charts to be filled in slightly differently. While the untyped charts are mainly
used in the very early phases in communication with people from the problem
domain (where a certain amount of manual handling can often be an advantage),
the graphical forms clearly need case tool support. Some people, being more
used to plain text specifications as in figures 3.10 and 3.11, might find the
symbols used in figure 3.12 somewhat confusing at first.
However, the symbols are really just abbreviations of their textual
counterparts, and the result is a much more compact specification. The graphical
specification elements have been designed to be simple also to draw by hand for
whiteboard engineering sessions and when no case tool is available. When used
for sketching, the rounded boxes are probably not drawn and the different
sections written as indented blocks delimited by the class name and textual labels
like inherit and invariant.
Once a reader has become familiar with the graphical elements, this form can
be read much faster. Care has been taken to only introduce graphical symbols
where they really serve a good purpose, and to keep them as simple as possible.
The use of ? and ! for pre- and postconditions, respectively, may require
some explanation. Obviously, when new symbols are introduced it will never be
possible to please everyone, since the associations they trigger depend too much
on general background and personal taste. However, the most important point is
that they are simple and unambiguous enough to feel natural after some time of
usage.
One may object that the interrogative/imperative meaning of the two symbols
could just as well have been used to specify whether a feature is a query or a
command. This may be true, but since types of results and arguments are
 GRAPHICAL AND TEXTUAL SPECIFICATION 49

specified using the familiar colon separator which can hardly be misunderstood,
the reader will quickly learn that the framed symbols signal assertions and not
feature signatures. What is then important in the long run, is that it should not be
possible to confuse the assertion symbols with each other.2
The graphical assertion symbols should be read as follows. Before an
operation is executed it is checked against the software contract:
? cond − − “Does the client fulfill his part of the deal?”

! cond − − “If so, the following condition is guaranteed!”

With this interpretation in mind, there should be no risk of confusion. (Using the
interrogation metaphor is reminiscent of Hoare’s CSP language [Hoare 1978],
where the same characters signify input and output.)3
This concludes the general description of how to write formal class interface
specifications in BON. Before we proceed to discuss system scalability, we will
spend the next sections on a more detailed description of the language used for
specifying software contracts. Readers who find the semantic specification
aspects of modeling boring may proceed directly to section 3.14.

3.11 THE BON ASSERTION LANGUAGE

Assertions in BON are statements about properties of objects. Externally, such
properties show up as system behavior, but viewed from the inside of an object,
properties of other objects can only manifest themselves through return values
from queries to the objects. Since an assertion must have a truth value, only
queries that return objects of type BOOLEAN may be used directly as assertions.
However, non-boolean query results are still useful in two ways: as targets of
new queries, or as input arguments to queries on other objects. As long as the
last query returns a BOOLEAN, any type of object may be part of such query
expressions. This enables system properties to be captured indirectly by the
successive combination of query results from individual objects.

2
This was a problem with older versions of the BON notation, and readers who have seen earlier
examples [Nerson 1991, Nerson 1992b, Nerson 1992a] may note quite a few improvements and
general simplifications of the graphical class interfaces in this book.
3
The main source of inspiration, however, was some correspondence between the French novel-
ist Victor Hugo and his publisher, said to be the shortest in history. Hugo, who was away, wrote
a letter inquiring about how his latest book Les Misérables had been received by the public.
When the publisher opened the envelope, he found a single question mark written across the
page. Since the book had been a great success, the reply was equally economic—a single excla-
mation point.
50 THE STATIC MODEL—CLASSES AND CLUSTERS

A number of specification elements are available for use in pre- and

postconditions and class invariants to express semantic properties of class
behavior. We will now return to figure 3.13 where the complete list is found,
and look at each group of related elements in order.

State changes
The first group in this table contains two special symbols that can only occur in
postconditions of command features. The delta symbol indicates that execution
of the command may lead to a state change affecting the feature name. This
means that the next time name is invoked, it may return another value than it
would, had it been called just before the current command was executed.
Name is thus assumed to be a function returning a value representing an
interesting state for the object that has the feature. Whether it actually
corresponds to a state variable or not is left to implementation, so the modified
state may very well be indirect through changes to several concealed state
variables.
The old symbol refers to the value of expr just before calling the command
(old captures the whole system state as it was when the call was made). Using
old makes it possible to specify how the values returned by functions may
change as a result of executing a command. Most often old is used to quantify
changes in abstract attributes. For example, count = old count + 1 expresses that
count was increased by one.

Basic object expressions

The second group contains symbols referring to the current result to be returned
by a query, the current object, and the special value for void reference. The last
two are central concepts of the object-oriented execution model. Any assertion
starts by querying the current object, or one or more possible input argument
objects in the case of pre- and postconditions. We also frequently need to
express whether an object is attached to a reference or not. The Result symbol is
needed to express postconditions in queries.
Since an unqualified query implicitly refers to the current object, no special
symbol is needed for such a call, but it is needed to supply the current object as
input argument to a query on some other object.
Constant values of basic types, such as 17, 3.14, false, ’C’, and "Fahrenheit
451", are also directly available. Each value represents an immutable object
accessible by a function whose name is precisely the value itself (expressed by
the notation above). So even if a constant occurs in several places, it is always
attached to the same object; there are no copies of basic values. (We may think
of the basic value functions as inherited from a top-level class ANY.)
 THE BON ASSERTION LANGUAGE 51

We have chosen INTEGER, REAL, BOOLEAN, CHARACTER, and STRING

as default basic value types. This set may be changed depending on
implementation language, and BON value types need not be represented as such.
For efficiency, the most basic string type is nearly always implemented as a
reference type, and defined by a special system class.

Numeric expressions
Objects of numeric type may be combined by the usual numeric operators, which
are in the third table segment. These (as well as the relational and boolean
operators in the following segments) are prefix and infix operators for ease of
expression, but using them is really no different from any query invocation. For
example,
a+5*b
should be thought of as
a.plus (5.times (b))
Adopting this view makes the object-oriented model extremely simple, since
almost everything can be expressed as feature invocation in the end. (The
implications of this are discussed further in the chapter on dynamic modeling.)

Relational expressions
The fourth table segment contains the standard relational operators. The results
of two queries may be combined using the = or ≠ operator to express that the
corresponding objects are, or are not, identical.4 If the return types of the queries
have a common ancestor defining an order relation, the results may also be
combined using the <, ≤, >, ≥ operators.

Basic boolean expressions

In the second last table segment, we find the standard logical operators for
building boolean expressions. As we have explained earlier, implication is semi-
strict, meaning that the implied part (the consequent) in P → Q will only be
evaluated if P (the antecedent) is true. If P is false, the result of the implication
is considered false regardless of Q. This is important for dealing with partial
functions.

4
Since objects may have reference type or value type, the precise semantics of the equality and
non-equality operators is allowed to depend on the target language.
52 THE STATIC MODEL—CLASSES AND CLUSTERS

We have chosen the symbols → and ↔ for logical implication and

equivalence, respectively, rather than ⇒ and ⇔. There are two main reasons for
this. First, in order to use only standard characters for portability, textual BON
requires multi-character operators. Combining => and <=> with the relational
operators <=, >=, =, /= in the same boolean expressions makes the result very
hard to read. Choosing single arrows avoids this confusion.
Second, in formal logic, the single arrow forms are more commonly used as
boolean connectives when building a single proposition, while the double arrows
are used when proving theorems from chains of propositions.
This leaves the last table segment containing the basic elements of first-order
predicate logic used in BON to reason about sets of objects.

3.12 THE BON PREDICATE LOGIC

Basic boolean expressions are easy to translate into efficient code, so assertions
built from them can be made optionally executable for testing purposes.
Contract violations may then trigger exception handlers, which may be used for
precise error reporting and recovery. However, the semantics that can be
conveyed with only basic boolean expressions is limited.
For lower-level implementation details, it goes a surprisingly long way and
may raise the level of correctness and understandability of a system significantly.
Replacing so-called defensive programming by clearly defined responsibilities
and a large number of executable assertions is perhaps the most generally
efficient way available today for reducing the cost of test and error repair. We
will discuss this issue in more detail in section 8.6.
But for analysis and design we need to go a step further. Whether it will ever
be possible to find a purely formal method to be applied by the average software
engineer to mainstream systems development projects remains to be seen.
Instead, we are trying to find a reasonable balance between simplicity and power
of expression. This is no easy task, since success depends very much on what
turns out to be useful in practice.
The possibility to use queries on other objects is crucial, since this is what
enables the recursive specification necessary for large systems. It also permits
dynamic extension of the specification language to support various application
areas. Classes (often generic) expressing fundamental semantic properties about
a problem domain then act as new specification elements.
The basic specification notation may of course also be extended, and more
experience from use in real systems will guide the future direction of the BON
assertion language. So far, we have taken a conservative approach and only
added the predicate logic extensions described in this section. However,
predicate logic can be used to form boolean expressions involving properties of
 THE BON PREDICATE LOGIC 53

object groups ranging over sets, which is significantly more than is possible with
only propositional logic.

Set operations
The sets used in assertions are usually standard library container classes, such as
SET, SEQUENCE, TABLE, and STACK. We assume they all contain a boolean
query returning whether a given object is an element in the set or not. This
operation and its inverse are defined by the infix symbols ∈ and ∉ (pronounced
“is in” and “is not in”) as seen in figure 3.13. (Although a client is only allowed
to remove the top element from a STACK container, the query elem ∈ my_stack
is still supposed to return true if elem is one of the hidden elements.)
Besides querying objects to obtain references to sets, there is also a way to
construct new sets by direct enumeration of elements enclosed in braces. The
“..” operator may be used to specify intervals for the two basic types INTEGER
and CHARACTER. For example,
i ∈ {2, 5, 8..13}
char ∈ { ’a’.. ’z’}
primary_color ∈ {"red " , "green " , "blue " }
nephew ∈ {huey, dewey, louie}
assert that the object referenced by i is one of the integers 2, 5, 8, 9, 10, 11, 12,
13; that char refers to a lower case character; that primary_color is attached to
one of the string constants "red", "green", or "blue"; and that nephew refers to
one of the objects huey, dewey, or louie.

Quantifiers
Using quantifiers and groups of variables, each ranging over a set of objects, we
may express properties that must hold for the object groups referred to by the
variables. There are two types of quantified statements: universal and
existential.
• Universal quantification asserts that every combination of the quantified
variables satisfying the conditions of a range expression will also satisfy a
given proposition.
• Existential quantification asserts that at least one combination of the
quantified variables satisfying the conditions of a range expression will
also satisfy a given proposition.
The initial quantifier ∀ or ∃ (pronounced “for all” and “there exists”) tells which
type of quantification we are dealing with. A general quantified statement has
54 THE STATIC MODEL—CLASSES AND CLUSTERS

the following abstract syntax (extracted from the grammar in appendix A):

Quantification ∆ Quantifier Range_expression [ Restriction ] Proposition

=

The quantifier is followed by a range expression which selects an initial set of

objects (or set of object groups, if there is more than one formal variable
present). Next an optional restriction follows, which may be used to select a
subset of the initial set, and finally comes the proposition. If the quantifier is ∀,
the quantification states that the proposition is true for all elements of the
resulting set. If the quantifier is ∃, it means instead that the proposition is true
for at least one of the elements.

Quantified variables
There are two types of specification elements that can be used in a range
expression: member ranges and type ranges. A type range states that each
variable in a list of formal variables must be of a given type, while a member
range states that the variables must all belong to a given set of objects. Below
are some examples:
∀ v: VEHICLE [Restriction ] Proposition
∀ c ∈ children [Restriction ] Proposition
∃ x, y: REAL [Restriction ] Proposition
The first ranges over all objects of type VEHICLE, the second over all children
in a set, and the third over all pairs of type REAL. The expression children (a
query on the current object) is assumed to return a container type whose
semantics defines the ∈ operator for the corresponding data structure.
Notice that the elements of the set denoted by a range expression are a
combination of one object for each of the variables specified. Thus, if the formal
variables are v1, v2, v3 ranging over the corresponding sets R1, R2, R3, the
proposition will apply to tuples in the cartesian product set R1 × R2 × R3.
Several range specifications may be combined in a range expression separated
by a semicolon, and the two types above may be combined. For example,
∃ c ∈ children; b ∈ baby_sitters [Restriction ] Proposition
∀ b: BOSS; o ∈ clerks [Restriction ] Proposition
talks about the existence of a combination of one child and one baby sitter, and
all combinations of one BOSS object and one object from the set clerks.
Finally, if the range part is omitted for a variable list, this is interpreted as an
abbreviation for type ANY. Thus, the first two and the last two of the following
assertions are equivalent:
 THE BON PREDICATE LOGIC 55

∀ x, y [Restriction ] Proposition
∀ x, y: ANY [Restriction ] Proposition
∀ a, r: REAL [Restriction ] Proposition
∀ a; r: REAL [Restriction ] Proposition
∀ a: ANY; r: REAL [Restriction ] Proposition

Propositions
The proposition clause consists of the symbol • (pronounced “it holds”)
followed by a boolean expression. It provides the condition that must be true for
all of the selected elements, or for at least one, depending on the quantifier.
In the invariant of a class representing a table of ELEM objects which are
inserted and retrieved by a key value of type KEY, the following are possible
(equivalent) assertions:
∀ e ∈ @ • (∃ k: KEY • item (k) = e)
∀ e: ELEM • e ∈ @ → (∃ k: KEY • item (k) = e)
They both state that each element in the table must have a key associated with it.
(The @ symbol signifying current object is used here to refer to the set
represented by the container class itself.)
The alert reader will have noted that the ∈ symbol has been used in two
slightly different ways. In the propositional part, ∈ simply means the infix form
of the “is in” query that must be defined for any container type participating in
BON assertions. Since the query returns a boolean result, it may be used like
any other boolean expression to build successively more complex propositions.
However, the notational elements of the primary range selection (just after the
quantifiers ∀ and ∃) are not boolean expressions. Therefore, in this context the
symbol ∈ has a different meaning and is instead pronounced “in” (the
corresponding type range symbol ":" is pronounced “of type”). This simplifies a
large number of standard assertions considerably. Instead of being forced to
write
∀ p: T • p ∈S → P
where the declaration of T is superfluous (since it is already implicit in the type
of p) we may simply write
∀ p ∈S •P
By analogy with the above, we also allow the range symbol ":" to be used as a
boolean operator in the propositional part (then pronounced “is of type”) to
express whether an object is of a given type or not. Thus a: T will return true if
the (dynamic) type of a is T and false if not.
56 THE STATIC MODEL—CLASSES AND CLUSTERS

This makes it possible to specify different things about objects returned by

queries depending on the dynamic type of the result, which is often necessary
when reasoning about polymorphic structures. For example, if the class
NIGHT_CLUB keeps its guest record as a SET [PERSON], but has different
admission rules for objects of type MAN and WOMAN, this can be captured by
the class invariant:
∀ g ∈ guests •
g: FEMALE → g.age ≥ 20 ;
g: MALE → g.age ≥ 25
Following the general object-oriented rule to search for dynamic binding, we
could instead have introduced a deferred feature is_old_enough_for_night_clubs
in a common ancestor class PERSON, and let MAN and WOMAN provide their
own versions. Then the simpler invariant
∀ g ∈ guests • g.is_old_enough_for_night_clubs
would suffice. However, this is generally not a good idea when the distributed
concept (as in this case, night club age limit) has nothing to do with the basic
abstractions, but instead reflects specific client usage, that is a possible supplier
role.

Range restriction
The optional restriction clause consists of the symbol | (pronounced “such
that”) followed by a boolean expression. It reduces the initial set of object tuples
by a subselection according to a boolean condition on the formal variables.
Strictly, the restriction clause is not necessary, since each assertion
Quantifier Primary_range | C •P
can be rewritten as
Quantifier Primary_range •C→P
However, we still think it is important to add a restriction construct, because it
simplifies many assertions and encourages a more natural way of thinking about
the propositions. Consider the assertions in figure 3.14. All six are logically
equivalent statements, but the syntactic groupings express quite different lines of
thought, which has a profound effect on human readability.
In the context of animals being transported together, the first two assertions
are statements about lions, the next two about hungry lions, and the last two
about lion−zebra combinations. Depending on where the focus of interest lies,
we may prefer one of these views to the others.
 THE BON PREDICATE LOGIC 57

Alternative views on joint transportation

∀ a, b ∈ animal_transport •
a: LION → ((a.hungry and b: ZEBRA) → a.eats (b))
∀ a, b ∈ animal_transport | a: LION •
(a.hungry and b: ZEBRA) → a.eats (b)

∀ a, b ∈ animal_transport •
(a: LION and a.hungry) → (b: ZEBRA → a.eats (b))
∀ a, b ∈ animal_transport | a: LION and a.hungry •
b: ZEBRA → a.eats (b)

∀ a, b ∈ animal_transport •
(a: LION and b: ZEBRA) → (a.hungry → a.eats (b))
∀ a, b ∈ animal_transport | a: LION and b: ZEBRA •
a.hungry → a.eats (b)

Figure 3.14 Equivalent assertions

In each of the three cases above, the form using a restriction clause is more
readable. There are two main reasons for this. First, it requires less parentheses
and avoids nested implications, which are always hard to understand. Second, as
soon as the restriction clause has been read, the proposition can be interpreted in
the new local context. This permits us to deal with one thing at a time as
opposed to unrestricted assertions, which must be understood as a whole.

3.13 ASSERTIONS AND PARTIAL FUNCTIONS

In the specification examples so far, we have implicitly assumed that all queries
used in logical expressions always return well-defined values. However, this is
not automatically the case, since we often need queries that are partial functions
to return interesting system states. Such queries return meaningful values only
when the corresponding preconditions are true.
Therefore, we must ensure that no precondition violation can occur when
queries are evaluated as part of an assertion. (We are not talking about
imperative execution here, only the logical evaluation of query expressions
needed to establish the meaning of an assertion.) Any precondition violation
would yield an undefined query result, making the assertion lose its truth value
for the corresponding object values.
For example, one might find the following assertion natural for expressing the
existence of at least one pair of matching elements with the same integer index in
two arrays A and B:
58 THE STATIC MODEL—CLASSES AND CLUSTERS

∃ i: INTEGER • A.item (i) = B.item (i)

However, since the quantification is only bound to integers in general, the
assertion has no meaning for indexes that fall outside the range of either A or B.
So we have to use a more restricted quantification ensuring that the values we
talk about are within the range of both data structures:
∃ i: INTEGER | A.has_index (i) and B.has_index (i) •
A.item (i) = B.item (i)
The reason why implication is defined as semi-strict in BON is the possibility to
use the operator as a kind of “if-statement” in a proposition. For example,
∀ p: PERSON • p ≠ Void → p.name ≠ Void
states that whenever a PERSON object is attached to p, that person must have a
name. With the usual definition of P → Q, the above assertion would not be
well defined. For similar reasons (as already explained in 3.8), the connectives
and and or are also semi-strict in BON, permitting assertions like:
p ≠ Void and p.name ≠ Void
p = Void or p.name ≠ Void
There are a number of possible approaches for dealing with non-denoting terms
that may appear as a result of partial function evaluation [Farmer 1990,
Parnas 1993], but they all introduce other complexities that have to be dealt with
instead. A discussion of these falls beyond the scope of this book. For an
introduction to the logic of programming and the semantics of programming
languages see [Hehner 1984, Meyer 1990a].
Although including the restrictions needed to guard against precondition
violations generally leads to longer assertions than what is possible with some of
the alternative semantic models, this is less of a problem in an object-oriented
environment.
First, more high-level features will generally be available to express the
restrictions more compactly. In a traditional language environment, the second
last assertion (transcribed into object notation) would more likely look like:
∃ i: INTEGER | A.min ≤ i and i ≤ A.max and B.min ≤ i and i ≤ B.max •
A.item (i) = B.item (i)
Second, a formal notation no matter how cleverly designed can only be used in
practice to specify a small subset of all interesting system properties. The real
leverage will come from successively extending the specification language with
elements in the form of queries to problem domain classes. Only the object-
oriented extensibility provides enough flexibility to express exactly what we
 ASSERTIONS AND PARTIAL FUNCTIONS 59

want; formal languages with comparable expressive power will be too complex
for routine use in the software industry.

3.14 SCALABLE SYSTEM DESCRIPTION

We now have a means of specifying the basic components in our system—the
classes—which captures both the syntax and semantics of the services they
provide. The rest of this chapter shows how the class components can be
grouped recursively into higher-level structures using clustering.
The possibility to compress and expand the graphical descriptive elements of
BON addresses the scaling and partitioning problems discussed earlier, which
always arise when large system architectures are modeled graphically. By
selecting a mixture of compressed and expanded forms, many different views
can be obtained at various levels of detail.

Levels of description
There are three basic levels of static description in BON:
• system level
• cluster level
• class level
The system level shows a set of clusters. The cluster level shows a set of classes
and possibly other clusters. Finally, the class level shows class interfaces with
their operations and contracts. The system and cluster levels may each comprise
several levels of nesting. For example, a large system may first be partitioned
into a number of subsystems (which will be clusters) and each subsystem may
again have subsystems (more clusters). As long as we only show clusters we are
still at the system level. When class headers start to appear we are at the cluster
level, and this continues until we show class interfaces. At that point we have
reached the class level.
Nothing prevents us from mixing the levels of detail in a view by showing,
say, a cluster with only a few class headers expanded into class interfaces. In
that case, the description is said to be at the lowest level contained in the view.

3.15 CLASS HEADERS

Building a static diagram often begins with the identification of a number of
classes and some relations between them, so class headers play an important role
in BON.
60 THE STATIC MODEL—CLASSES AND CLUSTERS

Graphical representation
A class header is represented graphically by an ellipse. The class name is an
alphanumeric string with possible underscores, whose first character is
alphabetic. The name, with possible annotations, is centered inside the ellipse
(upper case, bold italic font recommended).
The class header annotations translate to top annotations and bottom
annotations. Top annotations appear above the class name and bottom
annotations below the name inside the ellipse. The complete set of header
annotations was shown in figure 3.5. Figure 3.15 contains a set of annotated
class headers illustrating various uses and combinations of the class header
annotations (or just header annotations for short).

● ▲
CONTROL_PANEL TRANSACTION MAILER

HASH_TABLE * +
FLYING_OBJECT HELICOPTER
[T, U]

▲ VECTOR *▲
INPUT SESSION
[G]

Figure 3.15 Examples of annotated class headers

Compression of classes
A class header is the compressed form of a class interface.

3.16 CLUSTERS
A cluster represents a group of related classes (and possibly other clusters)
according to a selected point of view. Classes may be grouped differently
depending on the particular characteristics one wants to highlight. Possible
criteria are:
• subsystem functionality
• user categories
• application-domain-dependent factors
• reuse versus new development
• hardware or software platforms
 CLUSTERS 61

• license restrictions versus public domain

• abstraction level
• cohesion grouping or factorization
A class could for example be at the same time a user interface class, a reused
class, and a class supporting the X window platform. Such a class would then
qualify for membership in at least three different clusters. However, to avoid
ambiguity and confusion we need to lay down some rules for clustering.

System views
By a system view we mean a partitioning of the classes of a system into a set of
clusters, such that each class belongs to exactly one cluster. This does not mean
we have to use only one type of classification criteria for the clustering; some
clusters may represent subsystems, others general functionality, and still others
just groups of classes with the same ancestor. But for each individual class, one
of the possible groupings must be selected.
There may be several system views defined for a given system, but in practice
only one is used during analysis and design. Since the cluster structure will
evolve gradually as system development progresses, it usually becomes too
complicated to support several views simultaneously. However, if (as should
always be our goal) part of the classes developed in a software project are later
refined and merged with a general class library, their classification will probably
be quite different in that context.

Nested clusters
Besides classes, clusters may also contain other clusters. This makes it possible
to collect clusters which are only used locally, and group them with their client
classes. It also gives a facility for nesting subsystems of a large system to any
desired depth. In fact, very large systems will often have several layers of
clusters representing subsystems before any classes appear.
The cluster nesting is part of the system view. This means that not only
different partitioning of the same set of classes, but also different nesting of the
same set of clusters, will lead to different views. As modeling proceeds, patterns
are often detected that will cause the developers to restructure sets of existing
classes in order to better understand their roles.
Notice that the system (defined by a set of classes and its root class) and the
system view (defined by grouping the classes into possibly nested clusters) are
independent concepts. Changing the clustering does not affect the system, only
how we view it, while changing a class leads to a new system but does not affect
the clustering structure.
62 THE STATIC MODEL—CLASSES AND CLUSTERS

Graphical representation
A cluster is drawn as a box with rounded corners representing the cluster body
with the name in a separate rounded box (the cluster tag). The tag is positioned
just outside the cluster body touching its borderline. Cluster names follow the
class name convention: alphanumeric strings with possible underscores, starting
with an alphabetic character. Underlining the name signifies a reused cluster, in
analogy with the class header notation. The name is centered inside the tag, and
upper case roman font is recommended to make the label differ from class
names. A cluster tag may appear anywhere along the cluster borderline to
facilitate the presentation of complicated diagrams. An example of a data
structure cluster with two other clusters nested inside (one reused and one newly
developed) is shown in figure 3.16.
The line style of clusters must be different from the one chosen for classes.
The recommended practice is to use continuous borderlines for class interfaces
and class headers and dashed borderlines for cluster bodies and cluster tags.
BON was designed with case tool support in mind, and recommendations like
the one above are mainly directed to implementors of software generating BON
graphical diagrams. When used on paper or whiteboard, the fact that it is much
easier to draw continuous lines by hand may take precedence.

Layered clustering—not hierarchical development

Since every class belongs to exactly one cluster in a system view, the resulting
cluster structure is a hierarchy. However, this does not mean that BON favors
top-down hierarchical development in any way. On the contrary, such an
approach to object-oriented analysis and design suffers from many of the worst
disadvantages of the traditional methods. It forces the developer to make
premature decisions with far-reaching consequences very early in the project,
when problem understanding is still vague and many of the central abstractions
have not yet crystallized.
Instead, the overwhelming initial complexity may be gradually reduced by
successively modeling parts of the system that we do understand (and there are
nearly always such parts). When good abstractions have been found for these
areas, what is left will be somewhat easier to grasp thus allowing for new
patterns to be recognized and abstracted. Of course, we should not postpone the
hard parts, since this may give a false sense of rapid progress and later lead to
very unpleasant surprises. The general idea is to divide and conquer. We will
still discover errors at times and have to redo things, but the effects will usually
be more local and less catastrophic.
When more and more of the abstractions (classes) are in place, we may
gradually discover that some of them are only used in certain subparts of the
 CLUSTERS 63

DATA_STRUCTURES

SORTING

LINEAR_SORT QUICKSORT RADIX_SORT

INSERTION_SORT SHELLSORT HEAPSORT

TOPOLOGICAL_SORT SORT_MERGE

*
GRAPH WEIGHTED_GRAPH

DIRECTED_GRAPH UNDIRECTED_GRAPH

BIPARTITE_GRAPH DENSE_GRAPH

GRAPHS

Figure 3.16 A nested data structure cluster

system and reflect this by nested clustering. However, the potential reusability
of the classes of inner clusters is not affected by such a hierarchy. If we later
find more global use for them, we simply move the corresponding clusters out of
their local clusters. The cluster structure just serves to document how the classes
are used in this system, not how they could be reused in other systems.

Local naming
Clusters may correspond to different conceptual domains, so it is quite natural
that two classes in different clusters could have the same name but model
entirely different concepts.
64 THE STATIC MODEL—CLASSES AND CLUSTERS

For example, in an administrative system for a vocational school the class

BOARD could appear in the SCHOOL_ADMINISTRATION cluster and in the
MATERIALS cluster used in the carpentry classes. There is really no reason to
prohibit such name clashes during analysis and design, since they are easily
resolved by the context in which they appear.
Similarly, clusters which are nested inside different clusters may also be
permitted to have the same name. When the design is eventually turned into
implementation, the name clashes must be resolved through renaming or
qualification schemes in the object-oriented languages used, but this is merely a
syntactic problem.
To maintain a clear typographical difference between type names and role
names (type usage), we will consistently use upper case for class names and
lower case for class features in this book. A set of recommended naming rules
are discussed in section 8.4.

Compression of clusters
A cluster is compressed into its tag attached to an empty, iconized body. When a
cluster icon is expanded (or opened), its constituent parts first become visible in
compressed form, thus showing class headers and cluster icons (see figure 3.17).
The inner parts may then be opened recursively, until we reach the class
interfaces.

BEACH_SURVEILLANCE

EQUIPMENT
LIFE_GUARD

WATCHTOWER

Figure 3.17 An iconized cluster

4 Static relations

This chapter describes the relations between classes and clusters in the static
model. Only two kinds of static relations are needed in object-oriented systems,
inheritance relations and client relations. By combining them in various ways
and letting the resulting network of classes be guided by well-defined software
contracts, almost any type of semantic modeling needed can be achieved.
Moreover, we will see how the system scalability can be further refined by
extending the compression/expansion facilities to also include relations.

4.1 INHERITANCE RELATIONS

A class may inherit from one class (single inheritance), from several classes
(multiple inheritance), or several times from the same class (repeated
inheritance). Inheritance is simply defined as the inclusion in a class, called the
child, of operations and contract elements defined in other classes, its parents. A
class that is either parent or grandparent (recursively) of a class is called an
ancestor of the class. Similarly, a class that is either child or grandchild
(recursively) of a class is called a descendant of the class.
Inheritance may translate differently depending on the object-oriented
language used, so the definition is kept very general in BON. A special form of
complementary documentation called semantic links may be attached to
inheritance relations to show the designer’s intent in each case. These are
described in a later section.

Graphical representation
An inheritance relation is represented by a single arrow pointing from the child
to its parent, called an inheritance link. Inheritance links may be broken to avoid
crossing other elements, but the recommended practice for case tools is to only
use horizontal and vertical directions. A set of inheritance links representing
several children of the same class (or several parents of the same class) may be

65
66 STATIC RELATIONS

joined to simplify a diagram. In the case of direct repeated inheritance, instead

of drawing n inheritance links, one may use one link with a marker attached
(small lozenge) containing the count.
The diagrams in figure 4.1 illustrate different types of graphical inheritance
representation. From left to right and top to bottom it shows: single inheritance
using horizontal and vertical links, direct repeated inheritance, multiple
inheritance combined with indirect repeated inheritance, and single inheritance
with joined links. (Direct repeated inheritance makes sense if the multiply
inherited features are separated through renaming.)

*
COLD_STORE FREEZER INDEX FLYING_OBJECT

+ +
REFRIGERATOR TRIPLE_INDEX AIRCRAFT ROCKET

*
VEHICLE SPACE_SHUTTLE

+ + +
BICYCLE BOAT CAR

Figure 4.1 Different type of inheritance

Inheritance involving clusters

The inheritance relation can be generalized to apply also to clusters:
1. If all elements (classes or clusters) in a cluster X inherit (directly or
indirectly) from an element A outside of X, then the cluster X is said to
inherit from A. All direct inheritance links from elements in X to A can
then be compressed into one inheritance link from X to A.
2. If an element A outside a cluster Y inherits (directly or indirectly) from all
elements in Y, then A is said to inherit from Y. All direct inheritance links
from A to elements in Y can then be compressed into one link from A to Y.
 INHERITANCE RELATIONS 67

3. An inheritance link between two elements can be compressed into being

hidden (not shown in a diagram).
Rules 1 and 2 state that in order for a cluster to be child or parent of another
element, all elements inside the cluster must have a corresponding relation to this
element. (We shall see later that the corresponding rules for client relations only
require that some enclosed element is involved in the relation. This is not the
only possible convention, but it is the one we have found most useful in practice.
We will give some more justification for it in the discussion on client relations.)
Rule 3 makes further simplification possible by presenting less information.
Notice that we used the word element instead of class in rules 1 and 2 above to
allow for recursive compression. This is essential already for large inheritance
structures, but will become even more important for client relations. In fact,
recursive compression is the central key to achieving system scalability. Three
examples are shown in figure 4.2.

A A PARENTS

B C

B C B C D E

D E D E

CHILDREN CHILDREN F

Figure 4.2 Compressed inheritance relations

The leftmost diagram shows a cluster after application of rule 1. We can infer
that the classes B, C, D, and E all inherit either directly or indirectly from class
A, but we cannot strictly tell which inheritance is direct. For example, although
B appears to be a top-level class in the cluster (implying direct inheritance from
A), it need not be. It could be a child of E with the inheritance link hidden
through application of rule 3. However, in practice the developer knows enough
about the system structure to rule out such esoteric possibilities, and then
diagrams containing a large number of inheritance links can often be greatly
simplified without losing information (see 5.17).
The reason rules 1 and 2 only talk about replacement of direct inheritance
links is that we often want to keep local inheritance visible also after applying
one of these compressions, since it conveys additional information. To hide
more, we just continue the compression process by applying rule 3 a number of
68 STATIC RELATIONS

times, as illustrated by the middle diagram in figure 4.2. The rightmost diagram,
at last, shows an application of rule 2.
We do not need to define inheritance between two clusters separately. When
such a relation exists, it can be obtained by repeated application of rules 1 and 2
as shown in figure 4.3. From the initial view (upper left), we may either apply
rule 1 twice to get the upper right diagram, or apply rule 2 three times to get the
lower left diagram.

PARENTS PARENTS

A B A B

C D E C D E

CHILDREN CHILDREN

PARENTS PARENTS

A B A B

C D E C D E

CHILDREN CHILDREN

Figure 4.3 Recursive compression

If we choose the first alternative, applying rule 2 once more yields the final
result (lower right). Conversely, if we choose the second alternative, applying
rule 1 will yield the same result. In practice, a cluster inheriting from a class
occurs often, a class inheriting from a cluster is less common, and a cluster
inheriting from a cluster is rare.
 INHERITANCE RELATIONS 69

Compression of inheritance relations

Case tools may employ different strategies for the compression of inheritance
relations, as long as they comply with rules 1−3 above. For example, it may be
advantageous to hide all inheritance relations to certain groups of classes in order
to emphasize some overall aspect of the system. Typical candidates for
exclusion are general support classes, but also application classes considered less
important in the context studied.

4.2 CLIENT RELATIONS

A client / supplier relationship, or client relation for short, between a client class
A and a supplier class B means that A uses services supplied by B. This leads to
a class dependency between A and B, which means that if the specification of B
is changed, A may have to be changed too (its specification or implementation or
both). However, it also leads to a potential object dependency between instances
of A and B (or of descendants of A and B) when the system is executed. We will
return to the distinction between class and object dependencies in section 4.3.

Static relations between classes

It has become popular (and sometimes profitable) in recent years to compile
superficial articles and industry reports comparing various methods for object-
oriented analysis and design, which result in checkboxes with answers to the
questions: “Does notation X support concept Y?” The underlying assumption
seems to be that whenever a concept can be identified that may have an influence
on system behavior, being able to express that concept in an analysis and design
notation increases the power of the method. The BON approach disagrees with
this attitude.
Analysis and design is all about abstraction. We want to get away from details
that we do not necessarily need to keep track of, and leave them to underlying
structuring mechanisms and automatic support. This is why we group objects
into classes and classes into inheritance networks. In BON, contrary to many
other methods, there is no direct notation for expressing static dependencies
between objects—only between classes.
The reason is again reversibility. An object-oriented system is specified as a
structured collection of classes with precisely defined interfaces. It is through
the classes—and only through them—that object behavior is defined. Therefore,
if we make dependencies between individual objects (class instances) part of our
high-level static notation, there are two possibilities. Either the object structure
can be inferred from the class structure and vice versa (which means the notation
is not needed in the first place) or there is no simple mapping, in which case we
70 STATIC RELATIONS

break the seamlessness of the approach. When we need to emphasize structural

details about individual class instances, this is instead done through assertions in
the corresponding class interfaces.

Associations and aggregations

There are three types of client relation: association, shared association, and
aggregation. Each relation represents the supplier being used by the client to
fulfill a certain role. There can be several relations from one client to the same
supplier class, whose instances are then used to play multiple roles.
An association between a client class and a supplier class means that (at
system execution time) some instances of the client class may be attached to one
or more instances of the supplier class. A particular instance of the supplier class
may take part in many such attachments, thus permitting supplier objects to be
shared by different clients.
A shared association between a client class and a supplier class is a special
case meaning that whenever an instance of the client class is attached to an
instance of the supplier class, it will always be to the same supplier instance (or
one in a fixed set). This permits specific instances to be shared by all client
objects of a certain type. BON includes notation to express this type of
relationship, as we will see later.
An aggregation relation between a client class and a supplier class means that
each client instance may be attached to one or more supplier instances which
represent “integral parts” of the client instance. The parts may in turn have client
relations to other classes, which may be aggregations (subparts) or associations.
Aggregation (whole vs. parts) is an important semantic concept, useful when
thinking about the properties of objects in a system—which is why it has a
special notation in BON—but not so easy to define exactly. For example, some
operations applied to the whole will naturally propagate to its parts. Moving a
house will automatically move all the rooms inside. On the other hand, painting
a house does not in general mean that each room gets painted. Selling a house
may seem to imply that everything inside goes with it, but it is fully possible to
sell an apartment building without selling any of the owner apartments contained
in it.
Sometimes the internal parts of an object can only be accessed by the object
itself and therefore cannot be shared by any other client. In other cases, it may
be more practical for the enclosing class to export some of the parts as public
operations. Deleting an object usually means that all its aggregated parts will
also be deleted, but even in this case there are other possible strategies.
The conclusion is that the modeler must be allowed some freedom in using the
aggregation concept, and that its exact semantics will probably vary from one
 CLIENT RELATIONS 71

context to another also within the same system. Therefore, we cannot prescribe
any particular implementation of aggregation relations compared to ordinary
associations. That will depend on the objects involved, the language
environment, as well as general efficiency considerations.
However, we can state one absolute rule for aggregations: a particular object
can be an integral part of at most one other object. For example, an engine can
only be part of one vehicle (at a time) but may be referenced by other parties,
such as the manufacturer, the owner’s service garage, and the vehicle registration
authorities.
All types of client relations may be labeled with the names (in the client class)
of the features giving rise to the relation. These names should mirror the
corresponding role played by the supplier.

Graphical representation
A client relation is represented by a double line extending from the client to the
supplier. We call this line a client link. Client links may be broken to avoid
crossing other elements, but just as for inheritance relations the standard rule is
to only use horizontal and vertical directions. (Again, these are
recommendations for case tool construction; on whiteboard and paper we do not
impose any such restriction.) Association links end with an arrowhead pointing
to the supplier, as illustrated by figure 4.4.

residence location
PERSON ADDRESS CITY

Figure 4.4 Association relations

Aggregation links end with an open brace, as in figure 4.5. A client link can
be labeled with one or several names, corresponding to class features in the
client which cause the relation. In the case of multiple labels, the link represents
multiple relations of the same type, one for each name.
The reader may wonder why the particular directions for class relations were
chosen in BON. Well, there are also valid arguments for using the reverse
conventions. For example, inheritance links pointing from parent to child could
indicate feature propagation or extension of functionality, and links from
supplier to client could suggest provision of service. However, signaling

propulsion combustion_chamber
VEHICLE { MOTOR { CYLINDER

Figure 4.5 Aggregation relations

72 STATIC RELATIONS

dependency is usually more important: a child depends on its parents and needs
to know about them, but not the reverse. Similarly, a client depends on its
suppliers and needs to know about them, but not the reverse. In fact, a major
point in object-oriented development is that classes should not know beforehand
to what extent they will be used as suppliers or parents during their lifetimes.

Bidirectional links
A set of client association links in each direction between two classes may be
combined into one double link with an arrowhead at each end, as in figure 4.6.

brothers, sisters
children
CUSTOMER MOTHER CHILD
mom
shoppers: SET […]

preferred_mall
SON DAUGHTER
SHOPPING_CENTER

OFFSPRING

Figure 4.6 Bidirectional association links

In the case of labeled bidirectional links, we need a convention to show which

relation each label refers to. The rule is simply to put each label closer to the
supplier side. In figure 4.6 preferred_mall thus refers to a feature in class
CUSTOMER of type SHOPPING_CENTER, while shoppers refers to a feature in
SHOPPING_CENTER of type SET [CUSTOMER]. The latter is an example of
the use of generic classes in client relation labels, and will be explained later.
(Note that only bidirectional links require labels to be put closer to the supplier
side in order to remove ambiguity. For the vast majority of links, which are
unidirectional, labels may be put anywhere along the link.)
Bidirectional aggregation links might seem like a contradiction in terms: two
objects cannot both be part of each other. However, we must recall that client
relations are between classes, and it is not impossible for a class A to have an
integral part of type B and class B to have an integral part of type A, provided
that different class instances are involved. In fact, such recursive structures
often occur in practice. Figure 4.7 shows three examples: the first is the famous
Russian doll, and the second illustrates a common type of document, where each
section consists of an introduction followed by the main text. The third example
in figure 4.7 shows a bidirectional client link where one end is aggregation and
 CLIENT RELATIONS 73

introduction

{
SECTION TEAM
RUSSIAN_DOLL
preferred_team

{
subsections: SET […]

{
members: SET […]

{
my_matrusjka main_text

{
PLAYER
SECTION_LIST

Figure 4.7 Bidirectional aggregation links

the other association. A team has members, but some members may secretly
dream of playing in another team.

Multidirectional links
To simplify diagrams with many relations, we allow client links to be joined as
shown in figure 4.8. The five airplane parts in the figure could also have been
collected in a subcluster with the parts relations compressed into a cluster
relation. This is the recommended style when there are many parts.

AIRCRAFT

{ ROOF

HOUSE
{
{

{
WING FUSELAGE TAIL
{ CELLAR
{

LANDING_GEAR COCKPIT

Figure 4.8 Grouped client links

When several clients share the same supplier (figure 4.9), placing a label near
a client means that it only corresponds to a relation from that client, while near a
supplier means that all the clients are related to the supplier through this name.
Finally, we may also have several clients and several suppliers. Then the
semantics is as follows: each client has a relationship to each supplier of the type
indicated (arrowhead or brace). Figure 4.10 shows four network nodes sharing a
74 STATIC RELATIONS

STOCK_TRADING WATCH

black_thursday
power_source
DATE { BATTERY

milestone

PROJECT_PLANNING CD_PLAYER

Figure 4.9 Avoiding ambiguities by label placement

FILE_SERVER ERROR_LOG

{
host log

WORKSTATION PC TERMINAL PRINTER

Figure 4.10 Multiple clients and suppliers

file server, each node keeping a private error log. When there are several
suppliers, labels are only permitted to be put close to a supplier, since close to a
client would be ambiguous.
Joining only suppliers without any clients is not permitted for more than two
classes, since a network where all ends of the double lines have an arrow or a
brace attached looks too confusing: any interpretation would be ad hoc, and
difficult to remember. A number of bidirectional links (described in the previous
section) can be used instead. Joining only clients without any suppliers
obviously does not make sense.

Links involving generic suppliers

Generic classes are special in that they do not correspond to one type each, but to
a potential group of types. Each new combination of actual parameters given to
a generic class will derive a new type. (This is sometimes called type
instantiation, but we use the alternative term type derivation in this book to
avoid confusion with the term class instantiation, which means creating objects
as instances of classes.)
 CLIENT RELATIONS 75

One characteristic of genericity is that it is generally not possible to conclude

the client dependencies of a generic class by simply inspecting its interface (and
possibly private features), as can be done with non-generic classes in a typed
environment. The information needed is distributed among the clients, and the
dependencies are therefore deduced by the supporting compiler environment.
This calls for some special conventions with respect to the graphical notation.
Before embarking on a description of these conventions, we make it clear that
the BON notation does not support all possible uses of genericity. First,
recursive propagation of formal parameters from one generic class to another can
lead to very complex situations, where full understanding requires inspection of
the interfaces proper (and perhaps also the code). Second, a very detailed
graphical notation would require the target language to be fixed.
However, there are a few typical cases where the use of genericity can be very
helpful already during analysis and early design—most of them involving very
general data structures—so we want to provide some notational support but still
keep things simple. There are two main situations in which generic derivation
arises: through a generic supplier class or through a generic parent class. We
look first at generic suppliers.
Assume that a geometric figure has a list of points outlining its border. This
can be expressed using the notation described so far, as shown in figure 4.11.

border LIST item

FIGURE POINT
[T]

Figure 4.11 Indirect client dependency through generic derivation

To highlight that the client relation to class POINT (labeled item) corresponds
to a generic derivation (that is, a declaration of type item: T inside class LIST,
rather than item: POINT), we use the formal generic parameter as a label marker.
The marker is in a different font (bold, non-italic in this case) to show that it is
not a type but a formal name. The name or parenthesized list of names of the
client features whose signatures contain the generic derivation may precede the
marker, separated by a colon.
This is shown in figure 4.12, where we have also added a center point to the
figure. The label now shows that the type of the feature item in class LIST is the
center

border LIST item: T

FIGURE POINT
[T]

Figure 4.12 Generic label marker

76 STATIC RELATIONS

formal parameter T generically derived as POINT to form the client dependency

expressed by this link.

Compacted generic links

However, the generic classes involved in analysis and design are often
extensively reused library classes, representing high-level data structures. In
such cases, it is generally much more informative to emphasize the indirect
relationship between the client class and the data structure elements. Knowing
that a figure consists of points is normally more important than knowing if the
points are kept as a set or as a sequence. The structural properties can then
instead be put as additional information on the client/supplier links, thus playing
the role of semantic labels often found in entity−relationship diagrams.
To achieve this, we allow client relations arising from generic derivation to be
compacted. This means that links to the corresponding generic class are
suppressed, and the generic class name appears instead as part of a label on a
link connecting the client and the supplier corresponding to a derived generic
parameter.
The label of such a compacted link contains the name of the generic class
followed by its actual type parameters in square brackets, where the parameter
corresponding to the supplier has been replaced by an ellipsis. The name or
parenthesized list of names of the corresponding client features may precede the
generic class name, separated by a colon. Using this form simplifies the
previous diagram (figure 4.13).

border: LIST […] , center

FIGURE POINT

Figure 4.13 Compacted generic relation

When several client classes use the same generic supplier or there is more than
one generic parameter, the benefits of the compacted forms become even more
apparent. In figure 4.14, the upper part shows the full client dependencies
resulting from two derivations each of the generic classes LIST and TABLE,
while the corresponding compacted forms are in the lower part.
The TABLE class is an example of a very general class with more than one
generic parameter, used to store various elements for later retrieval through a
specified key of some type. Only classes that inherit from class KEY may be
used at the second generic parameter position in a TABLE derivation (shown by
the → KEY suffix), which guarantees the existence of an operation for
constructing a unique key code.
 CLIENT RELATIONS 77

BASEBALL_
CARD

(first, last): T

expenses LIST assets

SON PARENT
[T]

(first, last): T

ACCOUNT

PATIENT

patients key: V
NURSE item: U PATIENT_ID

TABLE
[U, V→KEY]

duties key: V
SURGEON item: U DATE

OPERATION

expenses: LIST […] BASEBALL_

SON
CARD

assets: LIST […]

PARENT ACCOUNT

patients: TABLE [… , PATIENT_ID]

NURSE PATIENT

duties: TABLE [… , DATE]

SURGEON OPERATION

Figure 4.14 Multiple generic derivation

78 STATIC RELATIONS

We see that besides being much simpler to read, the compacted relations also
remove the ambiguity regarding who is really related to what list or table
elements. Even if labels like expenses and patients give hints as to what they
refer to, we cannot be sure.
In contrast to the semantic labels used in ER modeling, whose meaning suffers
from the vagueness of natural language, the semantics of the generic classes used
in client relation labels can be precisely specified through formal assertions.
Carefully defining a set of high-level generic classes may thus serve as an
extension of the BON notation to tailor it to the needs of different application
areas.

Links involving generic parents

Besides the direct derivation of a generic supplier class, there is also the case
when a class inherits from a generic parent, and thus creates a client dependency
to the derived type parameters. This can be expressed as in figure 4.15, where
we again use a generic marker in the label, this time preceded by an arrow to
indicate the inheritance relation involved. The corresponding compacted form is
shown in the lower part of the figure.

SEQUENCE →T
BYTE
[T]

FILE

→ SEQUENCE […]
FILE BYTE

Figure 4.15 Generic client relation through inheritance

Role multiplicity
There can be many client relations between the same two classes A and B. So
far, we have expressed this by attaching several labels to the same graphical link
from A to B. In some cases (though not very often), one might want to
emphasize the number of links rather than the names. As an alternative to
labeling, it is therefore possible to use a marker (small lozenge, similar to the one
used for direct repeated inheritance) containing the number of links. Multiplicity
 CLIENT RELATIONS 79

markers may also be attached to bidirectional client relations. The lozenge is

then split in two parts to indicate the direction corresponding to each number.
This is shown in figure 4.16, which expresses the following:
• Arriving aliens hold two attachments to landing document.
• An apartment contains three attachments to room.
• A house has one architect and each architect has three house attachments.

VISITING_ALIEN 2 LANDING_DOCUMENT

APARTMENT 3 { ROOM

HOUSE 3 1 ARCHITECT

Figure 4.16 Multiplicity markers

Before proceeding, we want to make it completely clear that the above numerical
markers are not there to express multiplicity in the traditional data modeling
sense. There is a fundamental difference here: multiplicity notation in data
modeling is about the number of class instances (often called object instances in
traditional contexts), while BON’s multiplicity markers refer to the number of
different roles under which a certain class is used by another class (note the word
attachment in the list above explaining figure 4.16).
Since one normally gets a much clearer picture by naming the roles instead of
just counting them, multiplicity markers are used very sparingly with BON.
Usually, the previous diagram would be something along the lines of figure 4.17.
However, the multiplicity included in most other object-oriented analysis and
design notations simply mirrors the instance multiplicity from data modeling.
The reasons for emphasizing this information are historical: the multiplicity
affects the way database records are designed in traditional environments.
But with an object-oriented approach this is no longer a problem, since it
handles instance multiplicity as a duck takes to water. The designer whose
object-oriented systems must be extensively restructured because we must now
handle two objects of a certain class instead of just one should not be trusted
with industrial projects. Therefore, in our opinion, a constraint on the number of
instances of a certain class (of the type many-to-one, one-to-one, etc.) is
80 STATIC RELATIONS

immigration_form,
customs_form
VISITING_ALIEN LANDING_DOCUMENT

kitchen, bedroom: SET […] ,

living_room: SET […]
APARTMENT { ROOM

summer_house,
winter_cottage,
main_residence designer
HOUSE ARCHITECT

Figure 4.17 Naming, rather than enumerating, roles

irrelevant information at the system level, and should instead be placed where it
belongs: in the class interface specifications as part of the software contract.

Instance multiplicity
Whether we use multiplicity markers or labels in BON, they only express
potential associations and aggregations. For example, in figure 4.16 an instance
of class VISITING_ALIEN may (at different times during system execution) be
associated with zero, one, or two instances of class LANDING_DOCUMENT.
What the diagram really means is that class VISITING_ALIEN contains exactly
two entities of type LANDING_DOCUMENT, but how and when these entities
will be bound to objects of the corresponding type is decided dynamically.
Perhaps surprisingly, this also applies to aggregations. The APARTMENT
objects of figure 4.17 are probably thought of as containing their ROOM objects
directly when created, but this is not a general property of abstractions modeling
part-of relations. Clearly, an electronic document may be regarded as consisting
of its pages, sections, and paragraphs at any point in time, even if this structure is
allowed to change dynamically. The Russian doll in figure 4.7 may also be
considered an aggregation. Each doll has another smaller doll as a potential
subpart, but at some point we must find an empty doll. The same holds for any
recursive parts explosion: at least some of the classes used in the modeling must
describe objects that sometimes have subparts and sometimes not.
This shows that aggregation cannot be identified with what is known as value
types or expanded types in object-oriented languages, that is types whose objects
are fully expanded already when system execution starts. Such types are useful
for performance reasons (or in some cases, like heavy numerical processing,
even necessary), but they are issues of implementation, not of high-level
modeling.
 CLIENT RELATIONS 81

Instance sharing
So, in accordance with the above reasoning, the notation for multiplicity only
addresses the number of entities introduced in the client class. Contrary to
dynamic instance multiplicity, these entities are important, because they mirror a
conscious choice of roles assigned to classes in various contexts, which will
eventually be reflected in the implementation to be debugged and maintained.
More detailed constraints on individual instances may be specified in the
contracts of the corresponding classes.
There is, however, one special aspect of instance association which merits
special treatment, and that is object sharing. Very often groups of related
objects share a common supplier, which provides a certain service to the group.
Standard examples are window objects in a graphical user interface system all
sharing the same mouse object, or a set of diskless workstations sharing a
common file server.
By placing a special sharing marker (small circle) on a client link, the BON
user may specify that when a certain entity in a class is attached to an instance of
the supplier class, it will always be to the same instance. Figure 4.18 shows a
HOCKEY_PLAYER class whose instances all share the same instance of the class
PUCK. (Allowing each player to instantiate a private puck might lead to some
interesting surrealistic matches, but there would clearly be a consensus problem
regarding the final result.)

HOCKEY_PLAYER 1 PUCK

Figure 4.18 Necessary sharing of one instance

All instances of a client class may also share a fixed number of supplier
instances. This may occur in two ways. First, there may be several shared static
relations from one class to another. All instances of class STUDENT may, for
example, share one instance of class ROOM through a static link labeled kitchen,
and at the same time share another ROOM instance labeled bath_room.
Second, instead of being attached to the exact same instance, each static link
may be attached to any of a fixed number of instances. The STUDENT instances
may, for example, have access to one of three instances of class
TENNIS_TRAINER through a static link labeled current_trainer, but be attached
to at most one at a time. This is expressed by replacing the number 1 inside the
shared circle symbol by the corresponding number (in this case 3).
Several examples are shown in figure 4.19. In the left part of the figure, the
three diagrams express that all PC objects have two static links, each sharing an
82 STATIC RELATIONS

FILE_ FILE_
PC 2 1 PC 2
SERVER SERVER

server1
1 server
FILE_ FILE_
PC PC 2
1 SERVER SERVER
server2

server1,
server2
FILE_ FILE_
PC 1 PC 2 3
SERVER SERVER

Figure 4.19 Different ways to express sharing

instance of class FILE_SERVER. (These two instances could actually be the

same, but are probably not.) The links are unnamed in the upper diagram, and
named in the other two. The lower diagram is exactly equivalent to the middle
one, only more readable and compact.
Turning to the right part of the figure, the upper diagram shows a PC class
whose instances may be alternatively attached to one of two FILE_SERVER
objects. The middle diagram is exactly the same, only the link has now been
labeled. The lower diagram, at last, shows that each PC object has two shared
static relations to class FILE_SERVER, where each link may be attached to one
of three specific FILE_SERVER objects.
Finally, figure 4.20 shows how local sharing within several related groups can
be factored using inheritance. If a parent class contains a shared client relation to

* dividend_payer *
SHAREHOLDER 1 COMPANY

MICROSOFT_ ENEA_ MICROSOFT ENEA_DATA

SHAREHOLDER SHAREHOLDER

SUN_ SUN_
SHAREHOLDER MICROSYSTEMS

SHAREHOLDERS COMPANIES

Figure 4.20 Sharing through inheritance

 CLIENT RELATIONS 83

a supplier class, but the corresponding feature is deferred, this means that sharing
will be effected by descendant classes. In figure 4.20, the shareholder classes
corresponding to each company will make the deferred feature dividend_payer
effective by defining an implementation for it, which ensures that all instances of
this class (and possible descendant classes) will share the same supplier object
(in this case, an instance of the proper company).
If, by contrast, the dividend_payer feature had been declared effective already
in class SHARE_HOLDER, all instances of all the companies would have shared
the same supplier object, which is not what was wanted in this case. Once a
feature representing a shared relation has been effected, it cannot be redefined by
a descendant class since this would violate the global sharing.

Client relations involving clusters

Just like the inheritance relation, the client relation may be generalized to apply
also to clusters. Figure 4.21 shows an example. The ORGANIZATION cluster of
a conference system is a client of the CONFERENCE_DATA cluster, which in

ORGANIZATION

PROGRAM_ { program organization ORGANIZING_

COMMITTEE
CONFERENCE { COMMITTEE

●
COMMITTEE

chair, members: SET […] ,

referees: SET […] ,
conf_event

calls: SET […] ,

CONFERENCE_ answer *
SEND_OUT
EVENT

registrations: SET […]

…
{

attendee
PERSON REGISTRATION

CONFERENCE_DATA CORRESPONDENCE

Figure 4.21 Client relations involving clusters

84 STATIC RELATIONS

turn is a client of the class SEND_OUT in a third cluster. The labels refer to
features of classes in the client cluster that are clients of some class in the
supplier cluster, but we cannot tell which ones from the diagram. However, the
names often provide enough guidance for the analyst/designer who has worked
extensively with the model to know more or less which classes are involved. If
more information is needed, the client links may be expanded to present instead
a view like the diagram in figure 4.22.

ORGANIZATION

PROGRAM_ { program organization ORGANIZING_

COMMITTEE
CONFERENCE { COMMITTEE

referees: SET […]

●
COMMITTEE
chair,
members: SET […]

1 conf_event

CONFERENCE_ calls: SET […] *

SEND_OUT
EVENT

registrations: SET […]

…
{

attendee answer
PERSON REGISTRATION

CONFERENCE_DATA CORRESPONDENCE

Figure 4.22 Expanded client relations between clusters

Here, the individual class relations between clusters have become visible at
the expense of a more complicated diagram. This is a typical example of scaling
in BON, which is so fundamentally important for large systems. If we want
more detail, we expand down to the individual class interfaces; to get more
overview, we compress into clusters at higher levels. The labels may of course
be hidden if we are only interested in the general structure of a system.
As was pointed out earlier, the semantics of client relations involving clusters
is different from the corresponding semantics of inheritance relations. While an
 CLIENT RELATIONS 85

inheritance relation having a cluster at either end carries over to relations with all
classes in the cluster, client relations only refer to at least one class in each
cluster.
This might seem a bit strange at first sight, but is really quite straightforward.
The general semantics of inheritance is that of “all or nothing”—either you are
my son and then you have all my genes, or you are not my son at all. A child
class cannot select the operations to inherit from its parent; it receives
everything. (It can, however, modify the behavior of chosen operations, which
we may think of as computational mutation, but contrary to biological mutation
only modifications that are in accord with the software contract collectively laid
down by the ancestors are accepted.)
By contrast, the general semantics of being a client is that of “free choice”. A
client can request any service provided by a supplier class, but is not obliged to
use everything. So if we think of the classes of a cluster as somewhat analogous
to the operations of a class, the conventions used for cluster relations are easy to
remember. The rules for generalizing client relations to involve clusters may be
put more precisely:
1. If one or more elements (classes or clusters) in a cluster X are clients of an
element A outside of X, then the cluster X is said to be a client of A. All
the corresponding client links can then be compressed into one client link
from X to A.
2. If an element A outside a cluster Y is a client of one or more elements in Y,
then A is said to be a client of Y. All the client links can then be
compressed into one link from A to Y.
3. A client link between two elements can be compressed into being hidden
(not shown in a diagram).
All three cases occur frequently in practice, and have to do with the scaling level
chosen. Notice again the use of element instead of class in the above definition
to allow for recursive compression. As was shown earlier for inheritance
relations (figure 4.3), client relations between classes in two different clusters
may be successively compressed into client relations between the two clusters.
For large systems, many levels of cluster compression is common practice.

Compression of client relations

Case tools may employ different strategies for compression of client relations, as
long as they comply with rules 1−3 above. For example, it may be advantageous
to be able to hide all client relations to certain groups of classes in order to
emphasize some overall aspect of the system. Typical candidates for exclusion
86 STATIC RELATIONS

are general support classes, but also application classes considered less important
in the context studied.

Compression of labels
The labels of a client link between two classes are compressed as follows:
• Unidirectional links. If there is more than one label attached to the link,
the labels are compressed into the first label with a marker (ellipsis)
indicating that not all labels are shown. If there is only one label (marked
or not) it is compressed by being hidden.
• Bidirectional links. The unidirectional rules apply separately to the label
groups of each direction. Recall that labels on bidirectional links must be
placed closer to the supplier.
• Multidirectional links. The unidirectional rules apply separately to the
label groups closest to the same client or to the same supplier. Recall that
if there are multiple suppliers, then labels can only be put close to one or
more of the suppliers, which means the names correspond to features in all
the clients.

4.3 CLASS VS. OBJECT DEPENDENCIES

There is a fundamental difference between dependencies among the elements of
a system (the objects) and among their descriptive elements (the classes). The
full collection of objects at run-time is potentially a set of gargantuan
proportions, and would be totally unmanageable if the elements always had to be
dealt with individually. Fortunately, object-oriented development reduces this
set by orders of magnitude by grouping objects with similar behavior into
classes. The collective dependency of objects of type A on objects of type B may
then be captured through one dependency link from class A to class B.
However, even with this reduction the resulting number of dependencies may
become quite large. The good news is that another of the object-oriented key
features—inheritance with dynamic binding—can carry the reduction one step
further (one example being the factored sharing of figure 4.20). We will clarify
the principle with yet another example. The classes in figure 4.23 represent
some different vehicle types and a set of potential owners. Since any owner can
own any vehicle, there are nine class dependencies to keep track of. These have
been compressed into one cluster dependency for convenience. (Recall that a
client relation between clusters X and Y means that at least one class in X is a
client of at least one class in Y; here all classes in X happen to be clients of all
classes in Y, which is somewhat unusual.)
 CLASS VS. OBJECT DEPENDENCIES 87

CAR TRUCK PERSON COMPANY

BUS TRUST

VEHICLES OWNERS

Figure 4.23 Multiple class dependencies

If, instead, we let the vehicles and owners inherit from common ancestors
encapsulating the interaction as depicted in figure 4.24, the scene changes
radically. The nine class dependencies have now been reduced to just one.

* owner *
VEHICLE ASSET_HOLDER

CAR TRUCK PERSON COMPANY

BUS TRUST

VEHICLES
OWNERS

Figure 4.24 Factored class dependencies

The object dependencies have not changed—CAR objects are still clients of
COMPANY objects and so forth (and the deferred ancestors are never
instantiated)—but the big win is that the vehicle classes and the owner classes
have become totally independent of each other. Changes to one class do not
affect the correctness of the others, whereas in the first case changing, adding, or
deleting an owner class would potentially require modification of all vehicle
classes.
The bottom line is this: if we can only control our class definitions to make
sure all objects will behave correctly at execution time, we can forget about the
complexity of object dependencies—they will take care of themselves. That is
why we must concentrate on classes and downplay individual objects.
88 STATIC RELATIONS

4.4 INDIRECT CLIENT DEPENDENCIES

A client relation between a client class A and a supplier class B means A is
dependent on services supplied by B. The services of B may be used for
specification, as part of the public interface of A, or behind the scenes in the
implementation of A. A client link from A to B may correspond to one or more
of the following cases:
1. A public feature in A returns a value of type B (or a type derived from B, if
B is generic).
2. An input argument of a public feature in A is of type B.
3. An actual parameter in a generic type derivation specified in A as part of
the public interface is of type B.
4. The implementation of A uses an entity of type B. This includes the
signatures of private features, as well as state variables and local variables
used inside features.
In BON, as we have seen, dependencies arising from public query return values
and generic derivation (cases 1 and 3 above) may be captured directly in client
labels. Regarding public input arguments and implementation (cases 2 and 4
above), these can be captured by unlabeled client links. Extending the labeling
syntax to also cover these cases would confuse the notation and does not seem
worthwhile. When more precise information is needed, we can turn to the
formal class interfaces and/or the source code.
It is a general principle in BON to strive for direct correspondence between
the information expressed by client links in static diagrams and the formal
interfaces of the classes involved. Therefore, the colon separating query names
from their return type in class interfaces can be adorned to show the type of
relation (association, shared association, or aggregation). Since the two last
types of client relation have been explained at length in this chapter, we may
now present the full feature signature table (figure 4.25), for which a partial
version was earlier given in chapter 3.

FEATURE SIGNATURES (full table)

Graphical form Textual form Explanation

name: TYPE name: TYPE Result type

name:{ TYPE name:{ TYPE Aggregation result type
name:(n) TYPE name:(n) TYPE Shared result type (n objects)
– arg: TYPE −> arg: TYPE Input argument

Figure 4.25 Typed signatures revisited

 SEMANTIC LINKS 89

4.5 SEMANTIC LINKS

A semantic link is an annotation associated with an inheritance or client relation
between two classes that details the designer’s intent when defining the
relationship. Examples of semantic links are listed in figure 4.26, but the list is
extensible. The semantic links are stored as comments with the corresponding
relations.

Inheritance Client/Supplier

[DESCENDANT] is a [PARENT] [CLIENT] requests [SUPPLIER]

[DESCENDANT] behaves like [PARENT] [CLIENT] uses [SUPPLIER]
[DESCENDANT] implements [PARENT] [CLIENT] needs [SUPPLIER]
[DESCENDANT] combines [PARENT] [CLIENT] has a [SUPPLIER]
[PARENT] defers to [DESCENDANT] [CLIENT] consists of [SUPPLIER]
[PARENT] factors out [DESCENDANT] [SUPPLIER] provides [CLIENT]

Figure 4.26 Typical examples of semantic links

5 The dynamic model

5.1 INTRODUCTION
The static model of a system shows the high-level classes grouped into clusters
and the relationships between them (client and inheritance). It also shows the
class operations, their signatures and semantic specification through pre- and
postconditions, as well as the overall consistency rules for the corresponding
objects expressed as class invariants. These are the class contracts so
fundamentally important for reliable systems. Getting them right is a big step
towards mastering the task of systems development.
However, even if the semantics of the class operations is clearly specified
(which is certainly not the case for most systems today), a crucial detail still
remains: their implementation. Procedural functions do not magically disappear
just because we structure our systems as collections of classes. On the contrary,
we probably get many more of them than with traditional approaches, but with
two main differences. First, the operations are not global to the system but local
to the context in which they serve their purpose: the respective classes. Second,
they are normally much smaller and easier to understand, since the abstractions
they use as elements can be made to mirror the problem domain that much better.
But easy does not mean trivial. There is no such thing as an automatic
generator of class operations from a semantic description of a class (unless of
course the description is a full program in which case nothing has been gained).
That is why we still need programming languages to describe the procedural
steps of our class operations in enough detail to be executable by a machine.
Analysis and design does not include implementation, but it does include
ensuring that implementation is indeed possible at reasonable cost and with
adequate performance of the resulting product. Therefore, we need at least a
vague idea already at the early development stages of how the operations of our
high-level classes can fulfill their specifications by calling other operations (in
the same or in other classes). To help capture and communicate this idea is the
purpose of the BON dynamic model.

90
 INTRODUCTION 91

Object-oriented execution
One of the really nice things about object-oriented abstraction is that there are
very few basic concepts to keep track of. This has led to an extremely simple
execution model. A full execution of a system is simply the invocation of one
operation on a so-called root object. When this operation terminates, system
execution is over. The root object is instantiated automatically from the root
class by the system environment when the corresponding process is started.
A class operation performs a set of actions in sequence. There are only two
possibilities for each action: either terminate the operation by returning to the
caller, or invoke another operation. Since the decision whether to return or make
a new call—and if so, what other operation to call—may depend on the system
state, we also need a few basic control primitives. In fact, we only need two:
multiple choice and iteration, both based on a boolean condition.
Except for these control primitives, everything in an object-oriented execution
can be boiled down to operation invocation.5 This includes basic arithmetic and
logic, which traditionally uses operand/operator expressions that do not
resemble feature calls. However, this is just a syntactic convention. For
example, the expression
n+5<k∗2
can be rephrased as
n.plus (5).less_than (k.times (2))
without changing any semantics. Some object-oriented languages support prefix
and infix forms of class operations, making it possible to write calls like the
expression above in the familiar algebraic style.
What the above shows is that not only the object-oriented specification model,
but also the corresponding implementation model have an inherent coherence,
simplicity and beauty that are truly remarkable. This is crucial for our approach,
where the goal is to achieve true seamlessness not by reflecting low-level details
of code in the analysis and design, but by bringing high-level abstraction all the
way down to the executable instructions. When the artificial barriers between
the noble art of specification and the dirty (but unfortunately necessary)
workmanship of implementation have been torn down, and we realize that there
is no conceptual difference between the two, then we can hope for some real
progress in software productivity.

5
Well, almost; for completeness we do need a few more very basic primitives, like state variable
assignment and test for object equality.
92 THE DYNAMIC MODEL

What should be captured by a dynamic model?

The previous section showed that feature calls, or message passing between
objects, are what constitute a system execution. Consequently, this is what
should be expressed in a dynamic model as a complementary view to the static
model. Using a notation based on other concepts again loses reversibility and
breaks the seamlessness of the approach.
Nevertheless, we have yet seen very little useful notation for documenting
object communication in books on analysis and design. Mostly, they dwell for a
page or two on an oversimplified scenario of perhaps four or five objects whose
interactions are illustrated by the equivalent of an event trace diagram (a diagram
with objects on one axis and time on the other). Realizing that the expressive
power of such diagrams is much too weak for real-life systems, they then quickly
move on to methods outside the object-oriented world, mostly state transition
diagramming and/or data flow analysis.
In our view, functional modeling with data flow analysis should not be used at
all in an object-oriented method, since the impedance mismatch and model
confusion it introduces far outweigh any benefits gained. It can always be
replaced by better ways of arriving at good abstractions.
State transition diagramming, on the other hand, is a powerful technique that
is sometimes very useful also in object-oriented contexts. When large amounts
of discrete events need to be handled by an object, finite state machines (FSMs)
may be used to systematically take care of all possible combinations that can
occur, and to give each required action its natural place in a general structure. A
simple introduction to FSMs can be found in [Allworth 1990].
But it is important to understand that this technique is not a panacea. To
recommend that each class should model its internal behavior using state
machines is, in our view, totally inappropriate. It is like proposing that all
mathematical functions should be described by enumerating their return values
for each input value. This is fully possible on computers using a fixed storage
size to represent their numbers, but it is hard to conceive of a more simple-
minded approach, the side effects being increased complexity, waste of space,
and total lack of reusability.
On the contrary, the goal to strive for is always higher-level abstractions that
can be naturally described with much fewer parameters—where we do not need
to use what is known as the British Museum method of explicit enumeration.
Only for the unfortunate cases, where no nice continuous abstractions can be
found for the behavior, do we call upon the state machine brigade to come in and
sort out the disorder and transform the unstructured spaghetti logic into
controllable and structured spaghetti logic (all global dependencies still there, but
now sorted and tagged).
 INTRODUCTION 93

A typical such case, where state modeling can be of great help, is the design of
user interfaces, where many alternative actions can be chosen by the user in
various contexts and the system has no way of predicting what the next choice
will be. When state diagramming is needed, there are many well-proven
notations to choose from, so BON does not invent its own variant. Instead, we
recommend David Harel’s statechart formalism [Harel 1988], which has the
ability to structure large diagrams through nesting.
The conclusion is that although state diagramming may be useful in special
cases, the dynamic model of a general object-oriented method should instead
concentrate on describing behavior through simple message passing between
objects in the system. However, a more expressive notation than event trace
diagrams is needed.
The BON dynamic notation uses sequence labels to capture time, which frees
the second spatial dimension of a diagram to show the interaction between a
larger number of objects than is possible with event trace diagrams. Moreover, it
contains an object grouping facility that can capture more general scenarios and
also illustrate dynamic binding. Potential message passing between selected
objects is shown using a third kind of relationship, the message relation.
Besides object communication, the dynamic model contains three additional
BON modeling charts. These are: the event chart, which records incoming and
outgoing system events; the scenario chart, containing descriptions of selected
object communication scenarios; and the object creation chart, showing what
classes create new instances of what other classes. The informal charts and the
dynamic notation for expressing object interaction will be described in detail in
the rest of this chapter.

5.2 DYNAMIC CHARTS

System events
A system may be viewed as a “black box” whose behavior is defined by the
response it gives to various stimuli. A stimulus that can change the course of
action of a system and make it react is called a system event. System events are
of two kinds. An external event is triggered by something in the external world
over which the system has no control. Examples of incoming external events are
user terminal input, calls from other systems, arrival of data from sensory
devices, and hardware clock interrupts.
An internal event, on the other hand, is triggered by the system itself as part of
its reaction to one or more external events. Some of the internal events are
outgoing, which means that they send stimuli to the outside world triggering an
(outgoing) external event, mostly as a response to external input.
94 THE DYNAMIC MODEL

For example, a user at a workstation may click on a menu to print a certain

report. The receipt of the mouse click is then an external event, which may lead
to a whole series of internal events triggering various parts of the system to
search databases, compile information, and send the result to a printer. At least
one of these internal events (not necessarily the last one) will be an outgoing
internal event representing the system response, perhaps directed to a network
spooler.

Event charts
The object interactions which make up the execution of a system are ultimately
caused by external events. Therefore, it is often a good idea to compile a list of
external events that may trigger object communication representing principal
types of system behavior. The event list is then used as a guide for choosing the
scenarios to include in a dynamic model of the system. However, only a small
subset of all possible external events is interesting and representative enough to
be listed.
For example, to register a new subscriber for a monthly comic magazine, an
operator may first click on a principal subscription menu, then choose the entry
for the complete Carl Barks Library in color, then enter the name and address of
the subscriber along with the subscription period, and finally click on the button
“accept subscription”. In this series of operator actions, each mouse click will
constitute an external system event (in fact, usually two: one when the mouse
button is depressed and one when it is released). Depending on the supporting
system (widget handling, mouse tracking, etc.) each character keystroke may or
may not be an individual event, and each of these events may lead to a large
number of operations being applied to many objects.
But the only event in the above sequence likely to be included in a high-level
system dynamic model is the receipt of “accept subscription”. This event
corresponds to the commit statement of a transaction, and is probably a
significant part of the system behavior. The other events have to do with how
the necessary subscription data is actually collected and transmitted, and since
this can be implemented in many different ways, the details will in most cases
not be decided until later. Unless a novel interface model is part of the system
design, the menu selections will just follow the standard practice of some basic
general GUI library, and thus not be included in the dynamic model.
The idea is to capture a small number of external events each triggering a
principal type of system behavior, so that a representative set of scenarios can be
chosen. These scenarios will then sketch how the necessary actions can be
carried out by communicating objects (which are instances of the classes in the
static model) to achieve the desired behavior. Figure 5.1 shows a BON event
 DYNAMIC CHARTS 95

chart for a conference system, collecting some interesting incoming external

events. For each external event, the event chart lists the types of object that may
become involved as part of the system response.

EVENTS CONFERENCE_SUPPORT Part: 1/2

COMMENT INDEXING
Selected external events triggering created: 1993-02-15 kw
representative types of behavior. revised: 1993-04-07 kw

External (incoming) Involved object types

Request to register a submitted CONFERENCE, PROGRAM_COMMITTEE, PAPER

paper

Request to accept a paper CONFERENCE, PROGRAM_COMMITTEE, PAPER,

STATUS

Request to assign a paper to a CONFERENCE, PROGRAM_COMMITTEE,

session PROGRAM, PAPER, PAPER_SESSION

Selection of a session chairperson CONFERENCE, PROGRAM_COMMITTEE,

PROGRAM, PAPER_SESSION, PERSON

Request to register an attendee CONFERENCE, ORGANIZING_COMMITTEE,

REGISTRATION, PERSON

Request to print conference CONFERENCE, ORGANIZING_COMMITTEE,

attendee list REGISTRATION, PERSON, ATTENDEE_LIST

Figure 5.1 Event chart: incoming external events

Analogously, the event chart gives the same information for a list of important
internal outgoing events as shown in figure 5.2. Since all outgoing events are
indirectly triggered by incoming events, the outgoing events listed all have one
or more corresponding incoming external events.
For example, the outgoing event “call for papers is sent” was probably
triggered by an incoming event “request to send call for papers” resulting from
user keyboard input. Or else the calls may be sent automatically at some preset
date, but then the system clock interrupt may be considered as the incoming
external triggering event. Similarly, each incoming external event usually has a
corresponding outgoing event. A request to register a conference participant will
almost certainly yield some kind of confirmation being sent back to the user—at
least indirectly by issuing a standard system prompt as opposed to an error
message—indicating that the registration was successful.
In either case, both the incoming and the outgoing events will point to the
same scenario, so there is normally no need to list both related events in the
96 THE DYNAMIC MODEL

EVENTS CONFERENCE_SUPPORT Part: 2/2

COMMENT INDEXING
Selected internal events triggering system created: 1993-02-15 kw
responses leaving the system. revised: 1993-04-03 kw

Internal (outgoing) Involved object types

Call for papers is sent CONFERENCE, ORGANIZING_COMMITTEE,

PERSON, MAILING

Invitations are sent CONFERENCE, ORGANIZING_COMMITTEE,

PERSON, MAILING

A paper is sent to referees CONFERENCE, ORGANIZING_COMMITTEE,

PAPER, STATUS, REVIEW, PERSON

An invoice is sent CONFERENCE, ORGANIZING_COMMITTEE,

REGISTRATION, PERSON, INVOICE,
INVOICE_FORM

Warning issued for exceeding CONFERENCE, REGISTRATION, TUTORIAL

tutorial session capacity

An author notification is sent CONFERENCE, PROGRAM_COMMITTEE,

PERSON, PRINT_OUT*, LETTER_FORM

Figure 5.2 Event chart: outgoing internal events

event chart. Outgoing events like “warning issued for exceeding conference
capacity” are different, since these are triggered when the system state reaches
certain values. Such triggering states are of course also the indirect result of
incoming events, like trying to register one more attendee, but it is not always
easy to know exactly which ones, so we normally record this group of outgoing
events separately.

System scenarios
A system scenario is a description of a possible partial system execution. It can
be viewed as a sequence of events initiated by one or more triggering events
(internal or external) and showing the resulting events in order of occurrence.
Some of the events in a scenario will usually be external, but not always.
Particularly during design there may be many interesting, purely internal
scenarios that are worth capturing as part of the high-level system description.
Anyway, the great majority of events in most scenarios will be internal events;
that is, generated by the system itself. As we recall from the beginning of this
chapter, object-oriented system execution is really nothing but message passing
between objects, so all events except the incoming external ones are caused by
 DYNAMIC CHARTS 97

operations being applied to objects. Some of these operations change the system
state (the commands), and some of them do not (the queries).
When describing the internal events of our scenarios, we basically have two
choices:
1. View the system as a finite state machine and let the events correspond to
transitions between certain interesting system states.
2. View the system as a set of communicating objects, and let the events
correspond to message passing between them.
Most books on analysis and design containing any kind of elaborate notation for
dynamic modeling seem to take the first approach, and then use some kind of
state diagramming technique. However, this approach immediately runs into
trouble, because there is no natural mapping from the state diagrams to the static
model of the system classes.
BON instead takes the second approach, viewing internal events as operations
applied to objects. Then every message passed from one object to another can be
directly related to the corresponding classes and operations specified in the static
model. So rather than representing different worlds with no logical connection
between them other than the poor developer’s brain, as is the case with state
machine approaches, the static and dynamic models can now reinforce each
other and lead to a better understanding of the system under construction.
We will see in section 5.3 what dynamic diagrams with message passing may
look like.

Scenario charts
Using the system event chart as a starting point, we may select a set of
interesting system scenarios to illustrate important aspects of the overall system
behavior. A short description of each scenario (for which a dynamic diagram
may later be produced) is then collected in the system’s scenario chart. An
example is figure 5.3. Each entry has two short descriptions: the first is just a
few words which can also be used as a scenario name for quick reference, and
the second a few sentences explaining the scenario a little more.

Object creation charts

Investigating the creation of new objects may serve as a link between the static
and dynamic models. Some classes in a static diagram may be marked as
deferred, which means they contain operations that will never be implemented in
the class itself, but only in descendant classes. Such classes will never be
instantiated, so objects of the corresponding type cannot occur in the dynamic
98 THE DYNAMIC MODEL

SCENARIOS CONFERENCE_SUPPORT Part: 1/1

COMMENT INDEXING
Set of representative scenarios to show created: 1993-02-16 kw
important types of system behavior.

Send out calls and invitations:

Using mailing lists and records of previous conference attendees and speakers, prepare and
send out calls for papers and invitations to attend the conference.

Create sessions and chairs:

Partition the conference into sessions of suitable length; allocate session rooms and select a
chairperson for each session.

Register paper and start review process:

A paper is registered and three referees are selected; the paper is sent to each referee, and
the paper status is recorded.

Accept paper and notify authors:

A submitted paper is selected and an acceptance date is entered; a notification letter is
created and sent to the authors.

Assign paper to session:

A session suitable for the paper is selected and the paper is entered in the list of
presentations for that session.

Register attendee:
An attendee is registered with his/her address and selected tutorials are recorded.

Print conference attendee list:

All registrations are scanned and a list with attendee names, addresses and affiliations is
produced and sent to a printer.

Print badge:
An attendee is selected, and the corresponding badge is printed in appropriate format.

Figure 5.3 Scenario chart for a conference system

model. All other classes, however, must potentially have objects created at some
point during system execution, otherwise they are superfluous and should be
removed (unless, of course, we are developing libraries of reusable classes).
Thinking through how objects are created may thus help find possible fossil
classes, but it also helps the developer form an impression of how some of the
operations in the dynamic diagrams may be realized. The idea is to produce an
object creation chart, where for each class that may create other objects, the
types of these objects are listed. (Only high-level analysis classes are considered
here; keeping track of created lower-level objects is not the intention.) An
example for the conference system is shown in figure 5.4.
The class PRESENTATION in the creation chart is a deferred class with the
classes PAPER and TUTORIAL as descendants. Note that deferred classes may
occur in the left column of object creation charts, because a deferred class may
contain operations that create other objects. Regardless of whether these
 DYNAMIC CHARTS 99

CREATION CONFERENCE_SUPPORT Part: 1/1

COMMENT INDEXING
List of classes creating objects in the system. created: 1993-02-18 kw

Class Creates instances of

CONFERENCE PROGRAM_COMMITTEE, TECHNICAL_COMMITTEE,

ORGANIZATION_COMMITTEE, TIME_TABLE

PROGRAM_COMMITTEE PROGRAM, PAPER, PAPER_SESSION, PERSON

TECHNICAL_COMMITTEE TUTORIAL, TUTORIAL_SESSION, PERSON

ORGANIZATION_COMMITTEE MAILING, ADDRESS_LABEL, STICKY_FORM,

REGISTRATION, PERSON, INVOICE, INVOICE_FORM,
ATTENDEE_LIST, LIST_FORM, POSTER_SIGN,
POSTER_FORM, EVALUATION_SHEET,
EVALUATION_FORM, STATISTICS

PRESENTATION * STATUS, PERSON

PAPER REVIEW, ACCEPTANCE_LETTER, REJECTION_LETTER,

LETTER_FORM, AUTHOR_GUIDELINES

TUTORIAL ACCEPTANCE_LETTER, REJECTION_LETTER,

LETTER_FORM

REGISTRATION CONFIRMATION_LETTER, LETTER_FORM, BADGE,

BADGE_FORM

Figure 5.4 Object creation chart

operations will actually be implemented by the deferred class or by its children,

if we know their invocation may lead to the creation of objects of certain types,
this information should not have to be duplicated in all descendant classes. By
listing the classes STATUS and PERSON in the entry for PRESENTATION, we
can avoid repetition for the child classes PAPER and TUTORIAL, and may
instead focus on the differences.
It is usually best to exclude frequently reused library classes, such as SET,
TABLE, and DATE, from the creation chart, since it is rarely interesting to follow
the creation of such objects in detail.

5.3 DYNAMIC DIAGRAMS

Objects
We are now ready to describe the BON dynamic notation used in system
scenarios. A dynamic diagram consists of a set of communicating objects
100 THE DYNAMIC MODEL

passing messages to each other. An object is represented by its type, that is its
class name, and an optional object qualifier.

Graphical representation
Objects are represented graphically by rectangles to make them stand out from
the class ellipses. An effort is made in BON to make the dynamic and static
diagrams look different enough to preclude any confusion between them. There
is a deep semantic difference between a class and its instances that is sometimes
very obvious, but may at times be quite subtle, depending on the discussion
context and the general backgrounds of the people trying to communicate.
Mixing statements about classes with statements about class instances in the
same context is akin to mixing language and metalanguage in the same phrases;
it can easily lead to severe misunderstandings. Static and dynamic descriptions
should therefore be kept strictly apart and not be allowed to share the same
diagrams.
The name of the corresponding class is centered inside the rectangle in upper
case (bold italic font recommended). A single rectangle in a diagram always
refers to an individual object, so two single rectangles with the same name will
refer to two individual objects of the same type. A qualifying identifier may be
written below the class name, enclosed in parentheses (lower case
recommended), to distinguish several objects of the same type occurring at
different positions in a dynamic diagram.
An object rectangle may be double, in which case it refers to a set of objects of
the corresponding type. Passing a message to a double object rectangle means
calling some object in the set. Three single objects and one set of objects are
shown in figure 5.5.

CONTROL_TOWER FLIGHT

QUEUE QUEUE
(arrivals) (departures)

Figure 5.5 Objects are identified by class name, context, and optional id

Additional object qualifiers should only be used when really needed, since
having a lot of them may easily clutter a diagram and lead to more confusion
than help. Even if there is more than one object of a certain type in a diagram,
individual identification is not always interesting, or else the spatial context may
be enough to tell which object is which.
 DYNAMIC DIAGRAMS 101

Message relations
In this book, we switch freely between the message passing metaphor and the
feature call metaphor (objects invoking operations on each other) depending on
what view is most convenient for the particular issue being discussed. We may
thus at times talk about sending a message to an object, calling a class feature, or
invoking an operation on an object, but this will always mean the same thing.
For the relationship between a calling object and the callee, we use the term
message passing relation or message relation for short.

Graphical representation
A message sent from one object to another is represented graphically by a dashed
arrow extending from the calling object to the receiving object. The message
arrows (or message links) may be labeled with sequence numbers. These
numbers serve a double purpose. First, they represent time in the scenario, that is
the order in which calls are made. Second, they correspond to entries in a
scenario box where the role of each call may be described using free text.
In this way dynamic diagrams can be kept clean and easy to read, while the
natural language text in the scenario boxes provides the freedom to express
special semantic details at greater length when needed. The scenario box has a
header containing a scenario name, which is the corresponding short entry of the
scenario chart. An example is shown in figure 5.6.

Scenario: Borrow a car and go for a ride 1, 6

OWNER
1 Driver obtains keys from car owner
2 Driver turns ignition key
3 Engine starts DRIVER
4 Driver removes ignition key
5 Engine stops 2, 4 3, 5
6 Driver returns keys to owner CAR ENGINE

Figure 5.6 A BON dynamic diagram

A message relation is always potential; that is, we cannot tell from the
diagram whether the call will actually occur in a specific execution. For
example, if the car battery is flat, turning the ignition key may not lead to any
message being sent from the CAR object to the ENGINE object. Also, obtaining
keys from a car owner (label 1 in figure 5.6) may involve asking several owners
or the same owner several times before someone is persuaded to lend their car.
For reasons of simplicity and abstraction, we exclude all forms of conditional
control (such as multiple choice and iteration) from the BON dynamic notation,
102 THE DYNAMIC MODEL

and instead let it be implicit in the message relations. For the same reasons, we
also exclude any data flow indications. Some messages will implicitly cause
return values to be passed in the reverse direction, but this is not shown in the
diagrams. In case we need to express more, separate dynamic notations may be
used as a complement. The path expressions for high-level descriptions of
patterns of interaction between object groups described in [Adams 1994] may be
a candidate.

Bidirectional links
A set of message relations in each direction between two objects (or sets of
objects) may be expressed by a double link with an arrowhead at each end, like
in figure 5.7.

2 1
Scenario: Time watch HUNGRY_PERSON CLOCK
1 Notification at noon requested 4
2 Clock sends “time for lunch” message
3
3 Hungry person tells colleagues
4 Colleagues respond PERSON

Figure 5.7 Bidirectional message links

As in the static notation, this is just a compacted form of individual relations,

and we employ the same convention to show which direction each sequence
label corresponds to. A label is always put closer to the supplier object, that is
the object being called (the receiver). In figure 5.7, label 1 therefore represents a
call from the hungry person to the clock, and label 2 refers to the clock
responding. Similarly, message 3 refers to the hungry person summoning the
lunch team members, while message 4 represents their answers.

Multidirectional links
To simplify diagrams with many relations, we allow message links to be joined,
as shown in figure 5.8. A joined message link must have at least one calling

LEG
Scenario: Take a walk 1
BRAIN
1 Legs are told to move
LEG

Figure 5.8 Concurrency: send to many

 DYNAMIC DIAGRAMS 103

object and one receiving object attached, and it represents potential calls from all
the callers to all the receivers. Placing a label close to a caller means that it
refers to messages from this caller to all receivers in the diagram.
Joining multiple calls from the same caller, as in figure 5.8, is also a way of
grouping concurrent messages: we may not be interested in whether the right or
the left leg takes the first step, as long as it leads to a useful walk. (The notation
may mean true concurrency, or simply that we do not care.)
Similarly, placing a label close to a receiver refers to messages from all callers
in the diagram to this object (or set of objects). This is illustrated by figure 5.9,
where a pilot needs to receive clear signals from various sources both inside and
outside the airplane before taking off, but where no particular order is required.

CABIN_CREW
Scenario: Get ready for takeoff
1
1 Pilot receives “seat belts fastened”, PILOT
“cabin doors secured”, and “runway clear”
CONTROL_TOWER

Figure 5.9 Concurrency: receive from several

In figure 5.10, labels 1 and 2 represent messages from one parent to both
children, while labels 3 and 4 represent messages from both parents to one of the
children.

Scenario: Farewell at train station 1 4

FATHER DAUGHTER
1 Father gives children an extra $50 each
2 Mother gives children boxed lunches
3 Parents tell son never to ski out of bounds
2 3
4 Parents tell daughter what men are really after MOTHER SON

Figure 5.10 Concurrency: multiple send to many

Object grouping
We have seen that the simple labeling and joining conventions of the dynamic
diagrams introduced so far already give us the possibility to express many
different types of object communication. But we may take a significant step
further, which is particularly important for scenarios containing a larger set of
interacting objects. This step is the ability to send messages between groups of
objects.
104 THE DYNAMIC MODEL

An object group is a set of objects treated as a unit in some message passing

context. An object group may contain single objects (corresponding to single
rectangles), sets of objects (corresponding to double rectangles), and also other
object groups. A message link to an object group is simply the compression of
one or several message links to one or several of the elements in the group.
Similarly, a message link from an object group is the compression of one or
several message links from one or several of the group elements.

Graphical representation
Object groups are drawn as boxes with rounded corners representing the group
body with an optional group label in a separate rounded box. The group label (if
present) is positioned just outside the group body touching its borderline. BON
recommends letting dynamic group labels start with a capital letter with the rest
in lower case, to make them stand out from static cluster labels. A group label
may appear anywhere along the group borderline to facilitate the presentation of
complicated diagrams. Some examples are shown in figure 5.11.

Callers Receivers

A B RECEIVER A B A B

1
2

CALLER A B C D A B

Receivers Callers

Figure 5.11 Compressed message relations

The line style must be different from the one chosen for objects, and
preferably also from classes and clusters. The recommended practice is to use
continuous borderlines for objects, and to use dotted borderlines for object
groups and group labels.

Compression of objects and groups

An object group is compressed into its group label attached to an empty,
iconized body (see figure 5.12). Unlabeled object groups cannot be compressed.
When a group icon is expanded (or opened), its constituent parts first become
 DYNAMIC DIAGRAMS 105

Beach_surveillance

LIFE_GUARD Equipment

WATCHTOWER

Figure 5.12 An iconized object group

visible in compressed form, thus showing objects and other group icons. The
inner parts may then be opened recursively, until the objects are reached.

Messages involving object groups

Message relations between callers and receivers may be compressed as follows:
1. If one or more elements (objects, sets of objects, or groups) in a group
send messages to an element outside the group, all the message links
involved can be compressed into one message link from the group to the
outside element.
2. If one or more elements (objects, sets of objects, or groups) in a group
receive messages from an element outside the group, all the message links
involved can be compressed into one message link from the outside
element to the group.
3. A message link between two elements can be compressed into being
hidden (not shown in a diagram).

Compression of message relations

When object groups are being compressed or expanded in a dynamic diagram,
the labels and corresponding entries in the scenario boxes may have to be
renumbered to reflect the overall message sequencing. Case tools may employ
different strategies as long as they comply with rules 1−3 above.

Recursive messages
It is often important to express recursive message passing in a dynamic diagram.
Figure 5.13 shows three different cases. The first case represents a specific
object calling itself, and is therefore drawn using a single rectangle. The second
case shows a set of objects of a certain type calling objects in the same set.
106 THE DYNAMIC MODEL

CALLER CALLER CALLER A

Receivers

Figure 5.13 Recursive messages

Whether individual objects actually call themselves is not specified. Finally, the
third case uses grouping to express recursive calls to a group of objects which
may include the caller itself.

Active and passive objects

The objects in the dynamic model correspond exactly to the classes in the static
model, but the structures of static and dynamic diagrams describing the behavior
of a set of objects are often very different. The client relations in the static
diagrams are of course closely related to the message relations in the dynamic
diagrams; each call to an object is applied to an entity associated with the object,
and it is precisely these entities that indirectly define the static client relations.
Whether the object association is an address pointer produced by a compiler and
loader in a tightly integrated system running as a single task, or a symbolic
object identification used by a series of object brokers or traders to communicate
calls in a distributed system, is irrelevant to the discussion.
Regardless of the calling mechanism used, static diagrams show the access
structure to the services needed, while dynamic diagrams show control. The
difference is illustrated in the next two figures. Figure 5.14 shows an agent
dealing with real estate who has access to a repository of sales leads, each listing
a prospective buyer and the buyer’s address.

ESTATE_ info leads: SET […] addr

REPOSITORY BUYER ADDRESS
AGENT

Figure 5.14 Static diagram: estate sales data

The static diagram is a typically layered information structure. By contrast, a

corresponding dynamic diagram is shown in figure 5.15, where the scenario is to
produce a list of addresses of possible buyers. We see that the structure here is
completely different. Instead of showing layers of information, the dynamic
 DYNAMIC DIAGRAMS 107

BUYER

3
Scenario: List prospective buyer addresses
1 Get leads
4 ESTATE_ 2
2 Get next buyer in set of leads ADDRESS SET
3 Get address from buyer AGENT
4 Ask address to print itself
1

REPOSITORY

Figure 5.15 Dynamic diagram: active and passive objects

diagram shows the estate agent as the only caller in the scenario, while all other
objects simply act as data containers.
An object sending a message is said to be active, while the receiving objects
are passive. Being active or passive is a role; to carry out its task when called as
a passive object, the called object may have to call other objects and thus become
active. Many objects, perhaps most, will thus act in both capacities at different
times. However, it is often interesting in a dynamic context to separate the
objects playing the active parts from the passive ones, since the former are the
ones causing things to happen. In BON dynamic diagrams, this is immediately
visible through the occurrence and direction of the message relations.
When building a dynamic scenario with the aid of a corresponding static
diagram, it is then very important to realize that an indirect call chain to print the
address of the next buyer in the structure shown in figure 5.14, which can be
written:
info.leads.item.addr.print
does not mean that ESTATE_AGENT calls REPOSITORY, which calls SET,
which then calls BUYER, which in turn calls its ADDRESS object. Instead,
ESTATE_AGENT calls REPOSITORY to get a reference to a SET of leads.
Using the reference, ESTATE_AGENT then makes a second call to the SET
object to get the next BUYER object, then a third call to BUYER getting its
ADDRESS object, and then a last call to this ADDRESS asking it to print itself.
This is what is really expressed by the call above. Nested calls of this type are
only allowed for syntactic convenience, and are semantically equivalent to:
info: REPOSITORY
s: SET [BUYER]
b: BUYER
a: ADDRESS
108 THE DYNAMIC MODEL

s := info.leads
b := s.item
a := b.addr
a.print
If we use local variables for implementation as above, the corresponding static
structure becomes the one in figure 5.16. However, we normally do not want
this reflected in our high-level static diagrams, since it only makes them more
difficult to read. Therefore, an essential property of a case tool is to permit the
hiding of client dependencies resulting from local variables or private features.
a

ESTATE_ info leads: SET […] addr

REPOSITORY BUYER ADDRESS
AGENT

s: SET […] , b

Figure 5.16 Modified static diagram

A more complex example: moving graphical objects

To give the reader a general idea of the potential strength of the BON dynamic
notation, we will spend the rest of this chapter on two examples where the object
interaction is a little more complex.
The first example will show how the notation can be used to express dynamic
binding. Assume we want to model a simple graphical editor with grouping. A
static diagram for the editor is shown in figure 5.17. It can handle three types of
graphical object: pixel map objects for arbitrary images, graphical text objects
with variable font and point size, and geometric objects consisting of a set of
defining points and a shape. All the geometric figures inherit from the deferred
class FIGURE, which is associated with a set of graphical points.
Selected objects may be grouped and treated as a unit, for example when
moved, rotated, or resized. To be of any use, groups may also contain other
groups besides primitive objects. This is conveniently expressed in figure 5.17
by letting all graphical objects (including the FIGURE objects) inherit from a
second deferred class DISPLAY_OBJECT, and then letting the GROUP class
have an aggregation relation to a set of DISPLAY_OBJECT.
Since a DISPLAY_OBJECT can be either a primitive graphical object or a
group, we get the required recursion for free through dynamic binding.
Combining generic derivation with inheritance in this fashion is a very common
 DYNAMIC DIAGRAMS 109

*
DISPLAY_ {
POINT
OBJECT
members: SET […]
points: SET […]

PIXEL_
TEXT GROUP
MAP

LINE

* ELLIPSE RECTANGLE
FIGURE

CIRCLE SQUARE

GEOMETRIC_FIGURES

DISPLAY_ELEMENTS

Figure 5.17 Graphical objects with grouping (static diagram)

and powerful method to handle complex structures with object-oriented

modeling. We also note that by using clustering and compression of inheritance
links it is possible to express quite entangled classification structures with very
few arrows. Without the compression possibility, this diagram would be at least
an order of magnitude more difficult to read.
Now let us try to illustrate a dynamic scenario where a group object is being
moved. A group created by the editor is shown in figure 5.18.

Hello

Figure 5.18 Graphical editor example: a grouped object with nesting

110 THE DYNAMIC MODEL

The group contains a nested group and two geometric objects: a square and a
circle. The inner group contains one smaller circle and a graphical text object.
The thick, dashed boxes just illustrate the grouping, and are not part of the
graphical figures.

The need for two-dimensional object diagrams

The typical way to show object communication found in other approaches, such
as OMT [Rumbaugh 1991], the Booch method [Booch 1994], and OOSE
[Jacobson 1992], is through one-dimensional sequencing diagrams called event
trace diagrams or interaction diagrams. Figure 5.19 shows a diagram of this
type for our small example, without the graphical points that are part of the
geometric figures.

OUTER OUTER OUTER INNER INNER INNER

GROUP SET SQUARE CIRCLE GROUP SET CIRCLE TEXT

next group member requested

square asked to move
outer circle asked to move
inner group asked to move
next inner group member requested
inner circle asked to move
text asked to move

Figure 5.19 Move example group: event trace diagram

The eight objects involved (excluding the points)—two groups, two set objects
containing group elements, and four primitive figures—just about hit the limit
for what can be expressed with reasonable readability in one-dimensional
diagrams. Including the points would clutter the figure. Considering the
simplicity of the example, it is clear that diagrams of the type represented by
figure 5.19 are too weak for dynamic modeling. This may be compared with
figure 5.20 which uses the two-dimensional BON notation.
Freeing the time axis has made the actual communication diagram smaller,
while moving the temporal explanations to a separate place. The spatial
positioning of objects in two dimensions permits them to be locally grouped.
We have therefore been able to drop the qualifying identifiers for the SET and
CIRCLE objects, since the geometric contexts give enough information.
 DYNAMIC DIAGRAMS 111

Scenario 1: Move example group

1 Next outer group member requested
2 Square asked to move
3 Outer circle asked to move
4 Inner group asked to move
5 Next inner group member requested
6 Inner circle asked to move
7 Text asked to move

SET SET
1 5

2 GROUP 4 GROUP 7
SQUARE (outer) (inner) TEXT

3 6

CIRCLE CIRCLE

Figure 5.20 Move example group: two-dimensional BON diagram

The increased space available now permits us to go one step further to

illustrate the movement of geometric objects in more detail. We assume that
such an object is moved by moving, in turn, its defining points (a circle may be
defined by two points: its origin and one point on the circumference).
Figure 5.21 shows the resulting diagram. It now contains 11 single objects and 3
object sets without looking too crowded, while the corresponding event trace
diagram would no longer be readable.

General scenarios need object grouping

What is immediately striking, however, is that the scenario feels very special;
instead of showing individual move examples, one would want to capture the
general scenario of moving any group. To do this, we must employ the grouping
facility. A first attempt is illustrated in the upper diagram of figure 5.22.
Collecting the geometric figures in one group and the non-geometric ones in
another,6 makes it possible to illustrate the various cases of dynamic binding that
may occur. Call 1 of scenario 3 gets the next group member. Applying the

6
We assume it is clear from the context when the word group refers to an object group in the
dynamic notation, and when it denotes a group in the editor example.
112 THE DYNAMIC MODEL

Scenario 2: Move example group Scenario 2: (continued)

1 Next outer group member requested 8 Inner group asked to move
2 Square asked to move 9 Next inner group member requested
3 Next square point requested 10 Inner circle asked to move
4 Point asked to move 11 Next inner circle point requested
5 Outer circle asked to move 12 Point asked to move
6 Next outer circle point requested 13 Text asked to move
7 Point asked to move

SET SET SET

3 1 9

4 2 GROUP 8 GROUP 13
POINT SQUARE TEXT
(outer) (inner)

5 10
6 11
SET CIRCLE CIRCLE SET

7 12

POINT POINT

Figure 5.21 Move example group: extended to include points

move operation on this member then has to be split into calls 2−6, since different
things happen depending on the type of member. In fact, only one of call 2, call
3, or calls 4−6 will occur for a given member.
The diagram may be further improved as shown in scenario 4 (lower part of
figure 5.22) by nesting the groups that hold the polymorphic objects. We then
get a diagram which illustrates the dynamic binding of the move operation very
directly. If we compare scenario 4 with the static description in figure 5.17, we
see that the similarity is striking.
The deferred classes DISPLAY_OBJECT and FIGURE, which are never
instantiated, are replaced in the dynamic diagram by two object groups (outer
and inner). Moreover, the inheritance relations between the pairs
CIRCLE−ELLIPSE and SQUARE−RECTANGLE, respectively, are not shown in
the dynamic diagram. It would have been possible to use one more layer of
nested groups to include them, but when geometric shapes are manipulated,
circles and ellipses are usually not perceived as related objects, and the same
holds for squares vs. rectangles. Therefore, the scenario looks more natural
without this last grouping.
 DYNAMIC DIAGRAMS 113

Scenario 3: Move any group

1 Next group member requested
2 Subgroup member asked to move
3 Non-geometric figure asked to move
4 Geometric figure asked to move
5 Next point of geometric figure requested
6 Point asked to move

TEXT 3 4
GROUP LINE CIRCLE SQUARE

1
PIXEL_ RECTANGLE ELLIPSE
MAP 5
SET

POINT

Scenario 4: Move any group

1 Next group member requested
2 Group member asked to move
3 Next point of geometric figure requested
4 Point asked to move

SET POINT

1 3 4 Display_elements

2
GROUP

LINE CIRCLE SQUARE

TEXT

RECTANGLE ELLIPSE
PIXEL_
MAP

Figure 5.22 Polymorphism: needs object grouping

114 THE DYNAMIC MODEL

A different use of object grouping

Finally, we show a second example where grouping is useful, but where the
objects called are not necessarily related through common inheritance.
Figure 5.23 contains a typical scenario that can be divided into sequential
subtasks with some shared data between them.
It shows the general interactions needed to deal with an insurance company as
a result of a car accident without personal injury. Getting the car fixed may be
divided into three phases. First, various reports are filled out to document what
really happened and establish damages for which the insurance holder may be

Scenario 5: Settlement of claims for damages resulting from car accident

1−3 Owner obtains necessary statements and certificates from involved parties,
fills in damage report, and sends it to insurance company.
4−7 Insurance adjuster evaluates damage claims and sends settlement statement
back to owner.
8−9 Owner agrees on car rental and repair details based on settlement.

Evaluation

Accident_report INSURANCE

WITNESS STANDARD_
PRICE_LIST
POLICE

APPROVED_
INSURANCE_ GARAGES
INSPECTOR

Repair 5
1

RENTAL_ 9 7 3 INSURANCE_
OWNER
COMPANY ADJUSTER

GARAGE 2 4
DAMAGE_REPORT

8 6
SETTLEMENT

Figure 5.23 Grouping into sequential subtasks

 DYNAMIC DIAGRAMS 115

entitled to compensation. These reports are based on statements from people

being involved or witnessing the accident, and may need formal certification by
an insurance inspector and perhaps the police. This is captured by message links
from the OWNER object to the object group labeled “Accident_report”, to the
DAMAGE_REPORT object, and to the insurance company, representing formal
submission of the claims.
Second, an adjuster at the insurance company makes an evaluation of the
claims based on the accident report and the general terms of the applicable
insurances. A maximum cost is often specified based on agreed standards on
acceptable repair charges for various types of damage, and sometimes the
insurance company will only accept certain trusted garages to do the job. The
message link from INSURANCE_ADJUSTER to the “Evaluation” object group
illustrates this. The result of the evaluation is a formal settlement statement from
the insurance company, approving or disapproving parts or all of the owner’s
claims (message links to SETTLEMENT and back to OWNER).
Based on the outcome of this, the owner may then carry out the third phase of
the scenario, which is to negotiate with a suitable garage and perhaps also rent a
car to replace the one under repair (message link to “Repair” group).
This example is very different from the graphical editor for two reasons. First,
the messages sent to the objects in the groups may represent operations that are
totally unrelated and therefore do not mirror any inheritance structure in the
system. Second, the object groups may not correspond to any kind of grouping
in the static model of the system, but simply represent a convenient way of
viewing a somewhat complicated scenario to make it easier to understand.
There are many such cases in practice, where a system—perhaps consisting of
a number of loosely connected heterogeneous components as is becoming more
and more common in distributed environments—is used in a way that creates
abstractions that were never part of the system model. In fact, an important
characteristic of really useful systems is often that they provide a certain number
of basic facilities that the user can combine in different ways to create usage
patterns that the system designers could never anticipate (and should not try to).

Use cases
So system usage can give rise to abstractions of its own that may be worth
modeling, but we must also issue a word of warning in this context. Systematic
elaboration of use cases, as advocated in OOSE [Jacobson 1992], is often a good
way to gain more insight into what is really needed in a system. But using them
as the basis for system decomposition risks ending up with a product whose
structure reflects system usage more than long-term concepts. We will return to
this issue in chapter 6.
116 THE DYNAMIC MODEL

For these reasons, the dynamic diagrams in BON are normally not as stable as
the static model, and will frequently change as we gain more insight into the
problem. In fact, some dynamic diagrams will just serve as temporary help to
evolve the static model, and will be thrown away at later stages. Others may
survive longer and fill a role as complementary system documentation.
 Part III
The method
118 PART III
6 Issues of a general
method

6.1 INTRODUCTION
Developing high-quality professional software requires great intellectual skills
and long experience. It is a creative process that can never be replaced by the
rigid procedures of an ultimate “methodology”. Unless we are working in a very
narrow application area, the possible variations in system views make the task of
finding a good model too close to general human problem solving, which just
does not lend itself to any useful formalization.
Despite the fact that everybody in the trade knows this (except perhaps for
some bubble and arrow specialists who have never really been forced to
implement their own diagrams), there is a remarkable gullibility in many parts of
the software community. Somewhat like instant cures for baldness, almost any
kind of simplistic ad hoc procedure seems to sell if it only promises to turn
general systems development into an orderly, controllable procedure. A tailored
variation of the generic approach in figure 6.1 is then often used.

The universal method M (apply recursively when required)

1. This is the important beginning step, where we solve the first part of the problem.
2. Welcome to the middle of the method: in this step we solve the second part.
3. Now all that remains is solving the last part, and we will soon be done.

Figure 6.1 The archetype of many software “methodologies”

So the manager’s secret dream of totally manageable (and easily

exchangeable) employees for software development is still far off, and the reader
should not expect any miracles from the BON method either in this respect. Its
goal is to provide the developer with a strong set of basic concepts and
corresponding notation to be used as a framework for creative thinking,

119
120 ISSUES OF A GENERAL METHOD

complemented by a set of guidelines and recommended procedures that may help

in the search for good models. BON can never substitute for talent, but it can
hope to serve as an inspiration for developing one’s personal skills.
The reader may also have noted that we use the term method for these
guidelines and procedures, rather than the more popular buzzword
“methodology”. This is for two reasons. First, using a longer name does not add
any substance to a method, and second, according to available dictionaries the
latter form simply does not have the intended meaning:
method, n.
1. a way of proceeding or doing something esp. a systematic or regular one.
2. orderliness of thought, action, etc.
3. (often pl.) the techniques or arrangement of work for a particular field or subject.

methodology, n.
1. the system of methods and principles used in a particular discipline.
2. the branch of philosophy concerned with the science of method and procedure.
(Collins English Dictionary, 3rd edn, 1992)
Since BON is presenting a view of its own rather than trying to unify what
everybody else in the field is doing, we will discuss a method in this book and
not worry too much about methodology. The BON method will introduce a
recommended analysis and design process consisting of nine major tasks to be
carried out. Each task uses a set of input sources and produces a set of well-
defined deliverables (either new or updated) as part of the end result.
To complete the tasks, a developer will be involved in a number of standard
activities (also nine as it happens), some of them more related to analysis
(problem domain) and some more related to design (solution space). These
activities may occur in varying degrees during several of the tasks, and are
therefore considered orthogonal to the process.
This chapter will discuss some issues that are considered important in the
BON method for analysis and design. The next chapter will then describe the
nine major tasks of the BON process: what should be performed and which
deliverables to produce. Finally, chapter 8 will discuss the typical analysis and
design activities that will occur as part of carrying out the individual tasks of the
process.

6.2 SIMULATING AN IDEAL PROCESS

Anyone who has been involved in intellectually taxing activities—trying to
understand and solve a complex problem—knows that the process of arriving at
a good solution is far from regular. On the contrary, the most common
 SIMULATING AN IDEAL PROCESS 121

impression during the course of the effort is often a sense of total disorder and
utter confusion. This is also true for cases where the final result eventually turns
out to be very simple and elegant, and the greatest sense of confusion is often
experienced shortly before the crucial perspective is discovered.
So the bad news is that a rational modeling process, where each step follows
logically from the previous ones and everything is done in the most economic
order, does not exist. Complex problem solving just does not work that way.
But the good news is that we can fake it. We can try to follow an established
procedure as closely as possible, and when we finally have our solution
(achieved as usual through numerous departures from the ideal process), we can
produce the documentation that would have resulted if we had followed the ideal
process [Parnas 1986].
This gives us a number of advantages:
• The process will guide us, even if we do not always follow it. When we
are overwhelmed by the complexity of a task, it can give us a good idea
about how to proceed.
• We will come closer to rational modeling if we try to follow a reasonable
procedure instead of just working ad hoc.
• It becomes much easier to measure progress. We can compare what has
been produced to what the ideal process calls for, and identify areas where
we are behind (or ahead).
Such an ideal process is part of BON and will be presented in the next chapter,
but first we will discuss some general modeling issues.

6.3 ENTERPRISE MODELING

Object-oriented analysis may be used as the top-level instrument for the design
and implementation of software, but its modeling capabilities extend far beyond
that. Any type of scenario consisting of a number of actors exhibiting certain
behaviors is potentially a candidate for object-oriented modeling, regardless of
whether computers will be involved at some point or not.
For example, human communication and mutual understanding within an
organization may be greatly enhanced by creating unified models of company-
wide concepts and tasks and their interactions. This type of analysis, often called
enterprise modeling, may very well turn out to be one of the major areas of
application for object-oriented techniques in the future. In large organizations,
one can hardly overestimate the importance of a common view of the terms and
concepts used as a basis for making decisions in various areas of responsibility.
122 ISSUES OF A GENERAL METHOD

Object-oriented techniques have a great potential for enterprise modeling,

which needs both the abstractional power to model very diverse organizational
entities, and the simplicity to be understandable to a large number of people with
various roles and backgrounds. The metaphor of describing reality as a set of
communicating objects that can be asked to behave in certain ways through a set
of applicable operations seems to be very appealing to many people outside the
computer industry. This is of course no coincidence, since it resembles the way
we normally classify things around us and make decisions by picturing
alternative future scenarios in our minds.
One may argue that enterprise modeling needs more than the basic object-
oriented concepts to be expressive enough. This may very well be true for
complicated cases, but the additional needs are probably quite different for
different types of organization.

6.4 WHAT IS ANALYSIS AND WHAT IS DESIGN?

If we interpret the words analysis and design according to their normal use in
most disciplines, the question in the section title becomes trivial. Analysis
means taking things apart for examination, while design is to work out the
structure or form of something. Both activities are of course needed in all phases
of software development, although different parts and structures may be involved
at different times.
In recent years, however, the terms have gradually started to acquire an
alternative meaning in the object-oriented world. The approximate consensus in
many current books and articles seems to be the following:
• Object-oriented analysis means creating an object-oriented model from the
external requirements (a model of the problem).
• Object-oriented design means refining the analysis model into a high-level
object-oriented model of a solution.
In our opinion, more appropriate terms should be found for the activities above,
since the words analysis and design are much too general and useful to be robbed
of their original meaning. The ambiguity introduced creates even greater
confusion in a field that is difficult enough as it is. However, until something
better catches on, we accept the new interpretation as de facto usage.
In this book, we will therefore use analysis and design in both the new and the
traditional sense. Analysis classes and design classes will refer to abstractions at
different levels in our system (problem vs. solution). Both categories will of
course emerge through analysis of various prerequisite conditions. On the other
hand, the reader should not be surprised to find hybrid statements like “designing
 WHAT IS ANALYSIS AND WHAT IS DESIGN? 123

an analysis model”. The intended meaning of each term should be clear from the
context.

Recursive analysis and design

If we go back to normal language usage, analysis means investigating the details
of something, while design means building a new structure. Whenever we
decide the form of something, we are designing—no matter if we are building a
model for the real-time distribution of X-ray pictures in a wide area network of
hospitals or figuring out the best organization for storing prime numbers. And
every design needs analysis in order to be carried out properly.
Object-oriented technology yields a very flexible recursive model here. First,
the initial requirements are analyzed and an initial model is designed. The class
structure and public interfaces of this model then become a new set of
requirements for the next level of abstraction. The dual application of
analysis / design is then repeated the number of times needed to reach the desired
level. This usually means executable code at some point, but not necessarily.
Since class interfaces only specify how the corresponding objects should
behave—not what internal mechanism may be used to achieve this behavior—
the recursive model is not just theoretically pleasing, but highly realistic and
usable in practice. The idea that any software development should be divided
into exactly three parts (analysis, design, and implementation) is in fact arbitrary.

Multiple language metaphor

Since we are never interested in data per se in a system (only in various
manifestations of it) we should employ the philosophy of the abstract data type.
All data is viewed only indirectly through the behavior of certain black box
objects which can return values or change information. Whether the objects
supplying information actually store the values or just cheat by asking other
objects behind the scenes, the clients do not want to know about.
The collective interface of a set of classes specialized to solve a certain
general task is actually a language tailored to applications in the corresponding
area. Therefore, designing an object-oriented model of a certain problem domain
behavior can be viewed as designing a language in which it is easy to express
that type of behavior. (This perspective was presented already in 1969 by
Edsger Dijkstra in his seminal monograph “Notes on Structured Programming”
[Dijkstra 1972]. There, program development is viewed as the successive design
of languages for a series of abstract machines.)
Once the language is there, it is a simple matter to specify the high-level
behavior and also modify it later, provided we stick to the area which the
language was meant for. Instead of twisting our high-level concepts to fit an
124 ISSUES OF A GENERAL METHOD

obsolete standard language whose limitations we cannot do anything about (the

reality of most programmers for the last 30 years), we can simply change the
rules and invent new specialized languages as needed.
What remains is then to implement the operations of the new language. We
therefore create a second language consisting of classes with operations that are
particularly good at expressing the behavior needed by the operations of the first
language, so they can be easily implemented. This is repeated until the
innermost language can be understood and translated into executable code by
some compiler (or we stop earlier if no execution is desired). Of course, we take
care not to invent languages which are close to existing ones, and all the general
principles for constructing good class libraries apply [Johnson 1988,
Meyer 1994a].
Note that when we talk about multiple languages in the object-oriented sense,
we are referring to semantic content. The syntax of the languages remains
uniform at all levels (feature calls on objects), and seamlessness can thus be
preserved in spite of the great diversity.
So a system model consists of many layers of abstraction, each viewed as a
specialized language. Some of the layers may be hierarchically nested, while
others extend orthogonally across the system. Therefore, no natural borderline
between design and implementation exists; what is considered design will just
have to be decided from one case to another. And what is more—in spite of
many authors giving the opposite impression—there is no clear dividing line
between the analysis and design models either.

What is problem and what is solution?

The world has no a priori structure. A bird flying across a forest looking for a
good place to build a nest will certainly classify the trees very differently from
the lumberjack who is at the same time clearing his way below the foliage in
search of material for his employer’s new line of ship fittings. And if we could
talk to the neurons firing off the classification patterns in the lumberjack’s brain,
they would probably consider the whole concept of tree arbitrary and useless. So
anyone who thinks reality is just there for the analyst to discover is in for a
surprise.
To understand a problem, we need to impose a structure on its parts—but a
good structure is defined by what we want to achieve. If our task is to
computerize some manual work in an organization, we can often start from a set
of concepts and views more or less shared by the people involved. We may have
to study the area for a long time to detect these views, and when we do, it only
means we have a better understanding of the problem and how things are carried
out today. It does not necessarily mean we should keep the views (not even their
 WHAT IS ANALYSIS AND WHAT IS DESIGN? 125

core elements) when transferring to the electronic world. Sometimes traditional

concepts carry over very naturally; sometimes the new opportunities inherent in
automatic processing may change the picture completely.
So we need to create an analysis model whose abstractions make it easy to
understand the desired system behavior, but are still flexible enough to allow
many different realizations. This may result in classes representing tangible
objects, but may just as often result in classes that are abstractions of what users
consider independent concepts of the enterprise or of more general concepts the
users have not even thought about. In fact, physical objects in the problem
domain, which many books recommend as a starting point, are often too special
to be suitable as classes at the highest level in the resulting analysis model (see
for example the case study modeling a video recorder). We should often look for
something beyond the common view held by actors in the problem domain, since
our classification needs are not necessarily the same as theirs.

6.5 REUSE
Reusability should always be our goal, and the aim with shifting to object-
oriented development is that the amount of reuse should increase dramatically
compared with traditional approaches.

Levels of reuse
This increased reuse comes from several levels of abstraction. We can
distinguish at least three such levels:
• The basic level whose components are expected to be needed by almost
any application, regardless of type.
• The domain level covering typical needs in a certain application area.
• The corporate level containing reusable components tailored to the
specific needs of the enterprise.
At the basic level, expected to be part of any good object-oriented language
environment, there are the fundamental data structures of lists, sets, stacks,
arrays, hash tables, and trees that constitute the archetypical “wheel” of software
engineering since they are constantly reinvented day and night, at high cost, by
programmers all over the world using traditional languages.
The basic implementation algorithms are known to any literate programmer,
yet have enough variation to prevent effective reuse in all but the object-oriented
approaches. Large, coherent, and extensively used data structure libraries exist
in Smalltalk and Eiffel, while the hybrid character of C++ and its ad hoc memory
126 ISSUES OF A GENERAL METHOD

management makes the creation and combination of such libraries much harder.
Another prominent component at the basic level is support for standardized
graphical interface elements. Only very special applications can still afford to
repeat the enormous effort invested in GUI toolkits like X/Motif or Microsoft
Windows, yet most users expect graphical interfaces in modern systems.
Higher-level object-oriented models can be built on top of standard toolkits, and
provide developers with the means of creating very sophisticated user interfaces
at very low cost through easy reuse. The class libraries should then support
distinctive user metaphors like the MVC model for Smalltalk [Krasner 1988],
InterViews for C++ [Linton 1989], Nextstep’s Workspace for Objective-C, or
EiffelBuild and EiffelVision for Eiffel [Meyer 1994b].
At the domain level, we should expect in the future to be able to buy or
acquire reusable components for most of the familiar problems connected with a
certain application type rather than developing these ourselves. However, the
offerings so far are quite limited, so we may have to wait some time before this
becomes reality. In the meantime, many such components will have to be
developed and maintained as part of the corporate-level class library. Good
libraries at this level may become the most valuable assets a corporation may
have in the future, since they represent accumulated knowledge that can be
capitalized. This is perhaps especially important in the software business, whose
turnover rate is extremely high.

Reuse policy
Besides trying to locate already existing reusable components, it is also
important to take an active decision regarding the future reusability of the
software under development. Even if the sign of a good object-oriented
developer is in fact a Pavlovian reflex of almost physical nausea whenever a
familiar pattern must be repeated, there is always a short-term cost associated
with building reusable software. This cost may very well be regained already
through increased clarity and reuse within the scope of the project, but not
always.
At the basic reuse level, reusability should always be our first concern, since it
normally pays off immediately. Developing a well-structured set of classes for
building error messages, manipulating general strings, handling command line
arguments, etc., does not cost very much in a good object-oriented environment
(in case they are not already available). Repeating ad hoc solutions over and
over, on the other hand, will not only cost more in the long term, but also hinder
the fundamental shift of attitude so extremely important for the next generation
of programmers: that it is possible to take control instead of just being controlled
by the course of events (see the last section of this chapter).
 REUSE 127

When a developer builds a reusable component that encapsulates a recurring

problem pattern, this is not only an act of reducing future workloads—it also
represents a significant personal investment. By constructing components,
developers gain insight into parts of problem domains that will improve their
future ability to construct good models at all levels. It leads to greater job
satisfaction, since everybody wants to be on top of things instead of just being
used by the circumstances. The increased enthusiasm will wear off on others
through discussions at whiteboards and over coffee, and will gradually lead to a
substantial increase in the so-called human capital currently being recognized by
the economic sciences as the real asset of modern society.
So, we should in fact go for reusability even when it appears to cost a little
more in terms of time and salary, since in the long term it will not. However,
there is a limit to everything and it is particularly important that if we invest in
reusable components, the anticipated reuse should also happen in practice.
Since future reuse of more specialized elements at the corporate or domain
levels depends on many factors, these must be actively addressed before
determining how much should be invested here. If the company’s product
strategy changes, reusable components supporting what was produced yesterday
may quickly be rendered useless. High-quality libraries that remain aesthetic
monuments in a software database make no one happy, least of all the
originators.
Just as for any other project activities we must make a cost / benefit analysis to
decide where we should put our money and effort, and when. For the application
of this economic principle in a software engineering context, see for example
[Boehm 1981].

Planned vs. accidental reuse

There is a debate going on in some object-oriented circles about whether reuse
should be “planned” or “accidental”. The more orthodox school seems to
advocate that we should only allow reuse of software initially built to be
reusable. We find this line of argument very strange.
First, if all reuse were planned beforehand there would not be any point in
calling it “reuse”, since it would just be plain usage. We do not say “Today I
reused my word processor,” because the word processor was built to be executed
any number of times compiling various documents. A class library is no
different; if it was built as a general toolbox open to anyone, when we say we are
“reusing” a class from this library, we really mean “using”.
Second, the fact that the word reuse has a flavor of unplannedness about it is
no coincidence, but an important strategic point. The reason to move from
traditional to object-oriented techniques is not only to get more flexibility for
128 ISSUES OF A GENERAL METHOD

building families of abstractions from scratch, but also to get out of the
straitjacket of deciding too much too soon (cf. the open−closed principle in
[Meyer 1988a]).
In fact, the rapid changes in technology and software requirements guarantee
that we will only know (or even suspect) for a minority of cases what will be
needed in the future. Therefore, restricting reuse to what can be successfully
planned would not leave us much better off than with the old techniques.
The decision whether to reuse a piece of existing software should only be
based on how well it can solve the current problem. Anything else is beside the
point, and we would much rather reuse a good-quality class built without a
thought of ever being used outside its context, than employ a planned “reusable”
component of poor quality.
It is true that developing highly reusable software is not easy, and effective
reuse does not happen by itself just because we now have classes instead of
subroutines. Careless reuse of small heterogeneous pieces here and there may of
course risk losing consistency and miss global patterns. But this is no different
from the general problem with any evolving system. When successive additions
tend to degrade the initial system integrity we must restructure and generalize, no
matter if the added pieces were new or reused.
In practice, so-called accidental reuse will therefore mostly require some
modification to the set of classes in order to make them more general, more
robust, and better documented, but this is just a logical consequence of the
incremental nature of object-oriented development.
Sometimes we know that we are going to need a whole set of related
abstractions, and then we may plan early for a large library of reusable
components. In most cases, however, we simply cannot see this until the system
has evolved for some time. In the meantime, we are much better off with
“accidental reuse” than with no reuse at all.

Reuse manager
Reuse of other people’s products requires trust. If we do not have strong reasons
to believe that the quality and understandability of a reusable class is indeed high
enough, we will prefer to create a new class to gain full control. This is
particularly true in good object-oriented environments where the development of
new abstractions requires so much less effort. The “not invented here” syndrome
may sometimes play a part, but not nearly as often as many people seem to think.
In fact, it is our experience that good developers who are accustomed to
routinely creating reusable abstractions are also very eager to reuse existing
work, as long as they can trust the quality. The standards of a reusable library
must be therefore be high, so that people can always rely on the components and
 REUSE 129

do not need to think twice about using them (provided the functionality fits).
Since the cost of reuse also depends very much on the difficulty of identifying
suitable candidates, easy access to well-indexed class libraries through efficient
class browsers and good general documentation becomes extremely important as
the number of components grows large.
For both these reasons, if we want to promote a corporate database of reusable
components, particularly in large organizations where the people involved do not
all know each other, somebody must be assigned the responsibility of keeping
this database in good shape. The classes must be consistent and adhere to certain
company standards, so it becomes easy to judge whether a particular component
is suitable or not.
We call such a responsible person a reuse manager, and the role should be
assigned to somebody with high skills in three areas: abstract modeling, the
technology and problem domain embodied by the reusable components, and
human communication. Some even claim that without assigning a suitable
person to this role and setting aside a budget for it, reuse on a broad scale within
an organization will just never happen [Goldberg 1991].
Besides administrating the professional components, an important task of the
reuse manager is to take care of potentially reusable software. Very often good
ideas arising in projects are only carried to a point sufficient to cover the needs
of the current development, as time and budget will not permit the investment
needed to create a product usable in more general contexts. The work done may
nevertheless be quite substantial, and it is then important to pave the way for
future capitalization of the effort.
Therefore, a high-quality component library is not enough—there should also
be a less demanding formal collection of software artifacts, which may be called
a reuse repository. Project software with the potential of becoming fully fledged
reusable components, or simply representing general designs that may be of
interest to other projects, should be submitted to the reuse manager for inclusion
in the reuse repository.
The repository must be consistently organized and documented well enough
for users to assess the potentials of its elements. However, the quality of the
individual contributions must be allowed to vary much more than in the
corresponding component library.

6.6 USER CENTERED DESIGN

As computer-based tools are becoming part of the everyday life of a rapidly
increasing number of professionals, the concept of system usability is considered
more and more important. It has even been given an ISO definition [ISO 1990]:
130 ISSUES OF A GENERAL METHOD

A concept comprising the effectiveness, efficiency and satisfaction

with which specified users can achieve specified goals in a particular
environment.

In this section, we will briefly discuss a reference model for an approach called
user centered design presented by Hans Marmolin in [Marmolin 1993], aiming to
apply the results from the research on human−computer interaction to the design
of software.

Roles of users and computers

Since the 1960s there has been a gradual paradigm shift in the sciences of both
psychology and computing. The majority of researchers now view computers as
vehicles for enhancing the human intellect rather than ultimately replacing it.
(Some AI researchers still advocate, like Douglas Hofstadter [Hofstadter 1980],
that it may, at least theoretically, be possible to describe the human mind by
algorithms. However, the theoretical physicist Roger Penrose recently presented
a very intriguing and convincing reasoning to the contrary [Penrose 1989].)
Since users normally know best how to solve their own professional problems,
a system should not try to lay out a rigid work plan in advance. Rather it should
provide basic facilities to be combined freely by users, thus taking advantage of
the unique capabilities of humans to analyze complex situations, recognize
patterns, follow many threads simultaneously, and adapt quickly to new
requirements from the environment.

A reference model for user centered design

The model for human−machine interaction chosen in [Marmolin 1993] views
user behavior as a goal-oriented process, and human−machine interaction as a
stepwise transformation from goal to primitive actions through a number of
levels describing different aspects of the interaction. These levels are of course
in reality tightly interlaced, and must be chosen somewhat arbitrarily (up to 12
different levels have been suggested by other researchers). Marmolin uses only
three, thus keeping the model simple:
• The pragmatic level. A description in goal-oriented terms of the tasks a
user should be able to solve assisted by the system (the user requirements).
• The conceptual level. A description in system-related terms of functions,
objects, and processes that can be used to achieve the goals.
• The interface level. A description of how to deal with system functions
and objects and of the words and symbols that must be used to activate
 USER CENTERED DESIGN 131

them. This includes the physical operations needed, and the physical
presentation of information by the system.
The user achieves a goal by successively laying out a strategy for approaching it
stepwise through a series of actions which can be solved by the system. The
more knowledge the user has about the system properties on all its conceptual
levels, the easier it becomes to translate a goal into an effective set of physical
operations applied to the user interface.
For example, the user goal to clean up a text file directory (pragmatic level)
must be translated into a series of subgoals like creating safety backups,
removing old files, reorganizing and renaming (still pragmatic level). Each
subgoal must then be translated into functions applicable to system objects, like
removing a file by submitting its corresponding representation to a garbage bin
(conceptual level). Finally, these conceptual actions must be translated into
syntactically correct system actions, like depressing a mouse key on the file icon
and dragging it to the garbage bin icon (interface level).
During the course of the action, both goals and strategy are often modified
based on information about what has happened so far. For example, looking at
the set of file icons in the directory as it gradually becomes more readable may
reveal that a changed file structure (perhaps a merge with some other directory)
would in fact be preferable. Relevant system feedback to stop unnecessary work
or unwanted effects is therefore extremely important for usability.7
In terms of this model, the purpose of user centered design can be defined as
finding ways to decrease the cognitive distance between the successive levels.
By designing the conceptual level to map naturally to the goals represented by
the user requirements, finding relevant system functions for a task is facilitated;
by reflecting the system concepts in the user interface, choosing the right
keystrokes or mouse clicks to invoke a chosen function becomes obvious.
The model represents, of course, a highly simplified view of reality,
disregarding the strong dependencies between the levels. Nevertheless, it can
give new insights into the problem of designing usable systems. We will look
some more at the aspects of interaction represented by each level.

7
We certainly do not speak for the kind of feedback that used to be standard practice in many
early traditional environments (“Welcome to the File Copier XYZ, Copyright QED: 0001 files
copied”). Rather, we favor the UNIX principle to shut up when things proceed according to plan,
and only report anomalies or obvious dangers of unintended results. Routine verbosity always
hides relevant information.
132 ISSUES OF A GENERAL METHOD

The pragmatic aspect

The tasks at the pragmatic level are directly related to the user’s representation of
the problem domain, which usually contains much more than what is in the
system. So the basic concepts at this level are domain specific and system
independent, but the system concepts chosen must facilitate the transformation of
goals and tasks at the pragmatic level into a series of system actions.
The most characteristic property of such a transformation is that it is not a
sequential process guided by a well-defined plan. The order and application of
tasks vary; several tasks are often in process simultaneously, even if they cannot
be done in parallel. The human limitations in memory capacity and precision
usually prevent any effective long-term planning. The strategies tend to be
fuzzy, variable, and flexible depending on continuous reassessments of the
situation. Humans seem to solve problems by going back and forth between
various tasks in a seemingly “random walk” towards a goal [Cypher 1986].

The conceptual aspect

The domain-level tasks must be expressed by the user in terms of functions and
objects supplied by the system [Moran 1983]. Computer systems are regarded
by users as complex models of reality, which they are constantly trying to
understand by relating new experiences with the system to previous ones. The
user gradually builds a model of the system, and only the functions in accord
with this model are ever used. The user’s perception of a system is incomplete
and partly wrong, based on analogies and seldom corrected, since it is often
possible to find reasonable (but erroneous) explanations for the system behavior
in various situations [Carroll 1982].
Efficient interaction requires that the model offered by the system (on which
its design is based) closely matches the user’s mental model of the system. The
fact that users often perceive word processors in close analogy with their mental
model of a typewriter has been studied, showing how the matching parts were
easily mastered while the differences gave rise to many errors and slow typing
[Douglas 1983].
An important aspect of this level has to do with the user’s choice of system
functions to fulfill a certain task. A user does not always choose the best way of
solving a problem, but rather one which minimizes the subjective cost of usage.
This cost is based on the effectiveness of the functions, their perceived
complexity, how well they are understood, and the effort required to learn how to
use them. These factors are valued differently by different users.
So-called naive users tend to favor simple functions, even if they are not very
effective and may require unnecessary repetitions. They value minimal study
and often overrate the difficulty of learning new and more efficient functions.
 USER CENTERED DESIGN 133

They prefer to use well-known procedures, while experienced users tend to learn
more functions, valuing usability more than extreme simplicity. Usually, only a
small fraction of the functions offered by an environment is used routinely. It
was found that among the 400 commands generally available on UNIX systems
in the beginning of the 1980s, the 20 most frequently used covered 70% of all
usage [Kraut 1983].

The interface aspect

Having decided what system functions to use for a given domain-level task, the
next step a user has to perform is to invoke the selected functions by
transforming them into a series of physical operations (keying commands,
selecting from menus, dragging and dropping, pressing buttons, etc.). This may
be considered the syntactic part of the total translation from subgoal to action.
The user’s perception of the system is often strongly guided by the interface,
since the continuous process of using it tends to play a far more important role
than formal education (“if everything else fails—read the manual”).
This routine of “learning by doing” often leads to naive models and rigid
system usage: what is known to have worked before is repeated even when the
user suspects it is not the right way. It is important to recognize this practice,
and that users will learn by making mistakes. Therefore, care must be taken
regarding system responses to strange or unexpected user behavior.
Another thing to note is that learning is often unevenly distributed over the
system functions, so the average degree of user knowledge may not be all that
relevant. Occasional users may know some functions well, while experienced
professional users often are occasional users of some functions.
Finally, three different cognitive states of a user may be distinguished when
working with the interface model of a system: a verbal, a spatial, and a
controlling state. The verbal state corresponds to command dialogs, the spatial
to direct manipulation dialogs (drawing, pointing, dragging), and the controlling
state to multiple choice dialogs (menus, buttons, function keys).
For example, when editing a document in a multi-media environment the user
switches between various forms of interaction. Sometimes the text content is
changed, sometimes the document layout is manipulated, and sometimes
predefined document processing functions are invoked. Choosing an interface
that forces a cognitive state that does not match the current interaction form
closely (such as going through a menu hierarchy to move a figure) often leads to
time delays and increased error frequency because of the extra mental
transformations required. Particularly when a user is trying to solve a difficult
problem that requires deep concentration, cognitive state changes forced by the
computing tools can be very taxing.
134 ISSUES OF A GENERAL METHOD

User metaphors
A significant characteristic of a good system is that it presents a clear abstract
model of what is going on during execution. Instead of just offering chunks of
vaguely related functions and letting the user group them to create enough
structure to figure out what to do, the system should actively support the mental
picture of a virtual world populated with certain actors that behave according to
clearly understandable rules. This applies regardless of whether the user is
human or another program (since somewhere behind a client call there lurks a
programmer with the same basic needs to understand how to use the system
properly). We call this mental picture a user metaphor.
A metaphor can often be chosen to mimic analogous physical environments
that are part of the cultural heritage of most users and therefore natural to use and
easy to remember [Lakoff 1980]. For example, if a system for handling
documents supports the metaphor of a set of ordered racks with drawers, where
each drawer contains a number of folders with individual documents in it, it does
not take many seconds for a user to transpose the familiar layout of an office into
that of the screen.
Some of these metaphors may be quite elaborate and thus very expressive, but
since they are so well known, people do not perceive them as intricate. We
therefore get a full-blown communication language between the user and a
complex system for free, and we get immediate acceptance also from users who
would not touch anything that had the slightest air of complexity about it.
Graphical user interfaces have done much to popularize this idea, since in this
case the metaphor is so obvious that it can hardly escape anybody. When the
graphical interface paradigm—mostly developed at Xerox PARC in the 1970s
[Kay 1977]—was first presented to the people on a broad scale through the
introduction of the Macintosh computer in 1984, so-called naive computer users
could do the most amazing things right from the start, without even needing a
manual. This was a true revolution made possible by hardware advances, but
also by the necessary mental shift of software developers.
The important thing to remember about user metaphors, though, is that they
must be part of the user’s general knowledge. Therefore, it is not enough to talk
to other developers (no matter how experienced) to find out what mappings to
use. Interviews with several potential users, preferably in their normal
workplaces, can reveal what mental pictures most of them share regarding their
work, and what metaphors will therefore be likely to succeed in practice
[Lovgren 1994].
The fact that direct manipulation of graphical icons through dragging and
dropping using mouse buttons is often very appealing to inexperienced computer
users does not automatically mean that it is the best way to communicate for a
 USER CENTERED DESIGN 135

more sophisticated user. However, mouse clicks just represent command syntax;
if we prefer a good textual language, this may be offered as an alternative
without affecting the underlying metaphor.

Modeling of interaction levels

Marmolin has gone a step further and devised a user centered design method
based on the reference model described above. In his method, the three
interaction levels are mapped into three corresponding representations of the
system: a requirements model, a design model, and an interface model
[Marmolin 1993].
Such a general partitioning of the analysis model is not done in BON, because
its usefulness depends very much on the kind of application we are developing.
Maintaining a requirements model separate from the analysis model tends to
break the seamlessness of the approach. It may still be worthwhile for large
systems that are planned to be implemented in several different versions, or have
special demands for requirements tracing, but for most systems one model will
be enough.
As we will advocate in a following section, it is not clear where an interface
model stops and a corresponding conceptual model starts, and the importance of
the interface part will vary greatly from one system to another. However, the
interaction levels and metaphor concept described in this section should be part
of the developer’s mindset and actively guide the analysis and design process.

6.7 ROLES OF ANALYSIS OBJECTS

If our goal is flexible systems that are likely to evolve gracefully over time to
meet new requirements, we should not put all our emphasis on de facto objects in
the problem domain (physical things, organizations, items handled, common
domain terms, etc.).
Finding such objects is often a great help when trying to understand the
problem, and some of them will serve as initial candidates to become classes.
However, many of them are only manifestations of more general underlying
patterns, which are what we should try and capture instead. Also, moving to a
computerized solution usually opens up entirely new possibilities, which may
call for a modified (or radically changed) set of basic concepts.
Anyway, no matter what model we come up with, the degree of potential reuse
of the abstractions is always variable. Some classes will represent stable
abstractions, such as general information structures, basic interaction facilities,
and concepts fundamental to the enterprise or application domain. Others will
reflect specialized, volatile behavior, acting more like temporary glue between
136 ISSUES OF A GENERAL METHOD

comparatively stable components. The classes in a system will therefore play

different roles, which has lead some practitioners to look for a fixed role
separation that can form a basis for a general analysis method.
For example, OOSE [Jacobson 1992] recommends using three different types
of analysis classes (called entity objects, interface objects, and control objects)
depending on which one of three orthogonal system aspects (information,
presentation, or behavior) is considered most prominent for the corresponding
abstraction.
Entity objects are meant to represent more stable abstractions containing
longer-term information needed by the system, control objects encapsulate the
more volatile system behavior, while interface objects take care of external
communication. Grouping the analysis classes into these three categories leads,
in their view, to systems that are more adaptable to requirement changes.
Although a role separation is certainly useful in any system’s analysis model
(BON uses clustering for this) and will most probably lead to more flexible
designs if done correctly, imposing one standard initial classification on the
analysis abstractions in every system is too restrictive. Any predefined set of
classification criteria will be arbitrary, and therefore counterproductive in cases
where these criteria are less relevant.
As we will argue below, the external interface of a system usually consists of
several layers, and what exact parts of it are to be considered the stable “essence”
of the interface is by no means obvious. Therefore, assigning the interface
responsibility to a specific set of classes at an early stage usually means taking
premature design decisions which instead lead to less flexibility.
Moreover, what is important for achieving robust architectures is not to search
for “behavior” versus “information”; it is to look for abstractions which are as
stable as possible. This has to do with the respective generality (in the
application context) of the encapsulated behavior, not with whether it needs to
store information or not.
Object oriented abstraction supports free classification, and it must remain free
if we are to take full advantage of its potential. Instead, the professional
software engineer must be familiar with many different architectures, their
strengths and weaknesses, and be able to choose one that best fits the conditions
of a given project.

6.8 MODELING A SYSTEM INTERFACE

Systems as objects
At the highest level, a system can be viewed as a black box exhibiting certain
well-defined behaviors in response to various incoming stimuli. So if we can
 MODELING A SYSTEM INTERFACE 137

identify the main types of incoming stimuli and specify the actions that should be
taken by the system for each of them, we could in principle always view a
system as one object whose operations correspond to the various responses to
incoming stimuli from the external world.
In practice, however, the complete interface between a system and its
environment is much too large and contains too many different types of
communication to be modeled by just one class. Even if the number of top-level
services available to the users of a system (where a user could also be another
system or hardware device—not just a human operator) may be small, many
auxiliary functions are usually needed. Primary commands will often trigger
series of information exchanges between system and user to pass data through
screen forms or graphical widgets, read from secondary storage files, handle
external interrupts, or follow the conventions of various communication
protocols.
The full system interface is therefore mostly distributed over many classes,
some dealing with input from the keyboard and pointing devices, others
communicating with the underlying file system or some database handler, yet
others exchanging information with external processes. Although control is
always initiated from some top-level objects in the system (ultimately one per
system process—the root objects), the actual detailed communication with the
external world is delegated to more specialized objects.

Front and back ends of systems

There are two main aspects of a system interface: passive and active. The
passive system interface is the part that can be requested to perform some action
on behalf of an external user.8 This is the front end of the system. The active
system interface, on the other hand, is the part that asks for services from the
external world (files, databases, other systems), usually as part of servicing some
incoming request. This is the system’s back end.
These two types of interface play very different roles and cannot be seen as
just encapsulations of the incoming and outgoing events respectively. An
outgoing event could be either an active request (issued by the system) for an
external service, or a passive response to an external request (issued by a system
user). Analogously, an incoming event may be either an external request from a
system user, or a response from an external system to a previous request from the
system. Some external systems may play both roles: sometimes responding to
requests from the system, sometimes issuing requests to it.

8
Unless clearly stated in the context, a system user always refers to any external actor (human or
machine) requesting services from the system.
138 ISSUES OF A GENERAL METHOD

Most of the active external system interface is elaborated during design, since
it has more to do with how a service is to be fulfilled than what is to be done.
However, sometimes well-defined external systems are important prerequisites
that are really part of the system requirements, and then some of the back end
may have to be modeled already during analysis. Here, we will concentrate on
the passive interface, that is the collective services available to system users.

Several levels of interface

So what is it that really constitutes the interface to the system services?
Superficially it can be viewed as the union of all the parts of the various objects
handling external information interchange, since this is where the actual
interaction between the system and its environment takes place. However, this
view is usually too simplistic, because interfacing a system does not just mean
passing some data back and forth, but rather communicating with an underlying
metaphor.
Pinpointing exactly where this communication takes place is generally not
easy and sometimes not even possible when many classes in several layers
collectively model the metaphor. If the American ambassador in Moscow
communicates with the Russian president through an interpreter, does that mean
the interface object is the president or the interpreter? Well, it depends on what
is considered important in the situation.
For example, assume we have a library metaphor offering the mental picture
of a set of books to a certain type of system user. The internal system model
actually handling the books may view them differently—as sortable objects, as
copyrighted material, as stock items of certain size and weight, as assets with
certain values—but the users view them as books.
The next question is how they should be presented to the user—as text lines
on an alphanumeric terminal, as icons on a graphical display, as audio recordings
for visually handicapped users? Given a graphical representation, how should it
be detailed—books standing up or lying down, variable size or just one standard,
possible to manipulate graphically or just simple icons? Given a detailed view,
which look and feel should we pick—Macintosh style manipulation, X/Motif,
Windows?
Some of the detailed views are usually not addressed during analysis, but
sometimes they are part of the requirements and must be taken into account. As
has been argued before, there is no objective level of abstraction for analysis,
since what is special in one context may be quite general in another. What we
should always strive for is a set of abstractions that will be as resilient as possible
with respect to future requirement changes, but the nature of those changes is
completely dependent on what we are developing.
 MODELING A SYSTEM INTERFACE 139

No special analysis interface classes

What the previous section tries to show is that there is (in general) no
fundamental “system interface” that can always be isolated and modeled as
special classes. Rather, the communication between user and system passes
through many layers of abstraction which gradually shift the focus of interest
from sequences of basic mechanical data into the higher-level concepts actually
symbolized by the data.
The codes of individual keys on a keyboard and the button clicks and
corresponding coordinates of pointing devices are translated step by step over
higher-level concepts such as graphical menus, buttons, and icons into objects
collecting information that are beginning to mirror the objects of the chosen
metaphor. Where the actual interface lies (which can supposedly change and
therefore should be replaceable) varies radically from one system to another. In
fact, we can usually expect changes at many levels—the one closest to the
external world need not be the most volatile—so the classes collectively making
up the interface should be treated just like any other set of evolving abstractions.
The same general principles apply for finding common patterns and to separate
different concerns.
The most natural interface to a domain object is captured by the abstract
operations that a user would want to apply to it, and those operations should (in
general) be viewed as operations on the domain object itself, not as operations on
some artificial interface object. Typical domain objects in our last example
might be books and bookshelves. Suppose we want to allow a move operation
on books so they can be transferred by the user from one shelf to another.
Modeling a separate book interface object containing a version of this operation
would in most cases not make sense at all, but only increase complexity and
decrease understandability.
On the contrary, the interface aspect when moving the book is a minor one
that should not be allowed to influence the analysis model at all, but be
postponed to later design. What is important here is the user’s mental picture of
moving an abstract book between abstract shelves that is so easy to comprehend
because of its similarity to the physical reality in most cultures. This metaphor is
what will make the system understandable and accepted; the details involved
when telling the machine to do it can usually wait.
Of course the traditional GUI aspects are by no means unimportant for users,
but they belong to another domain more or less independent of our application.
So when we say that the details can wait, we assume that there will be adequate
packages to choose from during design. Considering the amount of effort
invested in major graphical interface packages, redoing such designs is not
affordable for normal applications anyway.
140 ISSUES OF A GENERAL METHOD

Again, this is not to say that interface objects are never interesting. On the
contrary, if the handling of many graphical views of the same underlying concept
is deemed important or other changes can be anticipated that would be much
easier to meet if the interface is modeled separately, interface objects can be very
useful. Since external interface modeling can be done in many different ways,
the normal procedure is to adopt a model supported by the development
environment chosen for implementation.
However, without special knowledge of the system to be modeled, there is no
a priori reason to favor a separate external interface abstraction before any other
good way of structuring the system. Designing robust software is always a
tradeoff between possible directions of future flexibility. Since a limited amount
of effort is available for each development, rigidly favoring just a few of these
directions will only decrease the chances of finding the abstractions that really
pay off for a particular system.

6.9 SYSTEM USAGE

There are two aspects of any systems development: short-term requirements and
long-term investment. Both are important for systems development.
The short-term requirements are often the direct reason to start a project, and
fulfilling them is necessary to finance the development. These requirements are
mostly expressed in terms of how this particular system is going to be used. The
long-term investment has to do with how the system should be structured both
internally and externally to be correct and reliable, easy to use, and easy to adapt
to future requirement changes. Part of this investment has to do with finding
components that can be reused not only in modified versions of this system, but
also in other systems.
A good analysis and design method should concentrate on the investment
aspect, since this will yield the greatest payoff in the long run. However, a
prerequisite is of course that the short-term conditions be met. Therefore system
usage is always an important factor in all developments.
A scenario in BON describes a characteristic type of system usage whose
details may vary, like registering an attendee in a conference system, or
preprogramming a recording on a VCR unit. Systematically going through
scenarios is a good way of finding new classes and of checking the completeness
of the evolving static structure of abstractions. However, always using scenarios
as the top-level abstractions from which the system architecture is successively
derived, as advocated in OOSE [Jacobson 1992] (where scenarios are called use
cases), is going too far.
The goal is to find stable abstractions that lead to robust systems, but these
abstractions may be found using different approaches. In some applications,
 SYSTEM USAGE 141

certain usage patterns may be very stable (sometimes even regulated by formal
business rules or standards), and then it may be quite helpful to model these
patterns as separate classes in the system. A case in point is the telephone
business, where a phone call has survived many different technologies and been
executed by the user in more or less the same fashion since the beginning of the
century. Under such circumstances the use cases “come for free”, and we may
benefit directly from the information they contain.
But when the requirements (as is usually the case) are vague and incomplete,
it may simply not be cost-effective to spend a large part of the project budget on
the systematic elaboration of many volatile use cases. Since use cases are only a
means to an end, understanding what underlying problem the user is trying to
solve is often more important. And the best way to reach this understanding is
not always by concentrating early on individual use cases (describing how the
user is going to work), but rather spending the initial effort on important
concepts that will affect this work.
Once a clear understanding of the problem is reached and the right concepts
established, a radically different set of use cases may instead fall out as a result
of this understanding, perhaps fairly late in the process. Always basing the
initial modeling on use cases suffers from many of the same disadvantages as the
more traditional approaches of functional decomposition and data flow analysis;
unless the use cases are very stable, much of the initial effort may later prove to
be wasted.
One may of course argue that once a good system model has been found (by
whatever means), the use cases should have become much more stable, and
modeling them as abstractions in their own right would then simplify
requirements tracing and make minor changes to individual use cases
comparatively easy to do in future releases of the system.
However, this is not necessarily so. Since the set of use cases only represents
one possible way of communicating with an underlying abstract model, allowing
this particular usage to affect the system structure in any deep sense may instead
make it more difficult to maintain a clear model and detect in what directions the
usage could change to become more effective. Everything has a cost attached,
and the important thing is to develop a nose for where the efforts should be
concentrated to yield the best return in a given situation. Sometimes use cases
are a major help when structuring a system, sometimes they are less important.
Of course, use cases may never be ignored altogether. Before modeling starts
the analyst has always formed at least a rough idea about the collective set of use
cases, detailed enough to allow an initial small set of classes to be selected.
Depending on available user metaphors and other problem domain knowledge,
the modeling of these classes may continue up to a point where further
elaboration becomes difficult. It is then often useful to group use cases and go
142 ISSUES OF A GENERAL METHOD

through them in detail to obtain more information about how to proceed.

However, this is not just a question of assigning operations to already
identified classes. On the contrary, the needed system behavior successively
revealed by the use cases should instead make the analyst go back to the drawing
board and check for new classes (abstractions) that will help create a more
understandable and reusable system. Recall that there is no a priori problem
domain reality; how you choose to view it depends entirely on what you want to
do with it. Therefore, the initial class structure will be far from fixed.
This switching between investigation of use cases and refinement of the
evolving class structure may be iterated several times. For each iteration, only
the operations which are very basic to the nature of the abstractions found should
be assigned; the bulk of the operations come later when the class structure has
reached some stability. Often many use cases are studied in parallel to detect
common patterns of needed behavior as early as possible, but the focus of
interest should always be on the underlying static model.
Class operations may sometimes fall out naturally from system usage, but
more often use cases will instead hint at some interesting underlying pattern,
perhaps leading to a modified system architecture. Whether this pattern is
detected or not depends on the analyst’s abstraction skills—not on any simple
translation scheme.

6.10 A PARADIGM SHIFT

We conclude this chapter on method issues by drawing attention to a
fundamental principle applicable to all software development, which lies at the
heart of the BON approach. The next two chapters will then develop the BON
method and the standard activities involved in performing its recommended
tasks.
Object-oriented modeling and implementation with software contracts is very
much a question of changed attitudes. Instead of making an attempt at a solution
and then testing the result to see whether you succeeded, you should learn to
trust your thoughts and structure your results enough to know that you are right
before testing begins. You should be willing to bet your last dollar that your
solution is (at least in principle) correct.
This is of course not fully possible or feasible in all cases, but the paradigm
shift behind it is absolutely crucial. Once you discover that you can be in control
instead of just being controlled by circumstances, your life changes radically and
there is no return. The new peace of mind and intellectual liberation is just too
great to ever let you fall back into the sloppy cut-and-try behavior practiced by
programmers who do not know there is an alternative. The importance of this
insight is very well expressed by Harlan Mills in [Mills 1975].
7 The BON process

With the discussion in the previous chapter as background, we now turn to the
rules and guidelines recommended in BON. This chapter will present a general
method for analysis and design, viewed as an idealized process with nine tasks.
The next chapter will concentrate on the standard activities that arise as part of
completing the tasks. Before embarking on a description of each process task in
BON, we will take a look at the standardized BON deliverables.

7.1 THE BON DELIVERABLES

The goal of the BON process is to gradually build the products or deliverables
shown in figure 7.1. The relative weights of these deliverables are far from
equal. Clearly, the leading roles are played by the static architecture diagrams
and the class interfaces. These represent the most precise specification of the
structure and behavior of the system that will exist prior to implementation, and
the idea is that in a suitable object-oriented language environment they should
seamlessly be evolved into a set of executable classes. The thrust of the BON
method is aimed at producing this static, declarative model.
Some of the object scenario diagrams may also be important, since they can
give the developer a good idea of how to implement major public operations in
the classes related to the problem domain. However, to a large extent the role of
the dynamic scenarios is to help discover classes and suitable structures in the
static model, and also serve as a check that what has been modeled is indeed
capable of realizing important aspects of the system behavior.
The static charts are used during early analysis, and for communication with
non-technical people and domain experts. Their relative importance varies
greatly depending on the problem area, how well the task is understood, and the
backgrounds of the people involved with problem analysis and others who need
to understand early system descriptions. Depending on the circumstances, these
charts may or may not be kept as part of the final system documentation. If kept,
they should be automatically updated to reflect the evolving system.

143
144 THE BON PROCESS

System chart SYSTEM CHART

Definition of system and list of associated clusters. Only one SSSSSSSSSSSSSS SSSSSSSSSSSSSS

system chart per project; subsystems are described through

SSSSSSSSSSSSSS SSSSSSSSSSSSSS

SSSSSSSSSSSSSS SSSSSSSSSSSSSS

corresponding cluster charts.

Cluster charts CLUSTER CHART
Definition of clusters and lists of associated classes and SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

subclusters, if any. A cluster may represent a full subsystem

SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

or just a group of classes.

Class charts CLASS CHART
Definition of analysis classes in terms of commands, queries, SSSSSSSSSSSS SSSSSSSSSSSS

and constraints, understandable by domain experts and non-

SSSSSSSSSSSS SSSSSSSSSSSS

SSSSSSSSSSSS SSSSSSSSSSSS

technical people.
Class dictionary Class dictionary
Alphabetically sorted list of all classes in the system, showing SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

the cluster of each class and a short description. Should be

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

generated automatically from the class charts/interfaces.

◊ Static architecture Static architecture
Set of diagrams representing possibly nested clusters, class
headers, and their relationships. Bird’s eye view of the
system (zoomable).

◊ Class interfaces Class interface

Typed definitions of classes with feature signatures and
formal contracts. Detailed view of the system.

Creation charts CREATION CHART

List of classes in charge of creating instances of other classes. SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

Usually only one per system, but may be repeated for

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

subsystems if desirable.
Event charts EVENT CHART
Set of incoming external events (stimuli) triggering SSSSSSSSSSSSS SSSSSSSSSSSSS

interesting system behavior and set of outgoing external

SSSSSSSSSSSSS SSSSSSSSSSSSS

SSSSSSSSSSSSS SSSSSSSSSSSSS

events forming interesting system responses. May be

repeated for subsystems.
◊ Scenario charts SCENARIO CHART
List of object scenarios used to illustrate interesting and SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

representative system behavior. Subsystems may contain

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

local scenario charts.

◊ Object scenarios Object scenario
Dynamic diagrams showing relevant object communication 1

for some or all of the scenarios in the scenario chart. 2

Figure 7.1 BON deliverables: ◊ indicates the most important ones

 THE BON DELIVERABLES 145

However, during system design the typed descriptions will dominate. In many
cases where the problem is well known and the technical level of the people
involved is high, one may prefer to skip the class charts altogether and go
directly for typed descriptions. This may be done either with full graphical
support through a case tool supporting BON, or by using a mixture of the
graphical and textual BON notations (with any personal simplifications or
additions that fit the situation). Regardless of case tools, the latter is common
during whiteboard sessions among engineers.
The dynamic model may also be of interest during early design. In large
systems, there may be many internal events that will lead to interesting non-
trivial behavior, so the BON user is free to extend the event and scenario charts
to also reflect corresponding internal scenarios that may benefit from a dynamic
description. In such cases, local event charts, scenario charts, and object
scenarios may be needed for some of the major subsystems.

Reversibility
The BON deliverables are not independent of each other. On the contrary, there
are close mappings between several of them, which is precisely the idea of
seamless, reversible software engineering. Depending on the situation (type of
system, its size, the people involved), the analysis information gathered by the
developers and the design decisions taken may enter the system model through
different deliverables. The modeling information captured may then be
propagated to other deliverables either manually or automatically, or through a
combination of both. An overview of all BON deliverables and their
interdependencies is given in figure 7.2.
Two complementary types of mapping are shown in the figure: those that
require non-trivial human decisions (single arrows), and those that can be
automated, at least in part (double arrows). The idea is that a developer should
be free to work with the type of deliverable that best fits the situation, and then
use mapping rules and automated support to generate or update other related
deliverables. For example, one project team may choose to initially produce a
large set of class charts, and later proceed to generate templates for the
corresponding typed class interfaces, while another project may start directly
with the typed versions, but still be able to generate the class charts for
documentation purposes.
What is perhaps even more important is that reversibility allows the
mechanism used for creating a given BON deliverable to change during the
course of a software project. A typical way of working may be to start with the
high-level untyped charts combined with some partial static diagrams and
dynamic scenarios. Later, when the requirements have become more stable and
146 THE BON PROCESS

LEGEND:
intellectual help to create/update
Class dictionary
possible (partial ) automatic SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS

generation SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

SYSTEM CHART CLUSTER CHART CLASS CHART Class interface

SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSS SSSSSSSSSSSS Class text
SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSS SSSSSSSSSSSS
(source code)
SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSS SSSSSSSSSSSS

Static architecture

STATIC MODEL
DYNAMIC MODEL
EVENT CHART SCENARIO CHART Object scenario CREATION CHART
SSSSSSSSSSSSS SSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS
1 SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

SSSSSSSSSSSSS SSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

2
SSSSSSSSSSSSS SSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

Figure 7.2 Dependencies between BON deliverables

the problem is better understood, typing will be needed to advance the system
model. The intellectual effort is then usually shifted to the formal class
interfaces, and subsequent updates carried out directly on these.
At that point in time (after the initial class charts have been converted into
class interface templates) there is a choice. We may choose to discard the
untyped static charts as we move into a more precise typed system description,
or we may choose to keep them as documentation. In the latter case, automatic
support is needed to maintain consistency between the charts and the evolving
interfaces.
Several BON deliverables thus have dual roles in systems development. They
may initially act as intellectual guides for finding a good architecture with
corresponding typed class interfaces, and later as high-level documentation of
the system.

Roles of deliverables
To make it clear how the deliverables are meant to be used in BON, we now look
in more detail at the dependency arrows in figure 7.2, and begin with the static
model. The chain of single line arrows extending from the system chart, over
 THE BON DELIVERABLES 147

cluster charts and class charts, into formal class interfaces (and eventually
implementation) indicates that manual creation of these deliverables and the
corresponding input of modeling information should proceed from left to right.
The same set of deliverables is also connected through a chain of double line
arrows, indicating that automatic generation is possible in both directions.
The idea is that information entered in one deliverable may be used as the
intellectual basis for creating or modifying the next deliverable to the right in the
chain. Templates can be automatically generated, but since deliverables to the
right represent refinements, new information needs to be added manually.
However, when changes are to be reversed (propagated right to left) in the chain,
a much higher degree of automatic support is possible and should be used. This
is why the single arrows only go in one direction.
As a project progresses, the deliverables directly modified by the development
team tend to be further to the right in the chain (typically formal interfaces
during design and source code during implementation). Automatic reverse
generation then becomes more essential for maintaining consistency, particularly
because of the increased volume and level of detail. The class dictionary, for
example, which contains an alphabetically sorted class index, should always be
created automatically from the class charts and/or the formal interfaces.
The static charts are also connected with the static architecture through single
arrows in both directions. Typically, initial charts give rise to partial static
diagrams, which may then be independently elaborated giving rise to new or
modified charts. Since feature names correspond to labels in static diagrams,
and inheritance as well as different types of client relations are part of both the
static architecture and the class interfaces, automatic propagation and
consistency checking is possible between them.
The double arrow from the class text to the static model indicates that it is
possible to generate unlabeled client links between classes to show a relation that
occurs for implementation reasons, but is not part of the public interface of any
class. Although BON is not concerned with how implementation will be carried
out, it may sometimes be essential to show that a certain set of important
implementation classes—for example, encapsulating an external database
interface that may be part of the system requirements—will be used behind the
scenes without specifying any details.
Finally, we look at the dynamic model. Here there are only single arrows,
since no automatic generation is possible. The event charts serve as
complementary help to select and describe scenarios of typical system usage,
which are collected in scenario charts. The scenario charts and the creation
charts (whose content is guided by the static architecture) then jointly form the
basis for the object diagrams. The construction of these diagrams may, in turn,
lead to changes in both event charts and creation charts.
148 THE BON PROCESS

Although the static and dynamic models are two very different types of system
description, they are closely related since the communicating objects in the
object diagrams correspond exactly to the classes in the static architecture
describing the behavior of each object. It is therefore possible to alternately
work on the static architecture and the object scenarios—whichever seems most
likely to yield new insights depending on the current degree of problem
understanding—and then propagate any additions and changes to the other
model.

7.2 CHARACTERISTICS OF THE PROCESS

Thus, having discussed the BON deliverables, we may turn to look at the system
development process whose aim it is to create them. A summary of the process
tasks is shown in figure 7.3. The tasks are listed in an approximate order of
execution, and represent the ideal process referred to in [Parnas 1986].
However, the BON user is free to change the order of the tasks or make any
other deviation from the ideal process that will help fulfill the goals of a
particular project, as long as the required static and dynamic models are
eventually produced.

Risk-driven development strategy

In fact, there are always aspects of a system development project that must be
allowed to take priority over any predefined work procedures. After all, the
direct aim of such a project is nearly always to deliver a product with adequate
functionality and performance to a customer on time and within budget.
Therefore, all software development procedures should be risk driven. This
means that the level of depth and detail in various parts of a problem
investigation should always be adapted to the potential future cost of
misunderstanding or modeling something incorrectly. Good high-level early
descriptions, particularly for large and complex systems, are therefore often
highly unbalanced in terms of elaboration of their different parts.
Instead of falling into the trap of modeling well-known concepts in great detail
during the early phases (and get a false sense of rapid progress), we should take a
deep breath and dive straight into the tricky problems. Eventually, we will have
to deal with them anyway, and postponement only increases the risk of
unpleasant surprises or even catastrophe later in the project.
Although this book is not about project management, we would like to
emphasize the importance of continually assessing and controlling the high-risk
elements in software development projects; see for example [Charette 1989]. In
the USA, the Department of Defense and NASA have long been strong
 CHARACTERISTICS OF THE PROCESS 149

TASK DESCRIPTION BON DELIVERABLES

Delineate system borderline. Find major
G 1 SYSTEM CHART, SCENARIO CHARTS
A
subsystems, user metaphors, use cases.
T
H List candidate classes. Create glossary of
2 CLUSTER CHARTS
E technical terms.
R
I SYSTEM CHART, CLUSTER CHARTS,
N Select classes and group into clusters.
G 3 STATIC ARCHITECTURE,
Classify; sketch principal collaborations.
CLASS DICTIONARY

D Define classes. Determine commands,

4 CLASS CHARTS
E queries, and constraints.
S
C Sketch system behaviors. Identify events, EVENT CHARTS, SCENARIO CHARTS,
R
I 5 object creation, and relevant scenarios CREATION CHARTS,
B drawn from system usage. OBJECT SCENARIOS
I
N Define public features. Specify typed CLASS INTERFACES,
G 6
signatures and formal contracts. STATIC ARCHITECTURE

CLASS INTERFACES,
Refine system. Find new design classes, STATIC ARCHITECTURE,
7
D add new features. CLASS DICTIONARY, EVENT CHARTS,
E OBJECT SCENARIOS
S
I CLASS INTERFACES,
G
N 8 Generalize. Factor out common behavior. STATIC ARCHITECTURE,
I CLASS DICTIONARY
N
G Complete and review system. Produce
Final static and dynamic models;
9 final static architecture with dynamic
all BON deliverables completed.
system behavior.

Figure 7.3 The BON process: initial version of deliverables underscored

advocates of risk analysis and management as a major instrument to control

software development.
The above does not mean we cannot start by modeling the more familiar parts
of a system. On the contrary, this is often a good way to reduce initial
complexity so we can begin to see the wood rather than just trees. However, the
easy modeling should only be taken far enough to cleanly separate certain
familiar patterns from the more difficult parts.
Once we understand these patterns, we can set them aside and return to them
later when the hard parts are also under control. This strategy increases the
chances of detecting potential trouble spots early in a project, so they can be
nipped in the bud.
150 THE BON PROCESS

Process tasks
As shown in figure 7.3, the tasks are loosely grouped in three phases with the
following aims:
• Gathering analysis information (tasks 1−3).
• Describing the gathered structure (tasks 4−6).
• Designing a computational model (tasks 7−9).
The partial ordering for some of the tasks is clearly fixed. For example, it would
not make sense to complete the final architecture before we have delineated the
system borderline or define features before the corresponding classes have been
selected. But some of the tasks are often reversed in practice, depending on the
nature of the problem at hand. The following tasks are flexible in this respect:
Task 1 (delineate system borderline) is sometimes postponed to after task 3
when the idea of what the system is supposed to do is initially very
vague. In such cases, we may need to design a first preliminary object
model before it can be decided what should fall within the scope of the
system and what should not.
Task 4 (untyped class definition) is often skipped in projects where all parties
involved have high technical skills and long experience and is instead
replaced by a typed, more formal definition. If desired, the class charts
may then be later generated from the formal interfaces.
Task 5 (sketching system behavior) is sometimes done very early, perhaps
already as part of task 1, for systems where a large part of the initial
requirements are expressed as complex global actions.
Task 8 (generalization) may be applied already in the early analysis phase in
case a fine grain abstract view of the problem domain is crucial for
subsequent system design and implementation.
Each task in the BON process has a set of input sources, is controlled by
acceptance criteria, and produces a set of deliverables that become part of the
analysis and design documentation. The deliverables that are created or updated
as a result of each task being carried out are listed in the rightmost column of
figure 7.3 (with the initial version of each deliverable underscored).
As was pointed out in the introductory chapter, BON concentrates on the
general aspects of analysis and design and does not address the full development
life cycle of object-oriented software engineering. Instead, the method is kept
small, while still covering the issues that are central to a wide variety of
applications in both large and small projects. Once these are understood and
 CHARACTERISTICS OF THE PROCESS 151

mastered, it should not be difficult to extend or modify parts of the method to

meet specific needs.
Besides analysis and design, there are also many more important issues
involved in conducting successful object-oriented software development
projects, many of them highly enterprise dependent. Therefore, the BON
process is meant to be integrated with other processes and tailored to suit the
specific rules and policies of the software production environment in which it is
going to be used. For example, since formal quality assurance procedures will
be highly variable from one project to another, the acceptance criteria listed in
the boxes accompanying the description of each process task (see the following
sections) are to be understood only as simple reviewing hints.
For the rest of this chapter, we will discuss each process task in more detail
with the emphasis on what should be produced and in what order. The next
chapter will then examine the typical analysis and design actions involved, and
will be more concerned with how to find the abstractions and structures needed
to produce the deliverables.
Although (as was argued at some length in the previous chapter) drawing a
precise line between analysis and design can only be done arbitrarily, we will use
a simple definition to roughly discriminate between the two:
1. As long as we only deal with classes that can be considered directly related
to the problem domain and restrict ourselves to the public features of these
classes, we are doing analysis. We call these classes analysis classes.
2. When we begin adding classes and features that are not directly related to
the problem domain, design has begun. We call the new classes design
classes.
With this definition, tasks 1−6 focus on analysis and tasks 7−9 deal with design.
The above view on analysis and design seems to work reasonably well in
practice, but the definition is of course cheating. In general, there is no way of
telling for sure whether a certain class can be considered “directly related to the
problem domain” or not. This is a matter of judgment, since there is no objective
reality to compare with—just as in politics, every view of a problem domain is
related to what you want to achieve.
However, even if we are not always sure whether we are doing analysis or
design (or implementation for that matter), so be it. We should still be able to
apply the same powerful principles of abstraction and not worry too much. After
all, it is the usefulness of the resulting models that counts, not what we think we
did to achieve them.
We will now define and discuss each of the nine process tasks in more detail.
Each task description will start by displaying a standardized box summarizing
152 THE BON PROCESS

the general results produced by the task, what specific BON deliverables are
affected, and what input sources and acceptance criteria to use. (The latter are
only hinted at, since quality assurance requirements vary greatly from one
organization to the next.)

7.3 TASK 1: DELINEATE SYSTEM BORDERLINE

GENERAL RESULTS BON DELIVERABLES

• Major subsystems SYSTEM CHART SCENARIO CHART

• User metaphors SSSSSSSSSSSSSS

SSSSSSSSSSSSSS
SSSSSSSSSSSSSS

SSSSSSSSSSSSSS
SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS
SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS

• Incoming and outgoing information flow SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

• Major system functionality

• Typical use cases
• Reused libraries and reuse policy
INPUT SOURCES ACCEPTANCE CRITERIA
• Customer requirements • Customer acceptance procedure
• Domain experts, end-users • Quality assurance plan
• Knowledge of implementation platforms
and run-time environments
• Comparable systems or prototypes
• Available reusable components
• General rules and regulations

This task is concerned with what is going to be modeled and what will be our
principal view of the world that we are trying to understand and formalize. It is
an extremely important task, since it sets the stage for everything else in the
subsequent modeling work, and should not be taken lightly. Particularly since
the initial difficulties may be quite different depending on system size and
modeling purpose, we will spend some time discussing various aspects of the
task.

System borderline
Analysis is the study of concepts and procedures in a problem domain. What
constitutes the problem domain in a certain modeling situation depends entirely
on the purpose of the investigation. As discussed in the previous section, object-
oriented analysis may be used to model with the intent of producing software, or
just to increase the understanding in some area of interest.
If we produce a very general analysis model of some business operations, it
would make perfect sense to talk about “design” and even “implementation” of
various organizations by elaborating this analysis model in different directions,
even if no software or computers were ever to be involved. Since design has to
 TASK 1: DELINEATE SYSTEM BORDERLINE 153

do with finding a form that suits a certain purpose, the borderline between it and
“the problem” is really just a matter of convention.
However, using design and implementation in this broad sense would
probably not serve any useful purpose when describing the BON process
(although it may certainly do so in other contexts). Therefore, when discussing
the process tasks and the corresponding standard activities in BON, analysis will
mean creating an object-oriented model of the problem, and design will mean
realizing this model in a computerized environment.
Implementation, finally, will mean translating design into program text using a
selected programming language. Usually, however, this language is known
already during design and is then likely to influence the design model chosen.
When BON is just used for enterprise modeling, design or implementation will
not be considered.
Delineate the system borderline means to decide what aspects of a problem
area should be modeled at the highest level, and what will be the communication
between elements in this model and the external world. It may include modeling
views of external systems, such as database management systems and GUI
packages, or this may be left to design. It all depends on what is considered
fixed and what may change.

Subsystems
For small systems, possible partitioning into loosely coupled subsystems can
often be postponed to the general class clustering of task 3, but if the system is
large, the organizational question regarding work groups becomes an important
issue already from the start. Since massive communication between many
people is extremely costly and error prone (even worse when geographical
distances are involved), having small groups of people working on separate parts
of a system is always a major principle for large developments.
Sometimes it even takes precedence over other principles, so that, for
example, a certain amount of fine grain reuse may be sacrificed in order to
achieve more independence between groups working in parallel. Software
production is a series of compromises, and being a good engineer means having
a nose for which compromises pay off and which ones do not. This kind of
knowledge does not come from books, but from a combination of talent and the
right sort of experience.
Thus, when analysis and design of a large system is to proceed concurrently,
the major subsystems should be decided already as part of the initial task, so an
appropriate work organization can be found. It is then important to make the
system parts as logically independent as possible to minimize group
interdependencies and facilitate final system integration.
154 THE BON PROCESS

Since there is always a risk of premature decisions if the problem is not very
well understood, it is usually better to produce a high-level model of a large
system first. When this model has been iterated several times, it will be much
easier to select the best partitioning into subsystems and then scale up and
proceed with a number of parallel developments.
In creating the initial model, we must be careful not to advance the design
“breadth first” without having fully investigated the most difficult parts of the
problem. This often means we need to go very deep on some parts of the
system, and often take one or two central ideas all the way down to
implementation (on a small scale) to make sure they will work.
The initial modeling should be carried out by a small group chosen from the
best people available. (What often happens when you try to gain time by
violating this rule may be read in Fred Brook’s classical essay collection The
Mythical Man-Month [Brooks 1975].) When a core design has been produced
and tested we are much safer, and the risk of steering in the wrong direction
when many people come aboard and costs really start to accelerate is greatly
reduced. This is precisely in line with the risk-driven development strategy
discussed before.

User considerations
One of the major factors to consider when outlining the system borderline is the
prospective computer user. Below are some guiding principles taken from
[Marmolin 1993].
• Identify usage. Describe the situation in which the system will be used: goals
of the enterprise, roles, tasks, resources, flows of information, work processes,
social relations. Try to understand the interplay between the computer-
supported activities and other tasks to be carried out.
• Identify user groups. Describe the different types of users involved in terms
of: knowledge, competence, experience, attitude, frequency of usage. Are the
groups homogeneous or variable? It may be necessary to design alternative
dialog forms depending on user type, variable help levels, allow personalized
configuration, plan for selective user training, etc. Is the system a central part
of the user’s tasks or just peripheral? Physically straining (long hours of
terminal input), or just a complement to more demanding intellectual work?
• Worst case principle. Design the system interface particularly for the users
who are expected to be in most need of assistance. The problem is of course
to strike a balance between the extreme simplicity required by some groups
and the rich functionality needed by others.
 TASK 1: DELINEATE SYSTEM BORDERLINE 155

• Metaphor principle. Base the design model on one or more metaphors from
the users’ professional environment or cultural background [Booth 1989].
These should be documented early, and used as input for the searching and
selection of candidate classes in the problem domain. (Be sure to also
document all important differences between the system model and what the
metaphors would indicate.) Some would regard selection of user metaphors as
design, but in our view it is clearly analysis (in the new sense of the word),
because the concepts involved will nearly always be close to either the
problem domain or some other external reality which is familiar to the user.
• Minimality principle. Include exactly the functionality needed to solve user
tasks—no more, no less. This recognizes the importance of striving for a
small number of orthogonal functions, which will be perceived as relevant by
the users and therefore often used. Functions seldom or never used yield a
more complex system model making the simple functions also more difficult
to understand.
The principle is of course a guideline rather than a strict rule. In practice, a
set of specialized functions is often needed for certain situations or by
advanced users, but these functions should then be kept in separate layers in
both the user interface and the documentation, so they do not interfere with
normal usage.
• Identify potential semantic errors. Try to think of possible semantic
misunderstandings that may cause the user to misinterpret the effect of a
function. Making the design model explicit and using metaphors decreases
the risk of semantic errors, but the problem is generally difficult. Mismatches
between the design model and the user’s mental model is one cause for
misunderstandings; semantic inconsistencies in the system is another.
Common semantic errors may be countered in three ways: better education,
better documentation (including on-line help), or modified metaphors.
Continuous feedback regarding current processes and system state is important
to help users detect errors. A problem is the inherent conflict between
providing enough information, on the one hand, and not disturbing the user’s
work by supplying useless information or forcing unnecessary confirmation of
intentional operations, on the other.
A feasible compromise in many situations is to concentrate on decreasing
the risk of irreparable damage by providing stepwise “undo/redo” facilities,
and possibilities to interrupt and resume ongoing processes.
• Use transparent modeless dialogs. Transparency means that the dialog has a
natural and direct relation to the interface. Ideally, the user should not be
aware of any interface, but only of objects that can be directly manipulated in
156 THE BON PROCESS

obvious ways [Shneiderman 1983]. Modeless means that the meaning of a

user function is the same regardless of system state. If modes are necessary to
structure a large set of different system behaviors, they should be naturally
related to the user task. The principle is to permit the user to do anything at
any time, since humans do not solve problems sequentially according to well-
defined plans.
The goal is not really to remove modes, but to remove what appears to be
modes to the user. For example, if the user must explicitly toggle between
modes (like drawing mode, text mode, and command mode) to make
subsequent input be interpreted as intended, this may lead to confusion and
difficulties in learning the system.
If, instead, there are three types of window, whose purposes are obvious by
their graphical appearance (like canvas, editor, and command interpreter), all
the user needs to do is place the pointing device in the desired environment
(which is then highlighted as confirmation) and start performing actions
natural in this context. The modes, which are still there logically for the
system to interpret, have disappeared from the user’s perception.
An analogous example is the ordinary telephone whose interface is so
familiar to people in the industrial world that it has become invisible (that is,
until you try to remember the baroque sequences of stars and number signs
needed to redirect your phone, or order a wake-up call).

External communication
When we have chosen what parts of the problem domain to model and what
general viewpoint to use, we should look at the communication between the
elements in the chosen model and the external world. This includes listing major
data flows coming into and leaving the system. Frequency rates and sizes are
important for systems that will handle heavy loads of on-line transactions or
capture large amounts of external sensory data in real time.
Data typically enters the system through terminal input (keyboard or mouse),
pipes, files, distributed objects, sensory devices, or notifications from other
systems, and leaves it through terminal output (screen, printers, or multimedia),
remote calls, pipes, files, and drivers for electronic devices. Data flows will be
part of the incoming and outgoing events that represent the abstract stimuli and
corresponding responses from the system viewed as a black box. These are
documented in an initial version of the system event chart.
Only the basic information content is considered at this stage, since the exact
form of data communication should not be decided too early. Major aspects of
the user interface, such as which user metaphors to choose, should be dealt with
early since this may have a strong influence on the overall design of the system,
 TASK 1: DELINEATE SYSTEM BORDERLINE 157

but the data flow details as well as what specific widgets and dialog sequences to
use are best postponed.

Major functionality
Starting from the user requirements and guided by the identified external
communication and user metaphors, we produce structured lists of desired
system functionality. Even if we have not yet decided on classes, most of the
functionality can often be grouped into sections relating to different concepts of
the system—user metaphors are particularly helpful here, since their elements
are obvious candidates for future classes.
A certain number of clusters are defined and documented in the system chart.
These represent subsystems or other groups of classes to be defined later, which
are separated according to various criteria. For large systems this work may be
very substantial and constitute the first real structuring of our model, while small
systems may initially contain only one cluster. We will discuss clustering in
more detail later under task 3.

Typical scenarios
From the listed data flows and initial event chart we also identify a number of
typical examples of system usage, or object scenarios. These are documented in
a first version of the system scenario chart, and will be used later to build the
corresponding dynamic scenario diagrams as complement to the static diagrams.
Depending on the nature of the problem, we may choose to do the dynamic
modeling already at the very beginning, and in that case we will also produce
first versions of the creation chart and the object scenarios in this task.

Incremental development
The initial requirements are often vague and inconsistent since customers rarely
know exactly what they want. This is not because customers are generally stupid
or unable to make firm decisions, but because the software to be developed is
mostly aimed at supporting rather special activities. Unless there already exists a
completely analogous system to compare with, only clairvoyant people will
know beforehand how useful certain initial ideas will turn out to be when
implemented in practice, and how much you can get at what cost.
If the modeling can be followed to its logical conclusion through use of a good
object-oriented language, there are great opportunities to provide early feedback
to users and customers. Instead of spending six months negotiating the details of
a 2000 page requirements document, and then proceeding for two years to
implement a system where half of the functionality is inadequate because of
158 THE BON PROCESS

early misunderstandings having been cast in concrete (and the other half possibly
obsolete at the time of delivery), a much more adaptive approach is possible.
By implementing only small portions at a time, starting with a very general
framework and proceeding to elaborate and flesh out functionality in the
directions shown by experience to be the most fruitful, enormous amounts of
wasted effort can be avoided; see for example [Boehm 1981, Gilb 1988]. This
approach is in complete harmony with the object-oriented philosophy of finding
well-defined reusable abstractions at all levels.

Look for reusable components

The project team should always search actively for possible libraries to reuse,
since it is there that the future potential lies for increasing productivity by an
order of magnitude. A decision to reuse is of course always a tradeoff between
how well the available functionality fits the current problem, what general
quality it has, how its future maintenance is taken care of, and so on. This must
be weighed against the corresponding factors and extra cost of developing new
software.
As the availability of reusable software increases, it becomes more and more
important to have a good knowledge of where to search for suitable components.
Being connected to a global network like the Internet [Quarterman 1990,
Krol 1992], both for general inquiries on product information and for easy access
to the very impressive stock of high-quality freeware that is currently being built
by universities and non-profit organizations (for example, MIT, Stanford, and the
Free Software Foundation), will soon become a strategic necessity for many
developers.

Establish reuse policy

We should make a conscious decision as early as possible regarding the level of
investment in the future reusability of the system we are going to design and
eventually build. This is important, since a high degree of reusability demands
much more effort in terms of careful design, generality, robustness, extensive
testing, and thorough documentation than is needed for a piece of software that is
only to be part of one particular context.
Reuse within a project should always be standard practice, since it pays off
directly and serves to foster the needed software culture. Reuse between projects
is of course a matter of the overall life cycle costs of the systems involved—now
and in the future. Heavy reuse of mediocre components is dangerous, even if
money appears to have been saved in the beginning, since it may create
dependencies that will delay production of the good abstractions.
 TASK 1: DELINEATE SYSTEM BORDERLINE 159

Often, a compromise works well here: build as much reusable software as

possible within the budget of the project, but with the emphasis on meeting the
specific requirements of this contract. This often means that many ideas are
sketched with only parts of their potentials realized, but care is taken to pave the
way for future generalization. Then, when the product has been delivered and
accepted, devote a special investment budget for capitalizing from the project by
turning the best ideas produced into reusable libraries.

Scaling
The system we are modeling may be very small or very large. A large system
will probably have to be split into many subsystems, but this does not necessarily
mean that modeling becomes more complex. If we can build a top-level
structure in which the subsystems are viewed only at a very high level, a model
of the whole system may still be understandable enough to be managed by a
small group of people. Reasonably independent subsystems can then be given
complete BON processes of their own, and only the uppermost level of each
subsystem need affect the behavior of the total system.
Depending on the goal of a project, the top-level modeling can be very
different. If the underlying implementation platform is fixed for the foreseeable
future, details of its architecture will probably affect the analysis model. There is
nothing a priori wrong or too specialized about this, since what is special in one
context may be quite general in another. The aim is to be as resilient as possible
to future requirement changes, but the nature of those changes will vary
depending on the type of development.
On the other hand, if you are designing a public service like a mobile
telephone system, what you want to model at the top level should probably be
completely independent of both the underlying platforms and their physical
distribution. Instead, it becomes all important to have a strong user metaphor,
which may survive the technological advances for a long time and will lend itself
to many highly different realizations. The model will then be expressed in terms
of the metaphor, and even if the underlying design and implementation may lead
to huge subprojects, the model can still be kept manageable.
As with all problem solving, the key to mastering complexity is scalable layers
of abstraction, enabling the thinker to concentrate on one thing at a time.

Environmental restrictions
Before proceeding to concentrate on the internal modeling of the delineated
system, make sure that the basic capabilities and limitations of the external
environment in which the system will run are well understood. An elegant
analysis and design model which is based on certain assumptions about available
160 THE BON PROCESS

services from an underlying operating system or database manager may suddenly

prove unimplementable in practice when it is later discovered that the
assumptions were wrong.
This is particularly dangerous when moving into a new type of environment.
For example, developers who have only been exposed to highly flexible
operating systems with few limitations, such as UNIX, may easily consider
certain basic facilities so fundamental that the thought of checking never occurs
to them.

7.4 TASK 2: LIST CANDIDATE CLASSES

GENERAL RESULTS BON DELIVERABLES

• First list of classes based on problem CLUSTER CHART

domain terminology SSSSSSSSSSSSSSS

SSSSSSSSSSSSSSS
SSSSSSSSSSSSSSS

SSSSSSSSSSSSSSS

• Glossary of technical terms and concepts SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

used in problem domain

INPUT SOURCES ACCEPTANCE CRITERIA
• Problem space • Approval by end-users and domain
• User metaphors experts
• Specifications of implementation
platform(s) and external interfaces

This task is the real starting point of the object-oriented development process.
From the problem domain we can often extract a number of de facto concepts
that are candidates to become classes. We will say more about this when the
BON activities are discussed in the next chapter. The various actors in the user
metaphors as well as documented interfaces to external systems are also obvious
class candidates.
What we produce here is a list of candidate classes, which means we should
not be too critical about including a concept from the problem domain—better to
have too many concepts than risk missing some of the important ones. In fact, it
is sometimes possible to choose between several disjoint sets of concepts, each
representing a different way of partitioning the problem space. Seeing the
overlapping candidates listed together can then make it easier to pick the best
consistent set. So the list will nearly always contain some candidates that will
not become classes, and others that will perhaps become several classes.

Classes vs. objects

You sometimes hear people argue that early analysis should concentrate on the
objects and not worry too much about the classes. Primarily, the reasoning goes,
 TASK 2: LIST CANDIDATE CLASSES 161

we detect a number of actors in the problem domain (the objects) and only later
does it become clear how these objects may be grouped into classes. However,
this is a misunderstanding of what objects and classes stand for.
In object-oriented analysis we are not concerned with objects per se. On the
contrary, our aim is to describe them only indirectly—through their behavior.
Everything else is irrelevant and should be excluded from the analysis model. In
an object-oriented context, we have one and only one way of describing object
behavior, and that is through classes (disregarding the hybrid approaches, against
which this book is trying to make a case). This means that even if there is only
one object of a certain type, the description of its behavior will still be called a
class. The name is well chosen, because no matter what object is modeled, it is
always possible to envision more objects of the same type—now or in the
future—and the class will then automatically cover the behavior of these objects
as well. (The class actually describes a pattern of behavior, since two individual
objects of the same type may behave differently depending on system state.)
So when we do not know initially whether two objects behave according to the
exact same rules or not, we simply model them as different classes until we
know better. Our initial attempts at classifying the problem domain will have to
be iterated many times anyway, and merging several classes into one is no worse
than splitting up a class found to represent more than one abstraction, or
changing the inheritance structure.
Since object-oriented development downplays individual objects (of which
there are often a great many) and instead concentrates on the different types of
behavior (of which there are much less), the term is really a misnomer.
However, starting to push class-oriented technology as a solution to the
problems of modern society might be misunderstood in some circles.

Glossary of terms
Besides a list of candidate classes, we also produce a glossary of technical terms.
This is particularly important in problem domains that use special terminology
whose meaning is not obvious to the analysts and designers. The glossary
should stress points where misunderstandings may occur, and will be used later
as a guide for naming the features of problem domain classes.
It is vital that analysts and users/domain experts understand each other’s terms
precisely, otherwise wrong interpretations may easily lead to incorrect system
requirements. In most cases, it is probably easier for the computer professional
(being familiar with applying modeling to widely different areas) to adjust the
terminology to that of the problem domain than vice versa. However, sometimes
common problem domain terms contain ambiguities or connotations that may be
necessary for discussing complicated matters between domain experts, but will
162 THE BON PROCESS

only be a hindrance in the abstract model. Then it may be better to agree on

some alternative terms to use in the joint work between analysts and domain
people.

7.5 TASK 3: SELECT CLASSES AND GROUP INTO CLUSTERS

GENERAL RESULTS BON DELIVERABLES

• Problem-domain-related classes SYSTEM CHART CLUSTER CHART

grouped into clusters SSSSSSSSSSSSSS

SSSSSSSSSSSSSS
SSSSSSSSSSSSSS

SSSSSSSSSSSSSS
SSSSSSSSSSSSSSS

SSSSSSSSSSSSSSS
SSSSSSSSSSSSSSS

SSSSSSSSSSSSSSS

• First static architecture with some SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

inheritance and client relations

INPUT SOURCES Static architecture
Class dictionary
• Possible subsystems SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

• List of candidate classes

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

• Glossary of technical terms

ACCEPTANCE CRITERIA
• Class and cluster management system
with automatic consistency checks

In this task we start working with the raw material collected in task 2. Beginning
with the list of candidates, we form an initial set of concepts to model as classes.

High-level class properties

At this point we may also start thinking about certain key properties of the
classes. Some objects may need to be persistent, which means that they must not
be deleted when the system execution terminates, but instead be saved for later
retrieval. Use of an external database for this may have been specified in task 1
(delineating system borderline) or we may decide storage form later, during
design.
Some of the selected classes may be deferred (in some circles called abstract),
which means they will never be fully implemented but instead rely on
descendant classes to fill in the necessary details. Deferred classes cannot be
instantiated and will therefore not be part of the object scenarios. The
corresponding actors at system execution time will be instances of effective
descendant classes.
Some classes may be externally interfaced, which means they encapsulate
some external communication (function calls, data retrieval, etc.).
Some classes may be parameterized (also called generic), which means they
give rise to different types depending on a set of type parameters. This is
common for various types of container classes.
 TASK 3: SELECT CLASSES AND GROUP INTO CLUSTERS 163

Similarity and collaboration

Some of the classes are usually easy to select, while others are more difficult.
When we get stuck, the best way to proceed is usually to start grouping the
classes we have selected according to the two main aspects supported by object-
oriented classification: similarity and collaboration.
Similarity refers to common patterns of behavior, and is captured by
inheritance relations. Collaboration, on the other hand, refers to objects using
the services of each other, and is expressed by client relations.
Do not be misled by those who claim that inheritance is implementation
oriented and should be postponed to later phases. On the contrary, inheritance is
classification of behavior just as much as is the grouping of objects into classes.
If the behavior pattern of two objects is exactly the same, we describe them by
one class; if it differs only slightly, we describe them by two classes related
through inheritance. In both cases we are reducing the amount of behavior
description, which helps us master complexity and get a clearer understanding of
the problem space. Of course inheritance can be used prematurely to reflect
things that are in fact solution oriented and should not occur in the analysis
model, but so can any other modeling facility.

Possible system views

Local modeling of smaller groups of classes to find relations between them will
often help us detect patterns that can guide us in selecting clustering criteria. A
given set of classes may be grouped into clusters according to many different
aspects: subsystem functionality, concepts of a certain application domain,
subparts of the same whole, children of a common ancestor, encapsulation of
external systems or devices, and so on. Figure 7.4 shows a possible system view
(partitioning into clusters) of a set of classes representing different cars and their
tires.
The same set of classes may also be viewed as in figure 7.5 without changing
any relations between the classes or any class interfaces. However, in practice it
is usually too much trouble to maintain several independent system views, so
only one is chosen. The retained view represents a compromise between
possible aspects that could be of interest. The grouping criteria are thus chosen
for each individual cluster, depending on what is deemed most important with
respect to the particular classes involved. Some clusters will represent
subsystems, others general libraries, yet others simple groupings of parts or
descendants of a common ancestor.
Note that system views in BON are views that may be expressed as a
hierarchical partitioning of the classes making up the system. There are also
other views of interest to the developer, which cannot in general be expressed as
164 THE BON PROCESS

BMW FERRARI

CAR

VOLKSWAGEN FIAT

CONTINENTAL TIRE PIRELLI

STOCK ITALIAN_IMPORT
GERMAN_IMPORT

Figure 7.4 Trade-oriented view of cars and tires

SPARE_PARTS

PIRELLI TIRE CONTINENTAL

BMW VOLKSWAGEN

CAR

FERRARI FIAT

DE_LUXE STANDARD

VEHICLES

Figure 7.5 Repair-oriented view of same classes

a single hierarchy. For example, the process structure of a system is orthogonal

to the classes and their relations, since each process may be built from various
overlapping subsets of the same set of classes making up the behavior
components.
An object-oriented process execution means instantiation of a root object and
the execution of its initialization routine. Each root object thus gets a separate
 TASK 3: SELECT CLASSES AND GROUP INTO CLUSTERS 165

thread of control, and will in most cases trigger the creation of a series of new
objects before its execution terminates. Some of these objects may in turn start
other processes by instantiating new root objects using the same or different root
classes. (For a thorough discussion of the object-oriented execution model, see
[Meyer 1988a].)
A single partitioning of the classes to show system structure is usually not
enough when many processes are involved, since the same class may be used
and viewed differently in different processes. Therefore, each process usually
gets its own static and dynamic model. However, it is also possible to let a
number of root classes share the same static diagram if a partitioning can be
found that suits them all.
Process structure is related to the general issue of object-oriented concurrency,
whose research is still in its infancy. Until more is understood about what is
really needed in concurrent modeling, BON does not try to invent yet another ad
hoc notation to be used in all cases. Instead, the BON user is encouraged to
employ complementary notations adapted to the particular project to show more
complicated concurrent behavior and process structure. A very interesting
attempt to merge the process and class concepts can be found in [Meyer 1993b].

Gradual evolution of a good model

The method of first selecting some initial groups of classes, then trying to gain
more understanding by working locally with classification within each group, is
reminiscent of the common strategy for doing jig-saw puzzles. First you walk
around the table collecting pieces of a certain color, a certain pattern, a certain
shape, and so forth. Then you sit down and try to relate these pieces to each
other. Some of the pieces will turn out to belong elsewhere, but others will fit
and little by little the outline of a rose bush, a stone wall, an old castle will
become visible. Then a glimpse of the bigger picture may flash by. Since a red
petal can be detected in one of the cracks of the stone wall, you begin to suspect
that the rose bush is in fact close to the wall…
After modeling the local groups for some time, new insights have been gained
and it may be possible to go back and select some more classes from the
candidate list and fit them into one of the groups. Then, gradually, some of the
groups may be joined into larger groups, while others become clusters of their
own. The initial cluster charts and set of static diagrams are built successively,
and the system chart updated to reflect the current status.
The clustering criteria will often change several times before a final model is
laid down. Some clusters may be there already from the first task (delineating
system borderline), because certain reasonably independent subsystems were
identified early from the system requirements. But most clusters will grow out
166 THE BON PROCESS

of the increased understanding that comes from refining the initial model, and
represent views that were not initially apparent.
In the puzzle analogy, we might start to connect sets of pieces based on
various criteria (a certain mixture of color nuances, certain interlocking shapes,
etc.), which may allow us to build rather sizable structures in the form of disjoint
islands. After some time, however, we may see quite different patterns emerge,
some within groups (a pond in the meadow, a cottage in the wood), and some
extending across groups (a cart track passing the pond into the wood, the blue
sky partly visible through the foliage), and these may then be viewed as more
important than the initial criteria for completing the global picture.
The analysis and design task is of course different, since its pieces can be
combined in so many ways resulting in pictures ranging from the grotesque to
the divine. The risk of going in the wrong direction is therefore greater, but we
also have the advantage of being able to carve the individual pieces to make
them fit.
The static architecture at the end of this task will still be fragmentary, since
there is no point in trying to connect everything too early—better to gradually
refine the modeling of each cluster until the corresponding roles have become
clear. The global picture will be decided later.

7.6 TASK 4: DEFINE CLASSES

GENERAL RESULTS BON DELIVERABLES

• Classes defined by queries, commands, CLASS CHART

and constraints SSSSSSSSSSSS

SSSSSSSSSSSS
SSSSSSSSSSSS

SSSSSSSSSSSS

• Updated glossary of terms SSSSSSSSSSSS SSSSSSSSSSSS

INPUT SOURCES ACCEPTANCE CRITERIA

• System and cluster charts • End-user / customer acceptance
• Typical use cases
• Glossary of terms

Having selected and grouped an initial set of classes, the next task is to define
each class in terms of the information it can be asked to return, the services it can
be asked to provide, and the general rules that must be obeyed by the class and
its clients.
This is done by filling in the BON class charts, where the above translates to
queries which are functions that return information about the system state
without changing it, commands which do not return any information but may
instead alter the state, and constraints listing general consistency conditions and
business rules to be obeyed.
 TASK 4: DEFINE CLASSES 167

A short class description along with brief indexing information is also entered
(see the class chart layout in chapter 3). Naming conventions are quite important
for class features, and we will return to this issue in the next chapter.
Class charts may initially be used as structured memos to collect information
for further elaboration into typed descriptions, and later as documentation kept
up to date through automatic extraction from the formal interfaces.

Analysis and polymorphism

A typical sign of the hybrid character of most object-oriented analysis and design
methods published to date is their neglect of the real hallmark of object-oriented
technology, namely polymorphism. Since polymorphism and dynamic binding
are what really distinguish true object-oriented languages from the ones merely
supporting modularization and data abstraction, like Ada and Modula 2, one
would expect any object-oriented method to give this issue the attention it
deserves.
However, looking closer at the literature reveals that little attention is paid to
it. Readers may easily get the impression that inheritance and dynamic binding
are implementation issues, when nothing could be further from the truth. In fact
polymorphism is a very important aid by which the complexity of requirements
specifications can be reduced.
Without polymorphism, descriptions of similar system functions applicable to
many different but related objects need to be repeated in the class of each
specific object type. Furthermore, clients of such related classes need to have
one separate static client relation for each of the related object variants it may
call during execution. With polymorphism, the dependencies can often be
reduced by an order of magnitude (cf. figure 4.24).
So we should actively look for polymorphism of operations as early as
possible in analysis and design to increase clarity and enhance maintainability.
Since names in the class charts are inherited through the “Inherits from:” clause,
polymorphism may already be expressed in these charts. Both the static and the
dynamic diagrams in BON support polymorphism and dynamic binding (cf.
figures 4.20, 5.17, 5.22).

Glossary of terms
The glossary of technical terms produced earlier serves as a general guide for
naming features of the analysis classes. However, sometimes a problem domain
term corresponds to several features or some other name is chosen for reasons of
better clarity or consistency with other classes. Such deviations from the feature
names that would normally be expected by end-users and problem domain
experts should be documented to facilitate communication.
168 THE BON PROCESS

7.7 TASK 5: SKETCH SYSTEM BEHAVIORS

GENERAL RESULTS BON DELIVERABLES

• Object creation chart EVENT CHART SCENARIO CHART

• Event and scenario charts SSSSSSSSSSSSS

SSSSSSSSSSSSS
SSSSSSSSSSSSS

SSSSSSSSSSSSS
SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS
SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS

• Object scenarios SSSSSSSSSSSSS SSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

INPUT SOURCES
CREATION CHART Object scenario
• Early scenario charts SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS 1
• Major system functionality SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

2
• Typical use cases
SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

• Incoming and outgoing information flow

ACCEPTANCE CRITERIA
• Consistency with static model
• End-user / customer acceptance

In this task we start elaborating the dynamic model of our system. As a result of
task 1 (delineating system borderline) we already have an initial scenario chart
capturing the most important types of system usage. A vague idea of this usage
is always a prerequisite for finding initial candidate classes and selecting
between alternative views of the problem domain. At some point, however, a
more detailed and exhaustive picture of potential system usage can be of great
help in refining and advancing the growing static model.
Depending on the nature of the problem, this task is of variable importance
and may also be carried out very early. Therefore, as was pointed out at the
beginning of this chapter, task 5 is often merged with task 1 when the initial
requirements are expressed through many complex actions, but with less
apparent structure.

Object creation
Executing an object-oriented system means starting up one or more processes in
some operating system environment. Initially only one class instance exists for
each process—its root object. The root objects will then, directly or indirectly,
create all other objects needed to serve the requests issued by the system users.
Investigating which classes are responsible for creating new instances of which
other classes serves as a link between the static and dynamic models. Finding
out when new objects are created may help the developer spot new static
dependencies and hint at possible realizations of system operations in the
dynamic scenarios. The result is entered into the system creation chart.
Checking for object creation may be done using two approaches, and it is
often best to use both in parallel. First, go through all classes and for each class
think of situations in which an instance of the class may be created. For all such
 TASK 5: SKETCH SYSTEM BEHAVIORS 169

situations, find out which class will act as creator and add the class of the created
instance to the corresponding entry list of the creator. Second, start at the root
class(es) and try to decide which other objects may be created by the
corresponding root object. For all classes of the created objects, try to decide
which other instances they may create, and so forth.
The potential creation of a certain type of object by a group of related classes
can sometimes be factored, so that this creation is always handled by a common
ancestor. If this is the case, the class of the created objects is inserted into the
creation entry only for the ancestor, to avoid unnecessary repetition. This
applies also when the ancestor is a deferred class with no instances of its own.
Objects are introduced in a system execution in two different ways: by
creating new objects or by retrieving old persistent objects, that is objects that
were created by an earlier system execution and saved in a database or file
system. For an introduction to object-oriented databases, see [Cattel 1991].

Event and scenario charts

A system event chart is produced and the first version of the scenario chart from
task 1 (delineating system borderline) is refined. For large systems, it may be
desirable to split these charts into several groups depending on context (type of
user, subsystem). This may also be done for the creation chart.
Since the partitioning of a chart is not necessarily tied to a specific cluster or
user group, we have no special entry for the context in these charts. Instead, a
parenthesized word or phrase indicating the context may be added to the system
name in the headers of charts that are split into several groups.

Finding scenarios
The BON method concentrates on object-oriented analysis and design, and does
not cover activities in the very early phases of feasibility studies and domain
analysis often necessary to obtain the initial requirements of the system
[Prieto-Diaz 1987, Arango 1989]. Structured interviewing techniques applied to
users and domain experts are often helpful here [Scott 1991] as well as
familiarity with the current research results in the field of man−machine
interaction [Marmolin 1993].
A technique to find the desired system behavior called Object Behavior
Analysis (OBA) is described in [Rubin 1992]. It starts by first identifying what
needs to take place in general, the system behaviors, and proceeds to assign this
behavior to initiators and participants. These play different roles in the system,
which will help us understand which parts of the system need to take
responsibility for providing various services and managing system information.
The result is recorded in scripts and glossaries.
170 THE BON PROCESS

How many scenarios?

Except for extremely specialized systems, the primitive requests that may be
issued by a user can be combined in an (at least for practical purposes) infinite
number of ways. So our aim when listing scenarios is certainly not to try and
capture every possible usage. Rather we group them into sets of significant types
of usage, and let each type be represented by a scenario entry in the scenario
chart. By selecting an appropriate level of abstraction for the scenarios, it is
always possible to cover the full system usage in this way even if many of the
details will be ignored for each category.
The number of different scenarios selected depends very much on the type of
system and how the requirements are expressed. If a flat scenario structure is not
enough, grouping by context may be used as indicated above. Some of the
scenarios will later be illustrated by one or more object diagrams, but not
necessarily all. We should make sure to list each principle type of usage in the
scenario charts even if later we do not consider it worthwhile to illustrate them
all by object diagrams. Less ambitious pencil and paper checking covering a
larger number of scenarios may also be a good way of ensuring completeness
and improve the static structure.

Scenarios to guide static modeling

The study of object scenarios is closely related to static modeling. As recalled
from the discussion on use cases in the previous chapter, scenarios are mainly
used in BON as complementary guides to choosing a proper view of the world
and to continuously keep modeling on the right track. Although meeting the
requirements embodied in the scenarios is necessary for a successful project, the
emphasis should always be on the evolving static model populated with
increasingly stable and flexible abstractions that will solve not only today’s
problems, but also some of tomorrow’s.
Planned reuse is of course important, and not preparing at all for changes that
seem likely to occur in the future should be as unthinkable for an object-oriented
developer as stubbing out a cigar on somebody’s living room carpet. How much
effort to invest in each case is another matter, but ignoring the issue is just not
part of good engineering behavior. However, the future changes envisioned are
always more or less limited to simple analogies with what we already have.
Only a few human beings are capable of producing truly original thoughts, and
then only occasionally, so we must realize that many changes will never be
anticipated.
To meet this we must strive to develop an instinct for the kind of abstractions
that will increase the probability of unexpected reuse. This is somewhat like the
sense of positional strength that good chess players acquire. Although they have
 TASK 5: SKETCH SYSTEM BEHAVIORS 171

no chance to sort out all possible ways in which they may be attacked in the
future, they still know that certain formations of forces give great advantages.

7.8 TASK 6: DEFINE PUBLIC FEATURES

GENERAL RESULTS BON DELIVERABLES

• Typed class interfaces with public Class interface Static architecture
feature signatures and software contracts
(pre- and postconditions, class
invariants)
• Updated static architecture
INPUT SOURCES ACCEPTANCE CRITERIA
• Class charts with queries, commands, • Consistency with static architecture
and constraints • Features cover external events in event
• System and cluster charts charts and selected scenarios in
• Glossary of terms scenario charts

In this task, the informal class descriptions found in the class charts resulting
from task 4 (define classes) are translated into fully typed class interfaces with
software contracts. The queries become functions which return information,
while the commands become procedures which may change the system state.
The constraints translate into pre- and postconditions on the operations and class
invariants for the whole class. The signature of each public operation is
specified (types of input arguments, if any, and type of returned value in the case
of functions).
Since we are still dealing with the public features of the problem domain
classes, this task is considered the last part of analysis. In practice, most
contracting details will be elaborated during design, since it is often much easier
to express exact formal assertions about classes and operations when a more
refined structure has been decided.
However, quantifiable facts regarding the problem domain that are known
already at the analysis stage should if possible be expressed formally, since it
helps communicate a more exact picture of what is going to be needed to meet
the system requirements. But in most cases there will not be that many
quantifiable facts very early in the process.
Typing of features leads to client relations being discovered “for free”, since
an input argument or a function return value of a certain class type automatically
means there is a static client dependency between the class containing the feature
and the corresponding supplier class. This may in turn lead to remodeling and
updating of the static architecture. Moreover, thinking through possible return
types of queries usually increases problem understanding considerably.
172 THE BON PROCESS

The typing is in fact very important for advancing the static model, so unless
close collaboration with many users and other problem domain representatives
dominates the analyst’s work (and sometimes even then), a simplified typed
notation is often used very early on as a supplement to (or replacement for) the
class charts. Also, since the BON graphical notation for class interfaces was
designed with case tools in mind, it needs automatic support. Therefore, if no
tool is available the user is free to use any hybrid form of textual and graphical
notation that feels comfortable. The underlying specification is what is
important.

7.9 TASK 7: REFINE SYSTEM

GENERAL RESULTS BON DELIVERABLES

• New design classes Class interface Static architecture
• Restricted features
• Detailed network of clusters and classes
• More complete software contracts
INPUT SOURCES
EVENT CHART Object scenario
• Class interfaces SSSSSSSSSSSSS SSSSSSSSSSSSS 1
• Static architecture SSSSSSSSSSSSS

SSSSSSSSSSSSS
SSSSSSSSSSSSS

SSSSSSSSSSSSS
2

ACCEPTANCE CRITERIA
• Consistency with object creation chart
• Consistency with object scenarios Class dictionary
SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

This task begins the design part of the BON process, and will therefore include a
repetition of many activities already performed for the analysis classes, now
applied to new design classes.
As we have pointed out several times before, there is no clear difference
between analysis and design, because it is always a matter of opinion which
classes should be considered modeling the problem domain (analysis) and which
are part of the solution space (design). Sometimes the concepts used by people
in the problem domain carry over more or less completely to an analysis model,
and then the distinction may seem obvious.
However, the feeling is deceptive since there is no objective reality to
“discover”; every analysis model is a conscious choice of viewing the problem
domain in a way that is believed favorable to achieve certain purposes (and is in
fact the first step towards a solution). Therefore, if the analysis model
corresponds very closely to the problem domain, it may mean we have a natural
 TASK 7: REFINE SYSTEM 173

straightforward mapping, but it may also mean we have superficially accepted

too many concepts without thinking through their roles in a software system.

Compressed design tasks

Judging by the amount of text and number of BON process tasks devoted to
analysis, the reader may get the impression that this is the overshadowing part of
the work that needs to be done before implementation. However, this is just a
result of the iterative nature of object-oriented development. Sometimes, when
the problem domain is unknown to the analyst and the resulting design job—
once the problem is understood—turns out to be small, there may of course be a
bias towards analysis. But usually more time is spent on design compared to
analysis, since many more details have to be taken into account.
However, the reason we only have three design tasks in BON is that many of
the general techniques of finding new abstractions, elaborating old ones, refining
the class relation structure, and adding more software contracts, have already
been applied during the analysis tasks, so we do not need to describe them again
in connection with design. Also, the tasks that need most guidelines are
probably the initial ones, when the developer is still overwhelmed by seemingly
unstructured complexity in the problem area and requirements text. Once a
system model begin to stabilize, much more insight has been gained and work
can often continue by analogy with what has already been accomplished.

Proceeding into design

Refinement of the analysis model into design is done by adding new design
classes and working out more detail:
• new design classes with detailed interfaces
• new class relations
• more detailed contracting conditions
• possibly restricted features in highly interdependent classes
• refined clustering
The system behavior descriptions may be complemented by design-oriented
scenarios, especially for larger subsystems. Sometimes essential parts of the
system’s interface, such as which GUI packages to use, are postponed until
design. Initial feedback from end-users may then again become important, once
such decisions are taken.
As the system structure grows larger, more detailed consistency checking can
be made between static model and dynamic scenarios, including the life cycles of
174 THE BON PROCESS

important objects. If an object-oriented language with a good garbage collector

is to be used, the developer need usually not worry about destruction of the
transient objects during design, but for persistent objects destruction is a logical
problem which must be accounted for as part of the system behavior.

7.10 TASK 8: GENERALIZE

GENERAL RESULTS BON DELIVERABLES

• Updated class interfaces and static Class interface Static architecture
architecture
• More deferred classes
INPUT SOURCES
• Static model
Class dictionary
ACCEPTANCE CRITERIA SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

• Consistency with object scenarios SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS

Generalization is a continuous procedure that should always go on in the back of

the developer’s mind, since it is so closely related to pattern detection. Much of
it must be applied very early in the process, since having enough generality built
in from the outset affects the basic flexibility and understandability of the system
model. However, there is also strong pressure to produce working results in a
project, and some of the early class structures may have to be altered many times
before the model becomes stable.
Therefore, it is often too expensive to do all generalization work very early in
the process, because the frequent changes may render much of this work useless.
The best way to proceed is often not to use a very deep inheritance structure at
the beginning, while many alternatives may still be tried. Later, when the set of
classes and operations become more clear, a better structure can usually be
inferred as several superimposed patterns become visible. Multiple inheritance
often requires the developer to choose between many different possible
combinations, where it is not obvious which view is best. Having a more
complete picture of what is needed in the system often helps when making such
decisions.
There are two important types of generalization. The first type yields a
structure that pays off immediately by making the current system smaller, easier
to understand, or more efficient, while the second type is an investment for the
future by yielding a more robust system. How much effort to spend on the
second type must be a conscious decision, since its payoff is very much
dependent on the expected lifetime of the product, what kind of changes can be
envisioned, and the overall generality of the system components considered.
 TASK 8: GENERALIZE 175

It may often be a good idea to have a fully separate generalization phase

applied to a product directly after release and acceptance by a customer. That
way it will not interfere with the project time schedules, but still be applied
before the project members are scattered into new assignments and gradually
forget the finer details of the system.
Such a separate phase requires a budget of its own, and of course a general
appreciation in the organization for an act that may on the surface seem to have
negative productivity in terms of produced source code; the resulting system will
often be smaller after a successful generalization process. However, it can be an
excellent opportunity to capitalize upon the otherwise volatile knowledge gained
by the development team, and to do it without too much time pressure.
Another possible point in time for a generalization phase is just before a new
project is scheduled to begin. The project members will then be likely to have a
high motivation, since the results will be immediately useful in the new
development.

7.11 TASK 9: COMPLETE AND REVIEW SYSTEM

GENERAL RESULTS BON DELIVERABLES

• Reviewed static and dynamic models Class interface Static architecture
• Reusable class interfaces
• Class text templates
INPUT SOURCES
• Static and dynamic models
EVENT CHART Object scenario
• Creation and event charts SSSSSSSSSSSSS SSSSSSSSSSSSS 1

ACCEPTANCE CRITERIA
SSSSSSSSSSSSS SSSSSSSSSSSSS

SSSSSSSSSSSSS SSSSSSSSSSSSS
2

• Syntactic validation of each class

• Consistency of inherited and combined
class invariants Class dictionary
• Consistency of pre- and postconditions SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS

of polymorphic routines SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

In this final task we polish and complete the evolving models. The overall
system consistency is thoroughly checked, and the final documentation
produced. What language or languages to use for implementation is usually
decided before design starts, since it is important to have this knowledge as early
as possible. Hybrid object-oriented languages may limit the design approaches
that can be used without breaking the seamlessness. For example, heavy use of
multiple or repeated inheritance during design is not a good idea if the
implementation language will only support single inheritance.
176 THE BON PROCESS

Class text templates are often generated from the BON specifications as part
of this final task. Such generation should be as automatic as possible, preferably
done by a case tool, and future changes to the class text should be fed back to the
specifications to achieve reversibility.

Root classes
Definition of the root class (or root classes, if many executable processes are
involved) is usually postponed until the end of design. Root classes should
normally be small and easy to change when systems requirements change. A
typical extension that should always be considered is the inclusion of the initial
system as a subsystem to some other system.
When this occurs, it is of great help if the system top level was already
structured to facilitate such a transition. Ideally, one should only need to change
one or more root classes to make the system start in its new role as a subsystem
and perhaps add a few classes for communication with the new environment.

Merging and splitting classes

We make a final check that our classification has not been taken too far. If some
classes are found to be redundant, they should be removed or merged with other
classes. Very large classes or routines should be split using intermediate classes.
Implementation dependencies between classes are examined, which may indicate
that some public features should instead be selectively exported (only to a group
of related classes). Such decisions are of course highly dependent on the kind of
support offered by the implementation language.

Contracting consistency
We also go through all classes and corresponding features to finalize their
software contracting clauses and check for consistency. Care should be taken
that multiple inheritance does not violate the contracts by combining abstractions
that do not fit together. In the case of polymorphic features, possible
redefinitions must abide by the contracting laws; that is, preconditions may only
be weakened and postconditions may only be strengthened.

Completeness and minimality

By going through the event chart, creation chart, and object scenarios once again,
we convince ourselves that all behavior needed to take care of user requests is
embodied by the classes in the static model, and that no unnecessary classes or
features are specified.
 TASK 9: COMPLETE AND REVIEW SYSTEM 177

If a class or an operation is not used in the system it does not automatically

mean it should be removed; we could be developing reusable libraries or some
unused features could be included for reasons of symmetry and abstractional
balance. However, anything that cannot by traced back to at least one scenario
should be questioned.

User documentation
Finally, we stress the importance of producing high-quality user documentation
for all aspects of the system which are visible to the external world. This is not
an easy task, and is usually quite expensive in terms of time and skill needed, but
it is still well worth the investment. Particularly for an elegantly designed
complex system, this documentation is the only guarantee you have that at least
some of the users will actually understand and use what you have spent so much
sweat and creativity to design.
Use the most talented writers you can find among the people who know
enough about the system. Better to have a good writer with a reasonable
technical background to create the manuals, guided by an expert, than put a
technical guru with low writing skills on the job. (And we are not talking about
writers who are used to producing a six foot stack of formal documentation
conforming to current military standards here, but the ones whose letter to the
editor of the local newspaper will cause the politicians to install a bump in the
road in front of the day care center first thing next morning.)

Requirements traceability
To facilitate the tracing of functional requirements on a system into various parts
of its design and implementation, a requirements cross-reference should be
maintained. For each individually identifiable requirement, the classes and
clusters realizing the behavior needed to fulfill the requirement should be listed.
8 BON standard activities

This chapter discusses nine standard activities that a developer will be involved
with as part of performing the tasks defined by the BON process. The activities
are as follows:
1. Finding classes
2. Classifying
3. Clustering
4. Defining class features
5. Selecting and describing object scenarios
6. Working out contracting conditions
7. Assessing reuse
8. Indexing and documenting
9. Evolving the system architecture
These standard activities are orthogonal to the process in the sense that they may
all occur to some degree as part of many of the BON process tasks. The first
four are continuously repeated during both analysis and design. The fifth occurs
mostly during analysis, but may also be repeated during design of large systems.
The last four, finally, are chiefly part of the design tasks. We will look now at
each activity in turn.

8.1 FINDING CLASSES

A common question asked by many newcomers to the field of object-oriented
development is something like: “Well, this encapsulation stuff—inheritance,
polymorphism, and all—sure seems powerful, but how do I find the classes?”
Our first reaction when faced with this question, often at the end of introductory

178
 FINDING CLASSES 179

courses on object-oriented concepts, was one of astonishment. The question

might as well have been: “I am thinking of starting this company; I do not know
yet what it is going to do, but I want it to be successful. How do I find the basic
ideas of the enterprise?”
It seemed strange that anybody would hope for a simple answer. (No doubt
there are many books and courses which are more than willing to share their
secret of how to enter the Fortune magazine’s Global 500 list, but the only
detectable correlation between them and practical success is probably the
authors’ royalties.)
Object-oriented abstraction is completely neutral; it will carry out any kind of
encapsulation—good or bad, abstract or concrete—with full discretion, no
questions asked. What it can provide is a very clear and structured way of
expressing and gradually evolving good ideas, but the semantic content of good
ideas for realizing system behavior remains the result of talented and
experienced engineering. Only superficial details can be captured by rigid
procedures, while the hard parts—the ones that really count—can only be
mastered through deep understanding of the problem combined with careful
selection among many possible approaches.
However, there are a great number of partial solutions to common problems
and other tricks of the trade that should be familiar to any developer. Knowing
about them as early as possible will help speed up the long period of hard work
and study which is a prerequisite for becoming a good designer. (The other
prerequisites of talent and general attitude are more difficult to change, but
exposure to powerful ideas may sometimes have a deep effect on the latter.)
Each partial solution addresses a certain aspect of a problem and may serve as
guideline or general principle when problems of a similar type are encountered.
Familiarity with a large number of such principles and standard solutions to
small problems enables the designer to detect many common patterns during
problem analysis and re-examination of the evolving system model. This will
then provide constant feedback about how to proceed.
Precisely because they focus on different problem aspects, many guidelines
and principles are contradictory, and the basic engineering task is to decide
which aspect is more important in a given situation. We will try to cover some
of the principles and guidelines in this chapter, and we now return to the question
of how to find the classes, which is the same as finding good abstractions.

What is an abstract data type?

When we say that an object-oriented system is viewed as a structured collection
of abstract data types (implemented as classes), it is important to understand the
essence of the term. An abstract data type is an encapsulation of behavior—the
180 BON STANDARD ACTIVITIES

emphasis is on “abstract”, not on “data”. Therefore, one should not fall into the
trap of viewing abstract data types as protected data records with accessing
operations attached, which may be tempting for people used to traditional data
modeling.
On the contrary, the main purpose of an abstract data type is to remove all
trace of internal structure by defining data exclusively through behavior, that is
by defining the semantics of applicable operations [Guttag 1977]. And the word
abstract means that the available operations should not reflect the internal
structure, but instead capture the external properties of the underlying concept.
Only those aspects of an object that you are willing to let all clients depend on
should be part of the public interface, otherwise the class will just implement a
concrete data type.
The principle of hiding internal structure and concentrating on external
behavior when defining classes has been called “responsibility-driven design” in
contrast to “data-driven design” [Wirfs-Brock 1989].

The importance of names

So a class, viewed as an implementation of an abstract data type, should capture
the essential properties of a concept through precisely defined external behavior
(rather than by information stored). As in all other disciplines, consistent
naming of the concepts handled by a system is a major key to proper
understanding and efficient communication between its users and developers.
Good names act as helpful guides which automatically coordinate human
thoughts in the right direction, while inappropriate or misleading names are
constant distractors highly increasing the risk of misunderstanding, no matter
how brilliant the underlying concepts.
In fact, the names of the classes and their operations make up the words and
phrases of the language in which the application is expressed. And anybody who
has been involved in language design knows how important even small details
can be for the resulting clarity and ease of expression. Therefore, naming is an
important part of systems design which should not be taken lightly. Designing
software means producing a description aimed for others to read and understand,
both humans and machines. Machines will not get confused by bad naming, but
humans will (and not just other people but also the designer six months later).
So understandability is always the first concern and is directly tied to the
resulting quality of the software. The personal egocentric naming style must
yield once and for all, and be replaced by a reversed attitude. It is not the reader
who is responsible for understanding what a program or a design is all about, it is
instead the responsibility of the writer to make sure that everything is expressed
so clearly that it will be impossible for any reader to get the wrong impression.
 FINDING CLASSES 181

Of course a certain background must be assumed from readers and there is

always a cost and a limit associated with everything. The importance of precise
understanding will also vary from one case to another, but the paradigm shift is
crucial. Describing software should be like writing a political article for a daily
newspaper: it had better be convincing to make a difference.
Consistency of style is extremely important for understandability. Therefore,
when you are maintaining someone else’s text, unless you are prepared to rewrite
everything, it is usually better to adopt the general style of the document even if
you consider it inferior to your own.

Naming classes
Classes describe objects that encapsulate behavior. An object does not represent
a behavior but should be thought of as an actor that will respond to various
messages with well-defined behaviors. In rare cases a class may only define one
public operation, but we should then always be able to envision more in the
future. If not, the class is probably not an independent concept and should
instead be merged with other classes.
For this reason, class names should be nouns with or without qualifying
adjectives, like WINDOW, BUTTON, or RED_BUTTON. Sometimes single
adjectives could be used as abbreviations, but then there is always an implicit
noun in the background. However, since class names are usually global to a
system, names like HELPFUL or LAST are almost always too general to be
appropriate. Clearly, this type of qualifier may be applied to more than one
concept in a system. For class features, on the other hand, where the enclosing
class defines a local context, names like the above are common.
Class names should always represent one object of the corresponding type,
and thus use singular form. The reason is simple: since every class describes a
possible set of objects, the alternative would be to always use the plural form.
This would not make sense, since it would only yield longer names without
adding any information. Also, since signature types are often specified one
entity at a time, using the plural form for class names would make normal type
declarations look utterly confusing:
favorite_book: BOOKS
gives the impression that not just one but a whole set of books are preferred. (It
is still possible to use the name BOOKS to signify a class handling a whole
collection of books, but in such cases BOOK_SET or SET [BOOK] is more
clear.)
Much more can be said about how to choose good and consistent names for
classes, but most of it applies to any naming used to classify concepts in some
182 BON STANDARD ACTIVITIES

problem domain (in fact, this is exactly what the classes are supposed to do). For
a general text on classification principles see for example [Ogden 1923].

Analysis classes and design classes

A class can represent any concept that may be viewed as a black box with a
specified behavior. In the analysis phase we form a perception of the problem
domain, which will then serve as a basis for subsequent design. This is the
analysis model and the classes are called the analysis classes of the system.
They should normally be expressed in problem domain terminology and the
corresponding concepts along with the visible operations defined should make
sense to users and domain experts with respect to the activities modeled.
As was pointed out before, this is not an objective model of reality but a
subjective view chosen with the intent of facilitating future design and
implementation of a system whose basic behavior is known. It is in fact the
beginning of a design. If too little is known about the desired system behavior,
we will not have enough information to decide how the problem domain should
be viewed, and then we must instead continue our study of the general
requirements.
The new classes that are added during design and implementation are
background classes, which deal with concepts that are not of immediate
relevance to the user’s perception of the system. These may be called design
classes and implementation classes respectively. If the separation of analysis
from design is not always easy (since analysis always encompasses an element
of design), drawing the line between design and implementation is even more
arbitrary. Since both activities deal with finding a suitable representation for an
analysis model on some computing equipment, any precise separation is mostly a
matter of taste.

Tangible objects
Some authors seem to imply that the analysis classes linked to the problem
domain should reflect tangible objects. Tangible objects (things you can
perceive with your physical senses) are important in the physical world, since
they convey a sense of reality and may affect us in a very direct way.9 However,
in the electronic world this difference disappears; a class representing a car is no
more tangible than one that models the job satisfaction of employees. What

9
In a deeper sense, a bus is of course also an abstract concept made up by the human mind for
practical reasons, but when it hits you crossing the abstract street without looking at the abstract
lights, you will know the difference.
 FINDING CLASSES 183

counts is how important the concepts are to the enterprise, and what you can do
with them. The assumption that tangible objects should be generally more stable
than non-tangible objects is also false. You only need to take a look at the piles
of last year’s models of physical products at the city dump to see how stable they
are.
Moreover, tangible objects are very often manifestations of more abstract
underlying ideas which are easily missed if modeling is centered on physical
appearance. Of course, software may be directly in control of expensive or
dangerous physical equipment whose classes may then play a special role in the
system, but this is no different from classes in charge of abstract information
whose correctness may be life-critical in, say, an intensive care unit of a hospital.

Deferred classes
Deferred classes, that is classes specifying one or more features without
providing implementation for them, serve as design skeletons to be completed by
descendant classes. These may then tailor the partially defined behavior in
desired directions. Specifying strong postconditions for such deferred features
restricts the semantics of the tailored versions, ensuring that important principles
will still apply in the descendants. Using correspondingly weak postconditions
will exert less control, but instead leave more room for variation.
The general principle is not to overspecify, but include as postcondition only
what must be expected of every conceivable version implementing this feature.
On the other hand, whenever such constraints do exist we should take care to
specify them since they represent important semantic information.
If all features are deferred, the class comes very close to an abstract data type.
As the system model evolves, more deferred classes tend to be added (usually by
insertion of more levels in the inheritance hierarchies), since this provides more
generality and makes it easier to meet future requirement changes.

The dual roles of a designer

It is important to realize that designing object-oriented software implies that two
very different roles have to be fulfilled at the same time: the client role and the
supplier role. Like the right and left hand of a pianist they must work in concert,
but at the same time independently.
The client role deals with selecting the basic building blocks to use, and the
supplier role deals with how to combine these building blocks to achieve a
desired behavior. Designing means constantly switching between these two
roles. Let us examine the implications a little closer.
Design usually proceeds by refining several classes at a time, collectively
representing some layer of abstraction. However, to get the general idea it
184 BON STANDARD ACTIVITIES

suffices to look at a single class. Assume the external behavior of this class has
been fully specified, and our job is to design its internal workings. A typical
approach would be the following.
Assume the supplier role. Look at the operations of already available classes
and try to tentatively combine them to see if they are sufficient as basic
components for the behavior we are trying to build. If we can envisage a
reasonably simple implementation using only these operations, we are done, else
we switch to the client role. (The duration of this first attempt may be anything
from days for a complicated class when we already have a large stock of suitable
supplier classes, down to less than a second if we have nothing appropriate.)
As client we now try to think of what new abstractions would be suitable for
conveniently expressing the missing parts of the behavior. As argued in a
previous chapter, this can be viewed as designing a mini-language for the type of
behavior we are modeling. Depending on the generality and potential power of
the new language we come up with, our focus of interest will switch from the
problem to the new abstractions for a longer or shorter period.
Recall that an essential part of object-oriented development should always be
investment for the future. Therefore, we should actively seek abstractions that
do at least a little more than just solve the present (mostly rather special)
problem. This means that refinement of new abstractions to take advantage of
their future potential is a natural part of the process. We must of course take care
not to spend a large part of the project budget on costly deviations from our basic
task, but even the short-term revenue is often enough to finance a little extra
work. In the long run, this attitude will pay off handsomely.
In fact, after some elaboration many such independent abstractions will yield
insights and new potentials that may even affect the initial requirements. If the
new outlook enables a whole new set of facilities more or less for free, which
will solve the user’s problems equally well or much better, even the most hard-
headed customer may rapidly switch attitude.
So after working locally with the abstractions until they are reasonably
understood, we again assume the supplier role and start combining the new
operations to fill the holes in our behavior model. If the behavior we are
designing for belongs to the analysis classes, part of our activities in the supplier
role will be to sketch object scenarios. We then keep switching roles back and
forth until we have completed the desired behavior.
When we think of new abstractions in the client role, an important point is not
to strive for an optimal solution to every single problem (as in traditional design
by stepwise refinement [Wirth 1971]), but to always take into account what we
already have. First, we need to take care not to solve the same problem twice
with only slight variations—much better then to invest some effort obtaining one
version general enough to be used in both cases. Second, choosing abstractions
 FINDING CLASSES 185

is always a tradeoff between quality and the cost of development. “The best is
the enemy of the good,” as observed by Voltaire.

A world of recursive behavior

Abstract data types transform data into behavior, and as client we deal
exclusively with the external behavior of our building blocks. In this role, we
should therefore never think of data as bits stored in fields inside various objects,
but only as something abstract which will manifest itself indirectly, through
well-defined behavior. Where the knowledge comes from that enables our
supplier objects to behave in the proper way is none of our concern as client.
Since we are also familiar with the supplier camp (our second nature as
developers), we do know what usually goes on behind the scenes: our suppliers
are in fact jotting down pieces of information here and there. However, if some
of them were able to use divine inspiration instead, we would not care as long as
they fulfill their contracts.
As suppliers, on the other hand, we are concerned with combining what we
have into some behavior that we have undertaken to deliver. This is a very
different role. To fulfill it we need two things: a set of references to a number of
useful objects to call, and the knowledge of how and when to call them. This is
the only information each object must have—nothing else is needed.
We will use an analogy with telephone calls as an illustration here, and we
will also allow ourselves to become anthropomorphic for a while (viewing the
objects as intelligent creatures). Contrary to Edsger Dijkstra’s well-known
assertion that “the use of anthropomorphic terminology when dealing with
computing systems is a symptom of professional immaturity” [Dijkstra 1982],
we believe it can be very useful at times.
So think of the set of recorded object references as numbers in the supplier’s
personal telephone book. Its content will be the information stored in each
object. Invoking an operation on an object corresponds to picking up the
receiver and dialing a number in the phone book, either to ask the answering
object to do something or just to receive a new telephone number. If required,
we start the conversation by reading some of the other numbers in our phone
book to the answering object.
Amazingly, the only thing an object ever does (inside the system) is call other
objects and exchange telephone numbers—nothing else. At first, this does not
seem to make any sense at all—where is the useful work? What benefit can we
possibly get from a system that can do only one thing: juggle telephone
numbers—albeit with stunning dexterity?
Well, the answer lies of course in the external communication. A completely
isolated object-oriented system is no more useful to the surrounding world than a
186 BON STANDARD ACTIVITIES

configuration of elementary particles that have never been measured and thus
never confronted by laboratory workers. Therefore, some telephone calls will
actually be answered by special kinds of objects that have some magic way of
making things happen in the physical world.
If asked to, some of them may convert a telephone number into patterns
visible to the human eye as glowing fluorescence on a video display screen or
traces of carbon powder burned into paper. Others will sound the lunch bell; yet
others may send a whole battery of missiles to sink a hostile fleet of warships.
So the system’s juggling of phone numbers has but one purpose: to make these
externally connected objects do the right things at the right time.
But what about internal value representation? Whatever happened to our
familiar integers, floating point numbers, and strings? The answer is that various
value representations are only useful concepts when thinking about internal
implementation algorithms. Every implementation has to choose some basic
representation, but when a system is viewed from the outside, there is no reason
to treat basic values any differently from other objects. They can only be
experienced indirectly anyway—through the physical behavior of some
externally connected device.

Encapsulated behavior
The reader may consider the last subsections as a deviation from the main
discussion of this first BON action, which is how to find classes. However, we
think the view presented helps to get a deeper understanding of what is really
behind the object-oriented ideas.
One thing that should be clear after this discussion is that it is very dangerous
to talk about object-oriented encapsulation as being centered on data, unless it is
perfectly clear to the audience that data is to be interpreted in the abstract sense,
as well-defined behavior. This is something entirely different from the view of
encapsulated records of data fields, which focus on internal structure.
Since data is nothing but potential behavior, there is nothing strange or
stylistically wrong with encapsulating a typical process in a class, even if the
system state affected by the process is not maintained by the objects of this class.
A class is not obliged to record hidden information to become a respected citizen
in the object world; if it can fulfill its behavior without taking notes, all the
better.
For example, if a parsing operation for some language syntax is needed, it is
not reasonable to make it a feature of a general string handling class, since the
fairly specialized parsing facility can hardly be regarded as a natural operation
that should be part of any string. On the contrary, an important principle of class
design is to include only features that are intimately linked with the abstraction
 FINDING CLASSES 187

maintained by the class. Mixing several abstractions in one makes reuse more
difficult.
One possibility would be to put the specialized behavior in a descendant class,
say PASCAL_STRING, which inherits from STRING and adds the ability to parse
Pascal syntax. However, this is not the natural way to view things, since one of
the main purposes of a parser is to decide whether a string actually conforms to
the given language. Therefore, the parser must be able to analyze any string, not
just those that are already known to be correct. (The parser would actually have
to be called already at object creation in order to ensure the class invariant!)
There are many cases like this, where behavior is best encapsulated with no
internal data, and services instead applied to objects passed as input arguments.
Such classes may be termed tool classes, as they supply general facilities
applicable to objects that need them (cf. facility inheritance in 8.2). Of course
the most common classes are still those which maintain some internal
information, but we are covering the less obvious cases here.

How many features?

The appropriate number of features in a class may vary a great deal. It all
depends on what abstraction is encapsulated. Some classes maintain information
that is often requested in many different forms by clients and may need quite a
few features to do it. This does not necessarily mean that such classes become
difficult to understand, since understandability is more a function of how many
different concepts you are faced with than sheer volume.
For example, a document handling class that contains 100 separate operations
to set various font options (we are not saying this is the right way of
encapsulating fonts, but not necessarily wrong either) may in fact only be dealing
with one or a few underlying concepts which are quite familiar and easy to grasp.
Ease of selecting the right operation is then reduced to having nicely organized
manual pages.
On the other hand, a class with an order of magnitude fewer features whose
semantics depend on each other in some intricate manner may be much more
difficult to understand and maintain. The most important principle is not to mix
separate abstractions in the same class. The average size may be 10−20 features
per class. If you have 50 features it may still be acceptable, but you should be
prepared to defend your position with good arguments.
A class should not contain too few features either, since this is often a sign of
weak abstraction. Either vital features may be missing, or the class should in
fact be merged with other classes. If a class has only one feature, something is
nearly always wrong. We are of course referring to effective classes here; top-
level deferred classes may have perfectly good reasons to have only one feature,
or even none at all.
188 BON STANDARD ACTIVITIES

Completeness and layered design

A class should be complete in the sense that enough operations are provided to
effectively use all aspects of the underlying abstraction. If a container class has
an insert operation, it must also have a corresponding remove operation,10 even
if that feature is not needed by the system giving rise to the new class. The extra
effort is a small price to pay for the enhanced future reusability.
A good principle is to design a basic layer of primitive features which are
sufficient to guarantee completeness of the class, but which are also necessary in
the sense that they cannot easily be expressed in terms of each other. This gives
a lower bound for what must be implemented right away, and a solid base for
understanding the abstraction.
For many classes, however, the basic layer will not be enough to express
commonly wanted behavior without tedious repetition of primitive operations.
Then we should add a second layer of convenience features, which are
exclusively phrased in terms of the primitives. These should be visible in the
class interface layout as a separate group, which are not logically needed to
understand the class.
Following the principle of layered design also makes it much easier to
redefine features inherited from complicated classes and to add new ones.

Class or operation?
A common decision which is not always easy to take is whether a certain
behavior should be implemented as a separate class or just as an operation in a
class. There is no simple answer to this question, since so many factors may
influence the choice. Some criteria are suggested in [Halbert 1987]. For
example, any of the following may be a sign that a new class should be created:
• The behavior can be shared and reused by more than one class.
• The behavior is complex and difficult to relate naturally to an already
existing class.
• The behavior is very specialized, and even if it could be related to an
existing class only some clients of that class will want to use it.
In the last case the behavior should probably be put in a descendant class.

10
Unless, of course, we are modeling a special container from which nothing should ever be
removed, like for example a (non-political) history log.
 FINDING CLASSES 189

Classes for high-level specification

Already at the topmost analysis level we need some basic concepts to describe
how things in the problem domain are structured. Support for this must be
included in any analysis notation, and BON uses the clustering mechanism
combined with client and inheritance relations. However, much of the additional
notation included in other approaches, such as semantic labeling or multiplicity
markers, is not necessary in an object-oriented approach.11 Since the class
concept (whose main purpose is to capture abstraction) is already there, it can be
used directly to express more detail. Therefore, we should define a standard set
of generic classes, such as
SET [T], SEQUENCE [T], SORTED_LIST [T],
TABLE [T, KEY], QUEUE [T]…
for use as specification elements for very high-level information structures. As
we have seen in chapter 4, such generic classes may then be attached as labels to
relations between other classes in BON static diagrams. Other candidates for
high-level specification are classes dealing with:
• finite state machines: STATE [T], TRANSITION [T]
• commands: COMMAND
• units of measurement: TIME, DATE, UNIT, AMOUNT
• various types of value: VALUE, NUMERIC, BOOLEAN
• domain-specific general categories
The precise set chosen depends very much on the application, but it is important
to make it consistent, orthogonal, and well defined. The great advantage of this
approach is that the specification elements are easily extensible and can be given
a precise definition through assertions in the corresponding classes, as opposed
to ambiguous labels expressed in natural language.

8.2 CLASSIFYING
The classification activity consists of combining a set of selected classes to see
how they may be related in terms of similarity and collaboration. This means
finding inheritance and client relations between classes that we already have.

11
BON allows semantic links as structured comments in more complicated cases, but these are
used sparingly and are not very important.
190 BON STANDARD ACTIVITIES

(Finding classes in the first place is of course also classification, but that is
considered an activity of its own and is not included in the BON classifying
activity.)

Inheritance vs. client usage

Very often we have a choice between reusing the facilities of some class by
either inheriting from it, or using it as client. The question of when to do which
is by no means trivial, since both have their pros and cons. The basic rule is that
when you are satisfied with the services offered by a class you should just use it
as client, while if you need to alter part of its behavior you should inherit from it.
As client you are comfortably shielded from the implementation of a supplier
class. You know that as long as the supplier interface (including the software
contract) does not change, its internal workings may evolve without affecting
you at all. On the other hand, the specified behavior is then all you get, while
inheritance gives you the possibility to tailor inherited features (as long as your
modifications do not violate inherited contract clauses).
Inheritance therefore implies greater freedom, but as a consequence you are no
longer protected from changes in the ancestor’s internal structure (freedom with
responsibility). Some object-oriented languages offer special mechanisms that
can be used by a class to limit the freedom of descendant classes (called private
mode in C++). However, using such protection mechanisms may do more harm
than good. As was argued in chapter 6, if we limit reuse to what can be foreseen
we are in trouble. Therefore, an ancestor should not try and figure out
beforehand in what ways its basic concepts may be reused in the future.
Prohibiting descendant classes from accessing certain parts of the ancestor’s
internal structure introduces restrictions that may easily lead to code repetition or
unnecessarily complicated circumventions in future classes. A descendant class
must still know exactly what it is doing and take full responsibility, since it is not
possible to modify inherited features in a safe way if the ancestor’s
implementation is treated like a black box. Important principles that should
never be violated by any descendant class are instead put in the contract clauses
of inherited specifications.
A typical situation when it is more advantageous to inherit rather than be a
client is when several features implemented by some other class are to be offered
directly by the new class as well. If we are just a client of the old class, every
reused feature that is going to be part of the new interface needs to be
reimplemented, since it is not possible to re-export features from a supplier class.
Even if each implementation is only a few lines (since it can call the
corresponding features of the old class), the repetition of many features can be
detrimental to clarity and future maintenance.
 CLASSIFYING 191

This often occurs in connection with multiple inheritance. A common

example is a class WINDOW that inherits both from TREE and RECTANGLE,
the first ancestor implementing the structural properties needed for nested
windowing, and the second implementing graphical properties like move, resize,
and so forth.
The effect of using inheritance here (even if none of the features inherited is
changed in any way) is that all the reused properties can be exported by
WINDOW without a single line of new code. Being client, on the other hand,
would imply reimplementation of a large number of features. (Such repetitions
are quite common with single inheritance, since only one aspect of a many-
faceted concept may then participate in an inheritance structure.)

Facility inheritance
Another typical situation in which inheritance may be preferable occurs when a
class encapsulates a set of functions whose results depend only on the input
arguments, not on any recorded internal state. For example, the standard
trigonometric functions are often contained in a class that is simply inherited by
any class requiring the facilities (which again assumes multiple inheritance).
The alternative—to declare and instantiate a dummy object just to use it for
calling these functions—is somewhat awkward.
p: TRIGONOMETRY
<create a new instance and attach to p>
..
.
a := p.cosine (b)
The required qualification and extra lines of code look confusing, and the
unnecessary object creation may be resource consuming if the encapsulated
feature is called in a tight loop. Inheriting from classes that add only functions
but no data fields may cause some extra work for the compiler, but should not
cost anything at run-time.
Since inherited operations become part of the local name space of a class (and
of all its descendants), the names of such general utility functions should be
easily recognizable to prevent confusion with names of features of the main
abstraction encapsulated by the class. If this is not the case, either renaming or
the dummy object approach should be considered.
When a class is designed for heavy reuse, it may also be important to limit the
amount of facility inheritance. Clients to the class will not be affected (unless
the inherited facilities are reexported), but the corresponding names must be
taken into account by descendant classes and may interfere with later extensions
in new directions.
192 BON STANDARD ACTIVITIES

Inheritance for grouping

It is often a good idea to isolate interface details to external software (such as
database systems and graphical packages) in separate classes to facilitate future
alternative implementations. However, the model offered by the external
operations may have to be completed by system abstractions before it can serve
as a coherent encapsulation of the external facilities in a class. These added
abstractions are probably not entirely independent of the external model, but
often independent of many of its smaller details.
Therefore, encapsulating the features that contain direct calls to the external
operations in a separate implementation class, and letting this class be inherited
by another class containing the necessary additional abstractions, can often
enhance clarity. Inheritance is then used for grouping rather than for potential
reuse, but if the external system is for example ORACLE version 6, having a
common prefix for each of the implementation classes, like for example
ORACLE_CURSOR and ORACLE_WARNINGS, makes it easy to see which
classes must immediately be checked when version 7 is later installed.

Inheritance of secondary properties

The possibility to inherit secondary properties from other classes is an essential
advantage. By secondary, we mean that these properties are not the ones
primarily associated with the abstraction encapsulated by the class. They may
nevertheless be important.
For example, if class FIGURE inherits from class PRINTABLE, this means the
instantiated graphical objects will know how to print themselves. Using the
name PRINTABLE rather then PRINTER for the class specifying the print
operation indicates that we do not really view classes inheriting from it (like
FIGURE above) as little printers, but as objects which among other things also
can take care of their own display.
The distinction is important, because a property like being printable is seldom
something we want to use as a main classification category for our classes.
Keeping a somewhat low profile is therefore essential so as not to disturb the
main conceptual structure, particularly since there may be quite a few such
properties around. (Classes inherited to acquire secondary properties are called
mixins in Common Lisp Object System, CLOS, environments [Gabriel 1991].)
Because the secondary properties are so general, the ancestor classes will
often just contain a specification for them, while their implementations are left to
descendant classes. In typed notations, such abstract specification plays an
additional major role by permitting constrained generic derivation.
Consider a general list class which stores its elements in sorted order. Such a
class must require that the elements to be stored have an order relation among
 CLASSIFYING 193

them, otherwise sorting will not make sense. To ensure this, it can be specified
as a constrained generic class SORTED_LIST [T −> COMPARABLE]. The
actual generic parameter T must then be a descendant of class COMPARABLE
known to contain the required feature "<" (infix form of less_than), which will
guarantee that sorting is possible. The "<" operation will be deferred in class
COMPARABLE, and later defined for each descendant class. So by simply
inheriting from COMPARABLE and defining one small operation, classes like
PASSENGER_AIRLINER and AIR_FREIGHTER may each take part in sorted
lists very elegantly. The first may, for example, define order by number of seats
and the second by tons of load capacity.

Beware of reverse modeling

A common trap that beginners in the object-oriented field sometime fall into is
modeling the world upside down. Figure 8.1 shows a typical case.

FACILITIES

DECK_CRANE STOWAGE_SPACE

WHEELHOUSE STEERING_ENGINE

CARGO_STEAMER

Figure 8.1 Incorrect modeling of attributes

Since multiple inheritance makes it possible to easily merge properties from

other classes, some designers may be tempted to use inheritance as a means of
gradually including all facilities needed in a class. This is somewhat tricky,
because it might actually work by accident (all the way through to
implementation), until future enhancement attempts show that something is very
wrong in the model.
So why is the design in figure 8.1 so wrong when we have just discussed
legitimate inheritance of secondary properties and general utilities? Well, the
crucial difference is that although we considered the earlier inherited properties
194 BON STANDARD ACTIVITIES

as secondary, they were still properties of the whole objects described by the
class. When class AIRLINER is made a descendant of class COMPARABLE, the
inherited features define an order relation applicable to the whole aircraft, while
the features inherited by CARGO_STEAMER were certainly not meant to apply
to a whole ship.
For example, if class DECK_CRANE has a public operation turn, this will
become part of the interface of CARGO_STEAMER. Therefore, when trying to
avoid an iceberg suddenly emerging out of the thick mist, all the skipper will
have time to see before being thrown head first into the Labrador Sea is the crane
cheerfully turning its jib across the deck.
The classes of the FACILITIES cluster represent parts of the freighter rather
than aspects of it, and should of course be used as suppliers instead of ancestors.
The turn operation in class CARGO_STEAMER would steer the vessel, while
another operation turn_crane is implemented by calling turn on an attribute of
type DECK_CRANE. In this way, the freighter is modeled as an aggregation of
its parts, which is the correct view.
Note that keeping the inheritance and fixing the name clash by renaming turn
into turn_crane would only hide the modeling error temporarily. Inheriting from
a set of classes means it must be reasonable to view the descendant as a special
case of each of the ancestor classes simultaneously. All inherited contractual
obligations (feature postconditions and class invariants) must also be fulfilled by
the descendant. Thus, the risk of conflict is high when combining abstractions
that do not represent orthogonal concepts.
However, there may not be that many contract clauses specified initially to
sound the alarm, and this makes the reverse modeling trap dangerous
(particularly if the situation is a bit less obvious than in the example above).
Things may seem perfectly normal at first, just as several people can work at the
same desk with entirely different problems until they start mixing things up and
use each other’s private notes as a basis for their own decisions—then the
seemingly innocent procedure suddenly turns into chaos.

8.3 CLUSTERING
Object-oriented structuring is a way of organizing functionality. Any large set of
concepts needs to be structured in several levels to be comprehensible to the
human mind. The first and most fundamental level is the class, which
encapsulates a set of related functions in such a way that the resulting group can
be viewed as describing a whole new concept. Such concepts, corresponding to
abstract objects with well-defined behavior, can then be used as larger mental
pieces by which to understand the world.
 CLUSTERING 195

The human brain has an amazing capacity to relate different behaviors to each
other by detecting patterns of similarity and inventing new names for whatever is
behind such similarities. Therefore, matching this capacity requires a very large
set of highly overlapping abstract concepts captured by classes. If each of the
overlapping classes were to be described from scratch, the resulting descriptions
would become too large, and we would not be able to remember enough of them
to explain a complicated scenario.
Here is where the second structuring level comes in, namely inheritance. By
describing the behavior of a new object type only in terms of differences from
the behavior of existing types, large groups of related abstractions can be
understood with an order of magnitude less effort. This already goes a long way
towards what is needed to build complicated systems, but is still not enough. We
need a third structuring facility to group sets of collaborating object types (not
only similar types) into smaller recognizable units, which can be used as
elements of a higher-level system description.
In BON, this unit of description is the cluster. Recursive clustering can be
used to impose a hierarchical structure on the set of classes in a system, which
we call a system view. Hierarchies have proven very useful in many
classification areas as an effective and simple way of organizing large sets of
concepts for easy overview and reference. Even very large sets require relatively
few hierarchical levels for complete classification of their elements.12
However, since the classes in a system need to be used freely by other classes
(the more general the abstraction, the more important the reuse), we cannot let
the hierarchical partitioning be strict. Doing this would mean letting in again
many of the drawbacks of traditional approaches through the back door.
Therefore, the system view of hierarchical clusters is imposed on the classes
independently of the basic client and inheritance relations between them.
Compared to classes and their relations, the clustering represents a much
looser structure that should be easy to modify from one system version to
another. This enables the combination of systematic order and adaptive freedom
needed for flexible systems. With a relatively small effort and perhaps some
new or modified classes as additional glue, clustering may be used for
programming-in-the-large [DeRemer 1976] using sets of existing classes.

12
For example, to classify 10,000 elements with an average of 22 elements per cluster, only 3
levels are needed.
196 BON STANDARD ACTIVITIES

Cohesion and coupling

That clustering is not bound by class relations does not mean there is no
connection between the two. On the contrary, several important principles
apply, but none of them is absolute. One such principle (mentioned already in
1968 in a short essay by Larry Constantine [Constantine 1968]) is to maximize
cluster cohesion and minimize cluster coupling.
Maximizing cohesion means putting classes which collaborate closely in the
same cluster, while minimizing coupling means having as few dependencies as
possible between classes in different clusters. The rationale behind the principle
is obvious: a cluster defines a local context in some sense, and it is desirable that
local contexts can evolve in parallel without interfering too much with each
other. On the other hand, to be comprehensive as a whole, related abstractions
should be kept together.
However, we stress again that the principle is only a guideline. Some
commonly needed abstractions (notably the fundamental data structures) need to
be easily available in many contexts and should therefore not be hindered by any
rule of low coupling. This calls for extra high quality and stability of such
abstractions, since large parts of any system are likely to become dependent on
them, but the resulting reuse is much more important in this case.
A general coupling rule which is less difficult to follow is to restrict cycles
between classes (where both classes use facilities from each other) to occur only
within the same cluster.

Interface classes
A good way to decrease dependencies between clusters (and emphasize the
needed ones more clearly) can be to collect all externally accessible features in
one or more classes serving as the cluster interface. This is particularly useful
for clusters encapsulating subsystems or black-box frameworks (see below).
The interface classes may also serve as the subsystem specification, which is
an advantage when clusters are assigned to parallel work groups. The rest of the
classes in the cluster are support classes whose purpose is to provide additional
internal abstractions needed by the interface classes. (If, as is often the case,
generally useful classes are discovered among the internal abstractions, these can
be moved to some library cluster for use in other parts of the system as well.)

Client−server interfaces
When even greater separation is desirable, as is becoming more and more
important in large distributed applications, the whole subsystem may be viewed
as just one class by all other system parts. An instance of this class becomes a
 CLUSTERING 197

server object with a set of visible operations available, and these operations may
be invoked through some standardized mechanism used for all client−server
communication in the system.
Such invocation is typically implemented by symbolic message passing, and
may involve general dispatcher objects known as traders or brokers
[Marshak 1991, OMG 1991] as well as client and server proxies if the
application is distributed. In the latter case, proxy classes can make
communication completely transparent to application classes, whether data just
passes between adjacent objects in the same address space or is in fact converted
up and down the communication layers and passed between heterogeneous
applications in a wide area network. In either case, the subsystem will only
appear as one abstract object to the rest of the world, and what happens behind
the scenes will not affect the clients at all, as long as each party obeys the server
contract.

Same level of abstraction

The classes in a cluster should usually be kept on roughly the same level of
abstraction with respect to the concepts they handle. If a cluster VEHICLES
contains the classes CAR, TRUCK, BUS, ENGINE, BODY, and WHEEL, one
might consider breaking out the last three and putting them in a separate cluster
VEHICLE_PARTS. If the classes of the VEHICLE cluster then turn out to be the
only users of the VEHICLE_PARTS classes in the system we are modeling, the
latter cluster may be nested inside the first.
To obtain uniform levels of abstraction, we usually need to create several
clusters that represent only conceptual groupings of classes. This is often good
during early analysis to render more visibility to the roles played by the analysis
classes. However, during detailed design and implementation this rule should
not be taken too far since the resulting fragmentation at lower levels may do
more harm than good.

Frameworks
Encapsulating common types of behavior in subroutine libraries only works for
behavior that can be expressed by simple combinations of fixed algorithms
selectable by a small number of input arguments. Whenever common patterns
are instead, as is often the case, more intricately laced with the surrounding
behavior, the only way to avoid complete repetition of the pattern—in traditional
approaches—is to copy some program skeleton and edit it manually.
The problem of course is not the editing, which is easy enough using modern
programming tools, but the fact that the borderline between the common pattern
and the adaptation disappears the moment we edit the skeleton. From then on,
198 BON STANDARD ACTIVITIES

the resulting code must be maintained as a separate whole, and we can no longer
benefit from the fact that most of it just represents standard behavior.
However, with inheritance we can keep the pattern intact, and still be able to
do all the required incremental modifications. A reusable object-oriented design
capturing a higher-level pattern will in most cases consist of several classes
working as a unit to achieve a certain type of tailorable behavior. Such a design
is called a framework, and contains a deferred class for each major component
viewed by the user, who is then supposed to tune the behavior by defining the
deferred operations in descendant classes.
Often a library of effective classes is also supplied for use when default
behavior is enough for some of the components. There are essentially two types
of frameworks each representing a different degree of maturity of the design.
The types may be called white-box and black-box frameworks, and are discussed
in a Smalltalk context by Johnson and Foote in an article proposing a set of
design rules, many of which agree with the BON approach [Johnson 1988].
In a white-box framework, tailoring is done by replacing chosen operations by
any new behavior seen fit. This gives fine grain control but often makes it
difficult to modify the default behavior without detailed knowledge of the
internal design of the framework classes.
In a black-box framework the user may only supply new behavior using
visible features defined in the framework. This gives less control, but makes the
user independent of the internal structure of the framework. Only the visible
interface of its components must be understood.

8.4 DEFINING CLASS FEATURES

Assignment of features never starts with a blank set of classes. In fact, if it does,
this is a sign of incorrect modeling. Since classes are only to be defined in terms
of behavior, unless we have at least a vague idea of some of the basic features of
a class (perhaps not yet written down), we do not know enough to decide
whether it should be a class. Classes and very basic operations go hand in hand.
However, our initial idea about the features will not necessarily be the best or
the only reasonable one. Therefore, trying to define in more detail the features
that lurk at the back of our mind—choosing names and signatures, writing down
their constraints—will give us more insight. We may then proceed to define
more features, either needed to fulfill system behavior or to make the abstraction
complete. The whole process becomes an iteration, shifting between finding
new classes, moving behavior between classes, and defining more and more
features and contracting details. The classification and clustering will also be a
natural part of this “multiple tactics” approach.
 DEFINING CLASS FEATURES 199

Initially, we should concentrate on what may be called the basic layer (cf. 8.1),
that is features that are essential for using the abstraction encapsulated by the
class, and which cannot be easily expressed in terms of other features. Later,
when the classes have stabilized, we may start adding convenience features.
Such features may in fact be essential for practical use, but are not necessary for
understanding the class and can therefore be postponed.

Naming features
The feature names are part of the language chosen for expressing an application.
In 8.1 we already stressed the importance of choosing appropriate names for our
abstractions, and we now proceed to give some rules for how to name various
types of behavior.
Since natural language is such an integrated part of our intellectual culture
carrying with it so many useful common connotations, we should only depart
from it when we have good reasons. One reason may be the greater precision
needed in formal languages (for example, when expressing contracting clauses),
another may be brevity. However, since the object-oriented model is particularly
good at capturing concepts usually expressed in natural language, the naming of
classes and operations generally does not need to depart very much from normal
description. This is a particularly important point for the analysis model (which
needs to be understandable by non-computer professionals), but also for the
subsequent design.
We should take advantage of this opportunity to stay close to natural language
when choosing our naming conventions. This does not mean our ideal should be
the style of the great epics, whose richness will be too much for our purposes,
but rather the terse dialect used in newspaper headlines. There, the phrases are
often carefully chosen to trigger very precise associations from the shortest
possible text.13
The style guidelines of Eiffel [Meyer 1992a] contain many good principles on
how syntactic conventions, spatial layout, and grammatical categories can be
used to balance the desired degrees of precision, brevity, and consistency. A
number of such principles are discussed in the following subsections.

Distinction from class names and clear word separation

There should be a clear syntactic difference between class names and feature
names, so that readers do not have to hesitate for a second about whether a name

13
The reader may not agree that conveying correct information is always the first concern in
headlines, but the potential is there.
200 BON STANDARD ACTIVITIES

refers to a type or to the use of a type. This distinction is extremely important,

since grave misunderstandings may result if the concept of class is confused with
the concept of class instance. The BON convention is to use only upper case
letters for type names, and only lower case for feature names.
There should be a way to incorporate the blanks of multiple word concepts
into names, so the individual words are still visible. Without this possibility,
expressive naming becomes very difficult. BON recommends using underscores
to signify blank spaces between words. Other common conventions use in-word
capitalizations, dashes, or combinations of these.
Of the three possibilities in figure 8.2, the second is very confusing in
arithmetic expressions and requires blanks to surround minus operators, while
the third deviates very much from natural language text.

easy_to_read_name: SOMETYPE −− BON standard

harder−to−read−name: SOMETYPE −− Not recommended
MuchHarderToReadName: SOMETYPE −− Not recommended

Figure 8.2 Possible naming syntax

Avoid abbreviation—but do not overqualify!

Since classes may collaborate with other classes representing entirely different
concepts, their names should be complete enough to avoid ambiguity. Feature
names always have the context of their class, and so can sometimes be made
shorter; on the other hand features are often more specialized and may therefore
require longer names. (We are talking about visible feature signatures here; local
variables with their limited scope may often use shorter names for simplicity
without risk of confusion.)
The overall goal is clarity and ease of reading. Therefore, we should make
names as long as needed to avoid ambiguity, but then always strive for short
names since excessive length per se is also one of the worst enemies of clarity.
One very important principle is never to include in a name facts that are
already clear from context. For example, if a class IBM includes a list of its
employees the feature could be named employees, but under no circumstances
ibm_employees. The IBM prefix is redundant and should not be used (unless of
course the class for some reason happens to also contain an employee list of
some competitor).
This type of error is rather common among beginners in the object-oriented
field who are used to global name qualification. We also strongly disagree with
the rules employed in many of Grady Booch’s earlier books, recommending
names like:
 DEFINING CLASS FEATURES 201

theSensor
aRoom
itsSemaphore
In our opinion, this is extremely bad style and can only decrease readability and
increase confusion when used routinely. (We are pleased to see that in the latest
edition describing the Booch method [Booch 1994] the number of names of the
above type has been reduced by an order of magnitude.)

Grammatical categories
Features that are commands represent actions taken, and should use verbs in
imperative form as names: enter, quit, remove. Queries, on the other hand,
represent access to information and should describe the result rather than the
process of fetching the information. Therefore, a noun possibly qualified by an
adjective should be used: speed, temperature, red_light. In contrast to class
names, feature names always have a type as local context and so may often omit
the noun of a qualified item without losing clarity: last, next, shortest.
Queries returning a boolean value representing a simple “yes” or “no” can use
two forms, either an interrogative phrase, is_full, is_empty, or just an adjective,
full, empty.

How many arguments?

If many input arguments are needed for a feature to fulfill its contract, this is
usually a sign that something is wrong in the system structure. A major principle
of object-oriented design is to reduce the need for heavy information flow
between local actors in the system (and do it without resorting to global
variables!). Therefore, the number of arguments of a class feature should
normally be fewer than for a corresponding subroutine in a traditional language.
Also, the calling syntax used by most object-oriented notations further
decreases this number by 1, since typical function calls like f (a, b) will instead
have the form a.f (b).
If many arguments still seem to be needed after several restructuring attempts,
a solution may be to create a new class as an aggregation of some of the
arguments, thus reducing the number of objects passed in one call. Also, never
use optional values as arguments, but instead define separate features to set
desired options.
As a result, the average number of feature arguments in an object-oriented
system can often be kept below one (the ISE Eiffel Base library of fundamental
data structures containing 1823 features reports an average of 0.4 arguments to a
feature [Meyer 1994a]).
202 BON STANDARD ACTIVITIES

Naming consistency
There is always an inherent conflict in any naming: should a name emphasize the
specific or the general properties of the concept it refers to? There is no simple
answer to this question; it all depends on what will usually give the most relevant
information when the name is presented to a user. Squeezing too much detail
into a name may be just as bad as having too little or useless information.
In UNIX environments, files are organized as directory hierarchies with local
naming within each directory. Software products are routinely distributed as
single directories (containing as many sublevels as needed) and then plugged
into a larger customer structure. When installing a specific product it may be
difficult to know what to do, since products are so heterogeneous. Some of them
may be directly executable, while others require several hours of detailed
configuration and compilation before they can be used.
The de facto standard that has evolved is to always supply a file named
README at the top level. This is an extreme example of a situation where
generality means everything and specificity will only come later. The name in
all its simplicity represents an enormously powerful semantic concept. Whether
the product consists of a few files that may be used to print an aphorism each
time a user logs off for lunch, or a fully fledged language environment whose
correct installation may be of the utmost importance to the organization, the
installer will know exactly what to do: just read the file and follow the
instructions!
In object-oriented systems, especially large ones, we are often faced with
similar problems. Scanning through the features of many classes to select the
appropriate ones for a particular task usually requires discarding many names,
while only picking out a few. If too much detail is part of the names (or worse,
only detail) it becomes difficult to find quickly the interesting categories.
Therefore, when naming features in libraries with many related types or in
frameworks, the higher-level category aspects are often far more important for
ease of use than detailed precision. This may require some unorthodox naming
at times.
Notable examples are the naming schemes in the data structure libraries of
Smalltalk-80 [Cook 1992] and Eiffel [Meyer 1990b, Meyer 1994a]. Some
standard names chosen in the latter environment capturing higher-level behavior
shared by many container classes are shown in figure 8.3.
It may seem strange, at first, to be faced with names like put and remove
instead of the usual push and pop when dealing with stacks, but the advantages
are usually grasped soon enough by programmers. In fact, once the power of
this principle has been understood, people will often start promoting it in many
contexts.
 SELECTING AND DESCRIBING OBJECT SCENARIOS 203

Name Behavior
item element access
count number of significant items in the structure
has membership test for given item
put insert or replace item
force like put but will always succeed if possible;
may for example resize structure if full
remove remove item
wipe_out remove all items
empty test for emptiness
full test for lack of space

Figure 8.3 Some standard feature names for container classes

8.5 SELECTING AND DESCRIBING OBJECT SCENARIOS

The amount of work needed to find representative object scenarios depends very
much on the type of system to be developed. In new situations, where there is no
previous manual task or users that can be studied, a more elaborate scenario
analysis may be required. The essence of such an analysis is to start from
existing technical and organizational conditions and try to envisage in detail how
a supporting system could be used.
First, a number of probable situations are selected and described as concretely
as possible in terms of a series of activities, user tasks and goals, input/output
requirements, and presentation formats. One or more of these descriptions are
then elaborated with respect to roles, tasks, and needs for computer support. A
systematic procedure described in [Gould 1988, Marmolin 1993] is shown in
figure 8.4.
Scenario analysis is chiefly used as a general help to generate or test new
ideas, particularly in new and unfamiliar situations. Therefore, it is very
important to choose representative scenarios that really show how the users
interact with the system. An advantage of this approach is that it gives a
complete picture of the situations chosen, while the disadvantage of course is
that it will be based on just a few, possibly irrelevant examples. An excerpt from
a scenario analysis is shown below.

The conditions are the following: a multi-media system providing

information about available traveling destinations is to be developed.
This system (called the travel guide) will be available for customers at
travel agencies, so they may browse among alternatives and make a
first selection while waiting to be serviced at the front desk.
204 BON STANDARD ACTIVITIES

• Determine technical and organizational conditions for the scenarios

• Identify tasks to be supported and possible system functions
• Describe concretely and in detail three imagined situations of system use
– one containing common simple tasks
– one containing common difficult tasks
– one containing rare but critical and difficult tasks
• For each situation
– describe the scene
– describe the actors
– describe who does what
– describe important events in the external world
• Describe concretely and in detail how the users might use the system
– describe the various working operations
– describe user roles, goals, tasks, actions, how these are controlled, etc.
– describe for each operation the system support, function, commands, etc.
– describe the course of events and system support using pencil and paper
simulation
• Analyze the scenarios and adapt the system support to derived requirements
– analyze work flows, information flows, and interaction
– analyze tasks and form subtasks (make a task model)
– analyze system usage problems and define support requirements
• Repeat with new situations and new tasks, and adapt the system continuously
to new requirements

Figure 8.4 Scenario analysis: a systematic approach

…On a cold day in January Mr P decides to surprise his wife on her

birthday with a week of unplanned vacation in a warmer climate. He
enters the local travel agency and walks up to the travel guide. The
message “Where would you like to go? (press here)” flashes on the
screen with a picture of a finger pointing to the text.
Mr P puts his finger on the message and a menu labeled “Select
by…” pops up. Mr P picks the alternative “Weather” and a panel
showing various weather conditions is displayed. Mr P then realizes
that besides romance the financial side is also a major factor, so he
presses “Cost” and a new panel consisting of several price groups
emerges. He chooses 2 weeks and 2 persons for < $2000 and then
returns to the weather panel. There he selects 95% probability of
sunshine and > 20 °C day-temperature.
He then tries his combination by pressing the “OK” button. After a
few seconds of background music the message “Sorry, nothing
available” appears on the screen. Mr P then tries 75% sunshine and >
18 °C, and the system answers with a list of possible travel destinations
 SELECTING AND DESCRIBING OBJECT SCENARIOS 205

and the text “For more information, select destination”. He presses

“Mallorca” and picks the first alternative from the resulting menu
“Major sights, Hotels, Recreations”. A video recording with samples
of the island’s cultural attractions appears on the screen…

Roles of scenarios
Scenarios can be of very different complexity. A scene like the one just
described illustrates how a complete problem, which we may called a user task,
is solved. Such scenarios can be broken down into more primitive pieces called
user actions (selecting weather conditions, submitting a choice).
Some scenarios representing major user tasks will often be worked out already
during task 1 in the BON process, since these can be of great help in guiding the
initial static modeling and convey a better understanding of what general view of
the problem domain to adopt. If the user problems to be solved are reasonably
understood at a bird’s eye level, the initial user task scenarios will hopefully be
realistic even if some of their details may change. So being as complete as
possible is of major importance here.
Later, when the basic class structure begins to stabilize, more scenarios
representing user actions will be added, enough to cover all types of primitive
user behavior at some level of abstraction. This can be done, since we now
understand the details better (usually during task 5 in BON).
However, the best partitioning may not always be two fixed levels—user tasks
and user actions. For simple systems there is perhaps no reason to separate the
two, while more complicated behavior may require more levels. Grouping
according to other criteria, such as user categories or subsystems, may also be
helpful.
Therefore, no explicit structuring is enforced by the BON charts—there is
only one type of scenario chart and scenario diagram. When a tailored structure
is needed, we simply recommend adding a category name in parentheses to the
name field of each scenario chart.

8.6 WORKING OUT CONTRACTING CONDITIONS

The BON approach focuses on the theory of software contracting, which we
consider the most important principle in existence today for constructing correct,
reliable software. No other technique proposed so far has the potential to turn
software development into the long-awaited discipline of engineering envisioned
by Douglas McIlroy in his contribution to the now famous NATO conference in
1968 [McIlroy 1976], and meet the challenge for an, as yet unrealized, software
components industry.
206 BON STANDARD ACTIVITIES

The theory, called design by contract [Meyer 1988a, Meyer 1992c], is actually
an elegant synthesis of the essential concepts in three major fields of computing
research: object-orientation [Nygaard 1970],14 abstract data types [Guttag 1977,
Gougen 1978], and the theory of program verification and systematic program
construction [Floyd 1967, Hoare 1969, Dijkstra 1976].

Design by contract
Most pieces of software text in today’s systems are only partially understood by
their developers. The central task of each subprogram may be clear enough (and
sometimes even documented), but as every experienced programmer knows it is
the unusual cases, or even the ones that are never supposed to occur, that present
the real problems. Since it is never exactly spelled out who is responsible for the
exceptional cases—the supplier of a subprogram, or its clients—important
prerequisites for various algorithms are often either checked in many places, or
not checked at all.
The general feeling of distrust resulting from this practice has lead to a
desperate style of blind checking known as defensive programming, which leads
to even less understandable software with more errors because of the complexity
introduced by the added redundant code. So the only solution is to create instead
an atmosphere of mutual trust by specifying precisely who is responsible for
what part of a complex system behavior. This is what design by contract is all
about.
The idea is to treat each subprogram as a subcontractor who undertakes to
deliver some service (specified by a postcondition), but only if certain
prerequisites are fulfilled (the precondition). The key opening the gate to future
trust and order—so that you can finally know that you are right when designing a
program instead of just guessing [Mills 1975]—is not as one may think the
postcondition (which specifies what the supplier will do), but instead the
precondition (which specifies what the supplier will not do).
To take an example, suppose you are to define a subprogram to calculate the
square root of a real number. If you expect this program to work under all
conditions, you are in fact mixing two completely different tasks into one:
• Finding and returning the square root of a non-negative number.

14
Simula, the first object-oriented language, not only introduced the remarkable concepts of
inheritance and dynamic binding already in 1967, but was also the direct inspiration of almost all
later work on abstract data types. It included the strong typing of Algol 60, but had generalized
the single stack model into a multiple stack machine, which enabled encapsulation of
autonomous objects.
 WORKING OUT CONTRACTING CONDITIONS 207

• Returning something reasonable when the input turns out to be negative

(assuming the output must be real).
For the first task we have a number of well-understood and efficient numerical
methods dating all the way back to Newton to choose from as supplier. For the
second task, we do not have a clue. Obviously the client has made a mistake,
and there is no way we can know what is a reasonable response.
Therefore, the only approach that makes any sense is to lift the second
problem off the shoulders of the supplier (who is not competent to handle it
anyway) and instead let it be the responsibility of the client not to ask impossible
questions. This may sometimes require explicit testing on the client side, but if it
does, there is no better place to do it. Usually the context will lead the client to
know without testing that the precondition is indeed fulfilled, something which is
never true for the supplier.

Contracting as a mutual benefit

The software contracting model has much in common with standard practices in
human society. For example, suppose you are in Stockholm and must deliver an
important package to an address at the other end of the city.15 Then you may
either deliver the package yourself, or engage a mail carrier to do it for you. If
you choose the latter alternative and employ the services of “Green Delivery”
(Stockholm’s bicycle courier), the standard agreement between you and the
courier looks like the one shown in figure 8.5. When two parties agree on
something in detail, the resulting contract protects both sides:
• It protects the client by specifying how much must be done.
• It protects the supplier by specifying how little is acceptable.
The obligations of one party become the benefits of the other. As an aside

Party Obligations Benefits

Client Provide package of maximum Get package delivered to recipient
weight 35 kg, maximum dimensions within central city limits in 30 minutes
parcel: 50 × 50 × 50 cm, or less without having to worry about
document: 60 × 80 cm. Pay 100 SEK. bad weather or traffic jams.
Supplier Deliver package to recipient No need to deal with deliveries
in 30 minutes or less, regardless too big, too heavy, or not prepaid.
of traffic and weather conditions.

Figure 8.5 A contract

15
The example is a slight modification of the one used in [Meyer 1992c].
208 BON STANDARD ACTIVITIES

(provided the contract covers everything) each obligation will also bring an
additional benefit: if the condition says you must do X, then X is all you need to
do. This may be called the No Hidden Clauses rule: sticking to the minimum
requirements of the contract is always safe for each party.
Regardless of the No Hidden Clauses principle there are usually external laws
and regulations whose purpose it is to prevent unfair contract clauses. For
example, if your package happens to contain a famous oil painting by Anders
Zorn the courier service is not permitted to drop it in the nearest garbage
container simply because it violates the precondition by measuring 80 by 90
centimeters.
Such external regulations, which are part of the general context in which the
contractors work, correspond to the class invariants of software contracts.

Laws of subcontracting
Polymorphism with dynamic binding is the main key to software flexibility. It
has the power to remove most of the discrete case analysis so error prone and
vulnerable to future changes—yet so abundant in traditional software. However,
flexibility is meaningless unless the resulting software is correct, and
polymorphism can be very dangerous in this respect.
Unless we are very careful when redefining an inherited operation, we may
easily end up with a system where only some of the implementations that may be
dynamically invoked will actually produce the expected result. What is there to
prevent a redefined area function from returning, in some cases, the diameter
instead? Without clear semantic rules, nothing but fuzzy naming conventions
and the folklore of software engineering.
The problem is more subtle than it may appear at first sight, because even if
every descendant class has a fully visible and correct specification of its
behavior, chaos may still ensue. For example, if we need to compute the area of
a list of geometric figures referred to by an entity of type LIST [FIGURE], all we
can look at as client is the specification of the area operation as it appears in
class FIGURE. During execution many different specialized versions of area
may be called dynamically, but we cannot check their corresponding
specifications when writing the list traversing code, if for no other reason than
because some of the corresponding classes may not yet exist!
Therefore, we must have strict rules that guarantee that any future descendant
class (whose operations may be invoked on our behalf whether we like it or not)
must fulfill the promises that were given by its ancestors. This leads directly to
the laws of subcontracting:
• A descendant class must fulfill the class invariant of the ancestor.
 WORKING OUT CONTRACTING CONDITIONS 209

• A descendant class may never weaken the postcondition of a redefined

feature (since this would mean delivering less than specified by the
ancestor).
• A descendant class may never strengthen the precondition of a redefined
feature (since this would mean imposing restrictions on the client not
specified by the ancestor).
Note that nothing prevents a descendant class from strengthening postconditions
(doing even better than promised) or weakening preconditions (imposing even
fewer restrictions).
Note also that the above rules must be obeyed for every ancestor in the case of
multiple inheritance, and will therefore prevent the combination of incompatible
abstractions. This is extremely important for building the complicated
inheritance lattices needed by, for example, the general data structure libraries of
strongly typed language environments.

Where to put consistency checking

Design by contract offers an alternative to the blind checking of defensive
programming by specifying a clear division of responsibility between client and
supplier regarding the checking of various details of system consistency:
• The client is responsible for guaranteeing the precondition when calling.
• The supplier is responsible for guaranteeing the postcondition when
returning.
Therefore, we have to choose where to put our tests for consistency. Either a
condition is part of the precondition and must be ensured by the client, or it is
removed from the precondition and must then be handled by the supplier.
Which alternative to choose must be decided case by case based on many
factors, but the guiding star should always be the resulting simplicity and
understandability of the system architecture. A rule of thumb is that if most
clients need their own special treatment depending on some condition, it is better
to put it in the precondition, while if the behavior alternatives are more or less
standard for most clients, it may be simpler for the supplier to deal with them.

Classes as specification elements

As was argued in chapter 2, we should not strive to make the specification of an
object-oriented system independent of its design, since this would defeat its
purpose. Maintaining two entirely different descriptions (one for the system
specification and one for its implementation) does not make sense, because
210 BON STANDARD ACTIVITIES

specifications of large systems become large no matter what language we

choose. Therefore, separating the two worlds will only give us inconsistency
problems and more difficult maintenance for rapidly evolving systems.
So the specification elements used must in the end be translatable into object-
oriented expressions, involving feature calls on objects of the classes which are
part of the system design. These classes are the only abstractions that can
capture the complex behavior of the system through simple notation (provided
the design is good) as using an independent system specification would
necessitate starting from scratch. Instead, specification and implementation must
share the same abstraction base, since the executable code should only be the
innermost part of a layered design of abstractions.
Partial recursive specification will not tell us the whole truth about a system,
but it will tell us nothing but the truth, and it has the flexibility and
incrementality we seek. Any other approach breaks the seamlessness and is in
our view doomed to fail as a road to mastering the industrial development of
large and complex systems.16 Complete specification of large industrial systems
will probably never become feasible anyway, if only because of the constant
changes involved.

Run-time monitoring of assertions

An important side effect of the recursive specification approach described is that
assertions may be translated into procedural code and monitored during
execution. Basic boolean expressions map directly to programming language
primitives, while the first-order quantifications of BON assertions may be
implemented as functions. We take the simple class PERSON in figure 8.6 as an
example. Its invariant expresses that if you are a person, then each of your
children has you for one of its parents. (@ is the BON graphical symbol for
current object.)
The corresponding Eiffel code, which may be generated by a case tool, is
shown in figure 8.7. The control structure of Eiffel, from−until−loop−end,
should be self-explanatory and the LINKED_LIST class is a cursor structure,
which may be traversed by invoking features start and successions of forth until
the state after is reached (meaning the cursor is positioned just after the list).
The standard feature item returns the list element at the cursor position and
Result is a predefined local variable containing a function’s return value.
(Entities of type BOOLEAN are initialized to false by default, so Result can be
used directly in the second function of figure 8.7.)

16
We are talking about general systems development here; certain critical or highly specialized
software may of course still at times benefit from other techniques.
 WORKING OUT CONTRACTING CONDITIONS 211

PERSON

name, address: VALUE

children, parents: LIST [PERSON]
Invariant
∀ c ∈ children • (∃ p ∈ c.parents • p = @)

Figure 8.6 Consistency requirement: your children are really yours

class PERSON
feature
name, address: VALUE
children, parents: LINKED_LIST [PERSON]
generated_assertion_1: BOOLEAN is
do
from
children.start; Result := true
until
children.after
loop
Result := Result and generated_subassertion_1 (children.item)
children.forth
end
end
generated_subassertion_1 (c: PERSON): BOOLEAN is
do
from
c.parents.start
until
c.parents.after
loop
Result := Result or (c.parent.item = Current)
c.parents.forth
end
end
invariant
generated_assertion_1
end

Figure 8.7 Class with generated assertion routines

212 BON STANDARD ACTIVITIES

When specification elements are implemented using procedural code, we must

be extremely careful not to introduce any side effects, since this may change the
semantics of the system when the assertions are monitored. Moreover, since any
feature returning interesting information about the system state is a potential
specification element, the rule of side-effect-free functions acquires an even
greater importance.

Systematic exception handling

The contract theory also enables a very powerful exception handling mechanism
to be applied during system execution. Since routines are not just small pieces of
reusable software text, but precisely specified individual implementations, it is
possible to introduce a notion of failure. Failure occurs when an execution of a
routine is for some reason unable to fulfill its part of the contract.
An exception in a routine can be triggered in one of three ways: a supplier
returns failure, an assertion is violated, or a signal is received from the
surrounding hardware/operating system. (Note that assertion violation includes
violation of the postcondition just before returning, as well as violation of a
supplier precondition just before calling, since the latter is the client’s
responsibility.)
Exceptions may be processed by handlers, which will restore the class
invariant for the current object and then either admit failure or else execute the
whole operation again (after, for example, setting some flags). Admitting failure
means triggering, in turn, a failure exception in the caller.
Since this book focuses on analysis and design and the details of exception
handling are closely linked with the programming environment, we will not go
further. The interested reader is referred to [Meyer 1992c, Meyer 1992a].
Finally, the violation of an assertion means that some implementation did not
behave according to the specification. It is important to understand that this is a
sign of a software (or possibly hardware) error. Things that may be expected to
happen, no matter how seldom, must be part of the system specification and
should therefore be handled in the main logic of the class features.

8.7 ASSESSING REUSE

Since BON is directly targeted at promoting reuse, we have chosen to view
various work related to it as a standard activity when completing the tasks
outlined in the BON process. This entails both the assessment of when existing
software may be reused for parts of the product under development, and
decisions regarding how much effort should be invested in future reusability of
the new software produced.
 ASSESSING REUSE 213

A changed attitude
An extremely important driving force in making large-scale reuse come true in
software projects is the general attitude of the developers and (perhaps even
more importantly) managers involved. A designer must actively seek reuse as
part of the routine work. Normally, as much time should be spent on reading old
code, looking at design descriptions, browsing through component indexes,
reading related literature, as is spent creating new code or designs.
Just as building an extensive network of human contacts has long been a main
strategy in many professions, we must learn how to make the most of available
channels to reduce the amount of new software development. In fact, for any
complex functionality needed there should be only two alternatives: either reuse
(and possibly adapt) existing software or, if this is not possible or feasible,
develop something which may be reused in the future. One-shot developments
should be restricted to the easy parts within a system.

Make the complex parts reusable

It is clear that large parts of the systems developed to support human activities
will always be very special and not worth the effort to be made highly reusable.
This is natural, since human endeavors are constantly changing and many ideas
(perhaps most) will just be abandoned before long in favor of others. But behind
every complex problem—no matter how special—there are always aspects of
more general patterns.
The trick is to detect these patterns and capture them as class abstractions.
With this strategy, every complex special problem will instead be viewed as a
less complex (non-reusable) combination of general (reusable) patterns, where
the latter will take care of the complexity. There will still be a substantial
amount of non-reusable code, but the important point is that this code is
comparatively straightforward, and we can afford to throw most of it away later.
In fact, for general systems, it is an important point of minimizing
development effort that significant amounts of trivial code should be thrown
away between releases of rapidly evolving products! This is the essence of
programming-in-the-large [DeRemer 1976], where the non-reusable code is
viewed as temporary glue to configure the high-quality standard units used to
build systems in a certain domain.
The alternative would be to develop too much in advance, without really
knowing whether it will ever be used in the future. Potentially reusable but
never used components developed at high cost represent as much waste of
money as the old tradition of doing the same things over again. However,
solving a complex problem without making at least some classes also applicable
in more general contexts should be looked upon by any developer as a failure.
214 BON STANDARD ACTIVITIES

Not only reuse of classes

The object-oriented technique is a very powerful vehicle for capturing and
reusing concepts. This means that not only existing classes, but any kind of
earlier effort in the problem area may have the potential for reuse, because reuse
is anything that will reduce the intellectual effort and time required to produce a
software system.
The traditional scavenging of old software to try and find useful pieces of code
and / or designs still applies. In fact, it may sometimes be more fruitful to go
through well-designed traditional programs solving related problems and re-
engineer the solutions found into class abstraction than to start from a mediocre
class library. Good designs in non-object-oriented languages tend to use (albeit
not always consciously) the ideas of encapsulation and virtual machines brought
forward by Dijkstra [Dijkstra 1972] and Parnas [Parnas 1972, Parnas 1979], and
are usually straightforward to carry over into the object-oriented world.
However, in the case of more obscure designs the benefits gained can easily
be offset by the effort needed to understand and rework. As expressed in
[Krueger 1992]:

For a software reuse technique to be effective, it must reduce the

cognitive distance between the initial concept of a system and its final
executable implementation.

Cognitive distance is defined as “the amount of intellectual effort that must be

expended by software developers to take a software system from one stage of
development to another.”

Encapsulation of non-object-oriented software

An especially important reuse technique, which may very well turn out to be a
major strategy in many application areas for gradual migration from old
technologies, is object-oriented encapsulation of existing components written in
traditional languages. This can be done in several ways, for example:
• by external subroutine calls from class features
• by calling class features from traditional software
• by invocation of external programs from class features
The first alternative is particularly suitable for well-documented and stable
external subroutine libraries. (A class that encapsulates a set of external routines
is often called a class wrapper.) It is often quite easy to make a first
encapsulation, which directly mirrors the conceptual model represented by the
 ASSESSING REUSE 215

library routines. As a side effect, the quality of the old software may often be
improved by including contracting elements. Depending on how well the
encapsulated concepts fit the problem area, these may then be extended by more
or less elaborate object-oriented models, using the primitive features as building
blocks. The second, inverted, approach may be advantageous when it is
desirable to let an old application keep central control, but facilitate the addition
of new functionality. Both approaches permit successive replacement of old
parts by new object-oriented ones.
The third alternative is very important in connecting already existing local
systems in networks using client−server techniques. Object-oriented languages
with good interfaces to other languages can act as a very efficient gluing
mechanism between heterogeneous components. With relatively little effort, an
object-oriented system can tie together the complex behavior of a large number
of existing products, and present a user interface that effectively removes the
conceptual walls between them. What used to be a world of completely separate
products, often with baroque historical interfaces, is then turned into one uniform
system with automatic information exchange behind the scenes.

Assess impact of future changes

An important principle in designing reusable software is to constantly make
thought experiments as to what would happen if certain requirements currently
taken for granted were to change in the future. This gives an idea of the
robustness of the product.
Possible areas to investigate include the information structure, the user
interface, the number and categories of users, speed requirements, persistency
requirements, security, platforms. Of course we cannot afford to take every far-
fetched possibility into consideration, but sometimes small preparations to
facilitate future changes in certain directions can be made at low cost if they are
already built in during the initial design.

8.8 INDEXING AND DOCUMENTING

Another important aspect of software reuse is the problem of selecting the proper
components from the base of available software. In this connection there are two
obvious truths (also pointed out in [Krueger 1992]), which may effectively
prevent reuse in practice:
• To select an artifact for reuse, you must know what it does.
• To reuse an artifact effectively, you must be able to find it faster than you
could build something of comparable quality.
216 BON STANDARD ACTIVITIES

The first truism shows the importance of abstract documentation (above the
source code level), which is of course equally important for maintenance. In
fact, maintenance is only a special case of reuse because whenever a change in
system behavior is contemplated, one must always consider what is most cost
effective: to modify the existing system or redesign it from scratch.
The second truism shows the need of various classification indexes, ranging
from very simple to very refined depending on the size and complexity of the
available software base. Because of their importance, we consider indexing and
abstract documentation a standard activity in BON.

Abstract documentation
Useful documentation needs to be abstract, which means that many details of
what is described are omitted. This is the only way to communicate complicated
ideas between people, since the human capacity of keeping many things in mind
simultaneously is extremely limited [Miller 1956, Miller 1975]. However, the
difference between being abstract as opposed to just incomplete is that the details
to be skipped are chosen with care. The idea is to find concepts that can be used
to group various similarities, so that details do not need to be individually
enumerated. Moreover, the concepts chosen must be easy to understand in
relation to the expected backgrounds of readers, and have enough precision for
the type of description.
At the heart of the documentation of an object-oriented software system are
the class interfaces. They are the final components that define the system
behavior and whose correctness is a prerequisite for usable software. Therefore,
the classes need abstract documentation to be understandable, and, what is more,
the documentation needs to be precise. The precision aspect becomes gradually
more important as more software is reused and systems grow larger. Without it,
we cannot see any possibility of a successful large-scale software components
industry.
Software contracting with strong typing has the potential to attain the required
precision, and do it without breaking the all-important seamlessness of software
development. The core of the documentation of a class should therefore be the
software contract between client and supplier. The specification of the contract
should be an integral part of each class, and not something maintained on the
side.
Besides class documentation, we also need documentation at the cluster level
(for those clusters representing subsystems or frameworks) and of the general
system architecture at yet higher levels for large systems. Such additional
technical documentation, as well as the various kinds of user manuals needed,
may be attached to the respective cluster levels in the system static architecture
by a case tool.
 INDEXING AND DOCUMENTING 217

Class interface documentation

Two important tools are needed to produce useful class interface documentation
from class texts: a class abstractor and a class flattener. The first one removes
all implementation details leaving only the signature, the pre- and postconditions
and descriptive comments for each feature, and the class invariant. The second
eliminates the inheritance structure of a class by including all inherited features
(with possible renaming and redefinition taken into account) so that its full
interface becomes visible. Combining the two produces a complete abstract
interface for a class which is guaranteed to be correct.
Without the class flattener a user may have to scan through a large number of
ancestors and compile the interface mentally. Doing this in a multiple
inheritance environment with 10−20 ancestors, with some renaming and
redefinition needed, and still hoping to understand everything, would seem
optimistic indeed.
Among the object-oriented languages, only Eiffel [Meyer 1992a] and Sather
[Omohundro 1991] include assertions as fully fledged language features, which
in combination with the strong typing makes it straightforward to implement the
abstraction tools mentioned above. To some extent, assertions can be emulated
by preprocessor macros in C++ [LeJacq 1991], but full run-time monitoring—
including enforcement of the laws of subcontracting—requires direct language
support.

Documenting frameworks
Object-oriented frameworks, often encapsulating fairly complex designs that can
be tailored by users, are generally quite difficult to document in a way that is
both precise and easy to comprehend. They often contain many layers of
abstraction and a large number of details that need to be understood at times by
some users, while a typical user may only need to know about a small subset.
The problem is how to structure the documentation in such a way that an
inexperienced user can get the required information without being disturbed by
too much detail, yet be able to go deeper whenever the need arises.
The possibility of improving current practice of framework documentation by
using an idea developed in another domain by the British architect Christopher
Alexander is currently being investigated [Johnson 1992]. Alexander used the
term pattern for a description of how to solve a particular type of problem. He
then designed and carefully structured a set of such patterns (253 design precepts
connected to urban development) into a “pattern language” that would capture
enough professional architectural knowledge in a document to enable laypersons
to use it for designing their own homes and communities [Alexander 1977].
218 BON STANDARD ACTIVITIES

Instead of describing all the facilities and concepts of a framework as a map of

its static structure, the pattern language technique focuses on standard ways of
using the framework to solve typical user problems. By choosing patterns that
cover the most important aspects of what users may want to do, a user can
combine these primitive patterns to solve more complex problems.
Within each pattern, the details believed to be needed most often by users are
described first, and special usage only later. The reader may choose how far to
go with each pattern before moving on to the next. When specialized behavior
concerning some system aspect is later needed, the corresponding pattern can be
read again in more detail.

Indexing guidelines
Each index entry in an indexing clause consists of a keyword serving as an
index, and a list of attached words serving as index values. The choice of
indexes and values is left open for a given library or installation to define its own
conventions. Some guidelines for such conventions taken from [Meyer 1994a]
are listed below:
• Keep indexing clauses short (3 to 8 entries is typical). May change in the
future as the needs of cataloging and retrieval tools are better understood.
• Avoid repeating information that may be automatically extracted from the
rest of the class.
• Use a set of standardized indexes for properties that apply to many classes
(such as choice of representation).
• For values, define a set of standardized possibilities for common cases.
• Include positive information only. For example, the indexing clause of a
class that does not have any representation should not contain the entry
representation: N/A, but simply no entry with keyword representation. A
reasonable query language will still make it possible to use query pairs like
<representation, NONE>.
The following are some examples of standard index terms and typical values for
a general data structure library. Index term description gives a short overview of
the abstraction represented by the class. Index term names records alternative
names for a structure. The abstraction implemented by a class LIST, for
example, may also be known as a “sequence”. Index term access records the
mode of access to the data structures. Standard values include one or more of
the following:
• fixed (only one element accessible at a time, as in a stack or queue)
 INDEXING AND DOCUMENTING 219

• fifo (first-in-first-out policy)

• lifo (last-in-first-out)
• index (access by an integer index)
• key (access by a non-integer key)
• cursor (access through a client-controlled internal cursor)
• membership (availability of membership test)
• min, max (availability of operations for accessing minimum or maximum)
Index term size indicates size limitation. Common values include:
• fixed (size fixed at creation, cannot be changed)
• resizable (initial size fixed; explicitly resizable possibly at some cost)
Index term representation indicates choice of representation, for example:
• array (contiguous, direct access)
• linked (linked structure)

Style of comments
Every class feature should have a header comment attached as part of the
abstract feature specification. The header comment should be informative, clear,
and concise. Brevity is in fact an extremely important quality for comments in
general (as well as for most text). To be effective, comments need to have a very
high signal-to-noise ratio, otherwise they lead to inattentive reading and thus loss
of information. Quoting one of the main rules in Strunk and White’s
masterpiece The Elements of Style [Strunk 1979]: “Omit needless words!”
The following rules taken from [Meyer 1992a] should help achieve brevity
without losing vital information. Avoid repeating information which is obvious
from the signature or from contract details already specified by the pre- and
postconditions. For example, the header comment should not have the form
tangent: LINE
− − Tangent to circle c through point p
– c: CIRCLE
– p: POINT
but simply
−− Tangent to c through p
220 BON STANDARD ACTIVITIES

Avoid noise words and phrases such as “Return the…” explaining the purpose
of queries, or “This routine updates…” explaining a state changing command.
Instead of
−− This routine updates the display according to the user’s last choice.
use
−− Update display according to last user choice.
Header comments should begin with a capital letter like a regular sentence, and
not contain word abbreviations. They should have the following syntactic form:
• For commands: imperative sentence ending with a period, as in the last
example.
• For non-boolean functions: nominal phrase, such as “Tangent…” above.
• For boolean functions: interrogation phrase ending with question mark, as
in “Is current node a leaf?”
Use consistent language; if one comment refers to “Length of string…” the next
should not say “Update string width…”.

8.9 EVOLVING THE SYSTEM ARCHITECTURE

System design proceeds by the gradual refinement of an initially incomplete,
abstract structure. This is the essence of object-oriented modeling: not to decide
too much too soon. However, refinement means not only working out more
detail and filling in missing parts—going from the abstract to the concrete—but
also changing parts of the model as our understanding increases. The system
will usually have aspects whose modeling is less obvious, and the best way to
gain more insight is then often to pick one possible approach and elaborate it in
more detail.
So the modeling can be viewed as a succession of iterations between
reasonably “consistent” system stages, each hopefully representing a better view
than the previous one, until we reach a point of acceptance. In this concluding
section on the last of the BON standard activities, we will discuss some general
principles for evolving an initial system model.

Avoid premature abstraction

Abstraction is strong medicine, which means that the view of the world
represented by the set of abstractions available to us has a very profound
influence on the way we think. Therefore, just as good abstractions will act as
 EVOLVING THE SYSTEM ARCHITECTURE 221

experienced local guides in their respective problem domains, automatically

showing us the best views, misleading or ad hoc abstractions can be very
dangerous.
Instead of trying to obtain as many layers of abstraction as possible, we should
search for the powerful abstractions. This may sound trivial (or void of content,
since we cannot objectively measure what is powerful) but it is still a crucial
point in modeling.
Each time we introduce a new abstraction by encapsulating some behavior (or
some data, which the reader should at this point be convinced is the same thing),
we protect its users from changes behind the scenes. But we do it at a cost: from
now on this new concept needs to be part of our vocabulary and will influence
our decisions on how to structure the software (in this, as well as in future
versions of the system). If the concept is a strong abstraction, which means it
can be used to solve many interesting variations of some problem and has
intuitively clear semantics, then this cost is more than offset by the benefits it
brings. In fact, having to remember and take new strong concepts into account
in future situations will be a strength rather than a burden, since they may very
quickly indicate solutions that are otherwise not easy to see. But for mediocre
concepts the payoff is not at all clear, and for weak concepts the net effect of
encapsulation is often negative.17
It is therefore often better to initially model a problem directly in terms of
somewhat lower-level strong abstractions than to prematurely introduce a set of
intermediate-level weak abstractions. The reason is that if we start thinking in
terms of the weak abstractions and reuse them heavily, it will distort our view of
the system and make it more difficult to find the higher-level strong abstractions
later.
If, on the other hand, we continue to employ lower-level strong abstractions as
long as we only see weak abstractions at the higher level, common usage
patterns of the small building blocks will eventually emerge, and indicate how
higher-level strong abstractions can be built to encapsulate them. Prematurely
introduced weak abstractions may prevent us from ever seeing these patterns.

Strong design means strong dependencies

System design cannot be, and should not try to be, equally flexible in all
directions. The characteristic of good design is precisely that it is expressed in

17
Note that a strong concept does not have to be potentially reusable in other systems. Since sys-
tem evolution (hopefully) means heavy reuse of the components of the previous system version,
a very special concept can still be a strong abstraction in the context of the system in which it
was conceived.
222 BON STANDARD ACTIVITIES

terms of a small number of well-defined strong concepts capturing the essence of

some problem area. Consistent modeling in terms of these concepts (and
consequently strong dependence on them) is what enables future requirement
changes to be met by simple incremental adaptation instead of extensive
redesign.
Trying to guard against all types of changes, on the other hand, will only lead
to weak design or no design at all, since the only order equally easy to modify in
all directions is random order. Therefore, minimizing the number of client
dependencies for each class by systematically encapsulating all collaborative
behavior needed in just a few supplier classes, as for example advocated by the
Law of Demeter [Lieberherr 1989], represents a driving force towards less
consistent global design. Instead of gradually leading to a system in which many
classes trust a smaller number of strong abstractions enough to call them directly,
the evolution tends to favor many weak abstractions each calling only a small
number of other weak abstractions.
Superficially, such a system may seem easier to maintain, since from a local
standpoint each class appears less vulnerable to changes in other classes, but this
is an illusion. Taking a look at the bigger picture will reveal that most classes
contain too many features (often representing flexibility in directions not needed)
which duplicate similar behavior and mix several abstractions. There will be
much more code to maintain, and the general building blocks, instead of being
small, clear, and orthogonal, tend to be large, fuzzy, and overlapping.

Specialization and generalization

There are an infinite number of strong concepts that may be potentially useful in
systems development. However, only a very small fraction of them will ever be
needed. Therefore—since we are not equipped with divine foresight—we should
resist the temptation to develop too much too soon, and instead let experience
guide the direction of our evolving conceptual base. A typical technique,
directly supported by object-oriented abstraction, is successive specialization and
generalization of the concepts found to be most useful in practice.
The corresponding classes will often start by offering an initial set of basic
behavior, which may be adapted in descendant classes (specialization). This is
the least painful process, since it requires no change at all to the old software.
After some time, however, it often turns out that some descendant class C only
needs part of the behavior inherited from B. The solution is then to isolate this
behavior in a separate class A, which now becomes an ancestor of both B′
(containing what is left of B after the extraction) and C.
This is an example of generalization capturing the common parts of two
classes. It is more of a corrective action than specialization, since some source
 EVOLVING THE SYSTEM ARCHITECTURE 223

code must be changed. However, if the external interface of B can be kept intact
(which is often the case) only minimal changes are needed, since old client code
using B will continue to work as before.

Deferred classes as program templates

Another important empirical input regarding what abstractions to choose is the
discovery of common patterns of behavior, not only in connection with
specialization / generalization as above, but also between previously unrelated
classes. Behavior patterns that are closely intertwined inside highly variable
behavior are often not possible to capture by traditional subroutines. It is
therefore common practice in many environments to use templates or
skeletons—pieces of software text representing the fixed parts of some
behavior—and then add the variable parts through a copy and edit procedure.
In object-oriented environments such templates can often be represented by
deferred classes with the variable parts filled in by descendant classes defining
implementations for the deferred features. This is a great advantage, since the
pattern can be maintained in one place and only the variable parts need to be
addressed individually.

Reuse through discovery rather than invention

Software design of well-understood small problems may often profit from the
technique of stepwise refinement [Dijkstra 1972, Wirth 1971], going from the
abstract to the concrete (provided an eye is kept on what abstractions are already
available, so we do not invent them again in slightly different form). But a major
part of large systems design, and particularly of system evolution, is to proceed
in the reverse direction: from the concrete to the abstract.
There are two good reasons for this. First, we need the benefit of reusing
classes already developed and proven in practice, which necessarily has a flavor
of “bottom-up” design to it. Second, many common patterns will only emerge
after several concrete problems have been implemented (or designed) and the
resulting behavior compared. This is to say that a large part of the reusable
abstractions in a system (perhaps the majority) will, in the long run, have been
discovered rather than invented.
This “empirical reuse” has the great advantage of being based on reality
instead of foresight, so the corresponding abstractions do not risk being left
unused. Of course we should always try to go beyond what is immediately
needed and find more general concepts behind the common patterns observed,
but not to look too far, because reality may soon enough take a new and
unexpected direction.
224 BON STANDARD ACTIVITIES

Implementation as a road to abstraction

Implementation is an extremely important vehicle in the search for reusable
designs, because most abstractions are not born in their final form, but rather
evolve gradually. We will have to accept working with software in a constant
state of evolution, and this fact must not be disregarded by languages and
methods.
Therefore, the implementation of a class should not be viewed as an arbitrary
low-level solution with no abstractional value for reuse. On the contrary, if done
properly, it is a great guide to how similar behavior can be implemented. As
opposed to many abstract designs on the drawing board, it also often stood the
test of actual usage.
Therefore, separating specification from representation in implementation
languages, by providing one inheritance mechanism for subtyping and another
for code sharing, as advocated by some researchers, does not promote
reusability. The inheritance of concrete classes without code sharing means
missing one of the main points of object-oriented development: solving each
particular problem only once.
This is perhaps even more apparent when several classes are involved. A
white-box framework is defined in [Johnson 1988] as a reusable object-oriented
design where the reusing developer is granted access (through inheritance) to the
internal structure of a set of classes, while in a black-box framework only visible
features may be used (cf. 8.3).
Most frameworks need to start in white-box form to permit fine grain
adaptation of behavior by using also the private routines of the inherited classes.
As a heavily reused design matures, it is often possible to evolve a white-box
framework into black-box form, and achieve the advantages of stronger
encapsulation and clearer abstraction. However, this is not possible or
worthwhile for all frameworks.
Nevertheless, inheriting a good implementation means getting access to an
abstract interface. Although this interface is at a somewhat lower level and
somewhat less clearly specified than the corresponding external client interfaces
(because of its greater level of control), it is still a substantial abstraction
representing a good deal of accumulated knowledge of the problem at hand. As
noted in [Johnson 1988], denying the value of sharing representation means
denying the value of white-box frameworks, both as a means for reuse and as a
vehicle for discovering more powerful abstractions.

Make regular conceptual summaries

Systems development is no easy task, at least not if the ambition is to find good
and clean solutions. Often the situation will be perceived as chaotic, but this
 EVOLVING THE SYSTEM ARCHITECTURE 225

seems to be a necessary ingredient in all complex problem solving and should

not be feared. However, when all the central concepts are in doubt and the
general confusion among the designers seems greater than ever, this could be the
sign of two very different states:
• The design process could be very close to a point where the thick mist will
suddenly disperse to reveal a new-born, clean solution in all its splendor.
• The process could be very far from a solution, perhaps pursuing a dead-
end route.
The trouble is, there is no easy way to tell which one. Therefore, as a means of
speeding up the process in the first state, or discovering and escaping from the
second state, a comprehensive overview of concepts and problems agreed upon
should be performed in writing at regular intervals.
To assure a globally consistent view (albeit subjective), this should be done by
one or a few designers, each covering a relatively independent set of concepts.
When the rest of the group compare the view to their own, previously unknown
disagreements will surface. Writing down design thoughts for others to read is
an excellent means of testing their degree of maturity; things you cannot explain,
usually you do not fully understand. Much of the error detection and
corresponding improvements of designs often results from merely systematic
documentation.

Seek polymorphism
Polymorphism with dynamic binding is the acid test for object-orientedness. It
provides the unique capability of eliminating case analysis before applying an
operation to an object in a family of related types, where each member may need
a slightly different version of the operation. This permits the application of a
whole group of similar operations to a set of similar objects in just a single
feature call, letting the system worry about connecting each object to the correct
implementation at run-time.
Polymorphism is the natural thing to use when we are only interested in the
similar aspects of a set of operations (which is very often the case). The
similarity is captured by giving each version the same name, and then thinking in
terms of applying one operation to one or more objects instead of the actual set
that may be involved behind the scenes.
A common ancestor of the object types, specifying this name and
corresponding signature, is enough to achieve the desired effect. Pre- and
postconditions capturing the essence of the similarity addressed will assure the
correct semantics for each version, present and future, through the laws of
software subcontracting.
226 BON STANDARD ACTIVITIES

This makes it possible to apply an important rule for maximizing the

separation of concerns in a system, which has been named the single choice
principle by Bertrand Meyer:
Whenever a software system must support a set of alternatives, the
exhaustive list should be known by only one component in the system.
Following this principle as closely as possible leads to considerably smaller
systems which are easier to maintain and understand. We can only do this by
actively promoting polymorphism as early as possible in the development
process.

On library design
We conclude this chapter by touching briefly on some library design principles.
An object-oriented library is meant to be used by many developers, and to cover
some problem area (small or large) in a reasonably complete, yet flexible way.
This automatically implies higher demands on correctness, documentation,
robustness, ease of use, and efficiency, than for systems developed with a special
environment and user group in mind. Great as the potential benefit of good
libraries may be, equally much damage can be caused by a heavily reused low-
quality piece of software, so extra effort has to be invested.

Precondition checking
Since heavily reused libraries are (hopefully) of better quality than average
software and are extensively tested, one may jump to the false conclusion that
run-time monitoring of assertions is not of much value once a library has
stabilized. However, what is then forgotten is that a violated precondition is not
a sign of error in the supplying feature, but in the client.
Therefore, it is extremely profitable to specify formally as many of the
precondition clauses as possible in library routines, and leave the checking on,
since this is a very effective way of catching errors in the applications that use
them. As opposed to postconditions, the checking of preconditions tends to be
cheap in terms of memory and CPU power consumed, since many conditions can
be expressed as simple boolean expressions.

Consistency
As mentioned earlier in 8.4, consistency becomes all important for large libraries
regarding naming, concepts, user metaphors, and functionality. Here, similarity
of related operations by choosing the same name for them extends beyond
polymorphism. When searching through a large number of partially unknown
 EVOLVING THE SYSTEM ARCHITECTURE 227

abstractions for some desired functionality, we need a way to first do a quick

categorization in order to come up with a crude selection. Then we can afford to
look a little closer at the specifications of the candidates in this smaller set.
Defining a basic layer of operations to cover the functionality of a complicated
class, and then complementing these by convenience features, may also be
important in this context.

Inversion of control
The purpose of a library is to capture and reuse knowledge of how certain types
of problems can be solved. Most libraries act as passive but resilient toolboxes
containing good abstractions that the user can choose from, combine, and adapt
in desired directions (for example, data structures, graphical figures, GUI
widgets, mathematical functions). The idea is that only the user—who has the
big picture—knows enough to select and combine the right components for the
job.
However, for certain well-understood problems the outline of a whole
component, or in some cases a whole application, can be captured as reusable
knowledge in a library, then often called a framework. In frameworks, the top-
level control is often reversed, so that the user instead plays the passive part.
The basic behavior is then furnished by the framework, but can be adapted by
information passed from the user, or by the framework calling user-supplied
routines. Such frameworks can play the role of structured templates for solving
complex problems using standard methods.

Toolkits
Some frameworks addressing complex problems require many similar details to
be supplied by the user. When a large number of classes need to be defined,
even if each one of them is straightforward enough, the sheer quantity may soon
become unmanageable.
Therefore, one may need to go a step further and use a dedicated application
that will generate most of the information required by the framework from much
simpler input (either interactively or from stored files). Such applications may
be called object-oriented toolkits. Typical areas that may benefit from toolkit-
generated class text include GUI applications, finite state machine applications,
and object-oriented parsers. We conclude this chapter with a few examples.
In Glazier [Alexander 1987], Smalltalk classes and methods are generated
using the MVC model to create new types of windows from interactive user
specifications combining a set of primitive window elements. The user does not
need to know the rather complicated details of the Smalltalk windowing
framework supporting the MVC paradigm to use Glazier.
228 BON STANDARD ACTIVITIES

In EiffelBuild [Meyer 1994b], complete Eiffel applications can be generated

using the EiffelVision encapsulation of Motif and OpenLook and an extensible
set of standard command classes. The user builds a GUI window hierarchy
interactively by selection and placement of standard widgets, and then connects
mouse and keyboard events in respective widgets to corresponding commands
by graphical drag-and-drop. A user extensible set of standard commands are
available as icons on a palette. Commands can be made state dependent and the
corresponding state transition graph defined using a graphical FSM editor.
In both these cases, a large number of tedious and error-prone details can be
hidden from the user, who can still tailor the result by editing the resulting
classes. This is particularly important when the requirements change constantly,
as they do in rapid prototyping or research applications. The third example
problem area, parsing, is perhaps even more typical in this respect.
The Eiffel parsing library [Hucklesby 1989, Meyer 1994a] supports automatic
parsing of languages expressed in LL(1) grammar form. Each grammatical
construct becomes a separate class, which gives interesting possibilities for the
clean separation between the syntax and semantics of a language.
However, the resulting number of classes for a realistic language is much too
large to maintain by hand, particularly when the language itself is evolving and
each version may be described using several alternative grammars. The toolkit
PG [Grape 1992] automatically generates classes directly from grammars
expressed in a compact extended BNF notation.
Finally, a whole object-oriented programming environment may also be
viewed as a toolkit whose purpose it is to support the generation of general
applications. Some interesting new principles for such a toolkit can be found in
[Meyer 1993a].
 Part IV
Case studies
230 PART IV
9 A conference
management system

This and the next two chapters will be devoted to showing the use of the BON
method and notation for object-oriented modeling in three significant case
studies coming from different problem areas. Two of them were inspired by
working systems, and partly extracted from effective implementations.
The three areas exemplified will be a simple information system, a high-level
hardware encapsulation, and the mapping between an object model and a
relational data model. This chapter contains the first case study, whose objective
is to model an information system to support the management of a technical
conference.
The system should help the organizers follow up a series of events taking
place during the preparation of the technical program, and to handle incoming
contributions and attendee registrations. The basic tasks of the conference
system are the following:
• Monitor the scheduled events and check that all actions required are
carried out in time.
• Automate the process of most conference tasks avoiding duplication of
effort, and produce warnings to make sure deadlines will not be missed.
• Serve as a repository of information useful to both the technical and
organizational committees.
The typical conference we have in mind consists of three main parts: a set of
technical sessions presenting recent research and development results, a number
of half-day lectures (tutorials) on various topics related to the conference theme,
and an exhibition part featuring industrial applications in the area, mostly
commercial products. The technical presentations are usually bound and
published in conference proceedings, which are distributed to the attendees upon
arrival at the registration desk, and also sold separately.

231
232 A CONFERENCE MANAGEMENT SYSTEM

9.1 SYSTEM BORDERLINE

Users
It is customary to split the conference tasks between three reasonably
independent parties:
• A program committee responsible for the technical program. This
includes inviting researchers and developers in the field to contribute
material (call for papers), assigning contributed papers and tutorials to
suitable reviewers, selecting the contributions to include in the final
program, and grouping them into sessions and subject tracks.
• An organizing committee responsible for general policies and logistics.
This includes advertising and general mailings (invitations to attend),
decisions on pricing and capacities, sponsor negotiations, rental of suitable
premises, conference lunches and special evening arrangements,
accommodation prebookings, registration handling (attendees and
exhibitors), confirmation letters, publishing, tutorial notes, participation
lists, and attendee badges.
• An accounting department taking care of the financial side. This includes
invoicing, follow up on payment, and subsequent balance statistics on
gross sales and expenditures.
In many disciplines the major conferences are international, which means that
committee members as well as speakers and attendees may come from any part
of the world. It is then very unlikely that the people responsible for the
conference will all share the same geographical location. Particularly the
program committee is in most cases a highly distributed group and will for
practical reasons usually have only one common meeting to decide what papers
to include in the final program. The organizing committee may be more
homogeneous, but is usually separated from the program committee. Finally, the
financial side may use a standard system already present at one of the organizing
parties, or leave the task to an independent commercial service bureau.
This usually results in the use of unrelated software by the different groups.
An invoicing and accounting system with conference add-ons will perhaps
satisfy the administrative people, while the program committee members may
use some generally available software, like a spread-sheet program, to record
submissions and review reports.
The program committee receives contributions and sends letters of acceptance
or rejection. Even if in practice the organizing committee may take over the
bookkeeping, duplication of information and parallel updates without careful
 SYSTEM BORDERLINE 233

synchronization inevitably lead to oversights resulting in the wrong letter being

sent, duplicate letters, or no letter at all in some cases.
To counter this, a more rigorous system may define exact formats for data
interchange between heterogeneous systems along with orderly procedures for
keeping all distributed data consistent. In this example, however, we will just
assume one single support system, which will have all facts available and be
responsible for the overall consistency. This may be achieved if distant
committee members only communicate messages to the central system, which
then handles all correspondence with attendees and contributors.

Incoming information
Typical information entering the system is:
• registration form
• payment (bank wire, credit card slip, check)
• submitted paper or tutorial
• purchase order
• reviewer report
Figure 9.1 shows an example of what to expect of a registration entry format.

<Title> Dr
<Last name> Maugham
<First names> Julia Rachel
<Affiliation> Advanced Spacecraft Intl.
<Country> USA
<Postal address> 1010 Bourbon Street
<City zip code> New Orleans LA 70100
<Telephone> 504 333 22 11
<Facsimile> 504 444 55 10
<Electronic mail> [email protected]
<Selected tutorials> T3, T13, T21, T10
<Conference days> 2
<Amount paid> $1675.00
<Entitled rate> A
<Registration number> 341
<Registration date> 1993-12-30

Figure 9.1 Registration sample

234 A CONFERENCE MANAGEMENT SYSTEM

Outgoing information
Typical information leaving the system is:
• call for papers, invitations, promotional material
• letters of confirmation, acceptance, rejection
• attendee tickets
• preliminary and final program
• invoices, reminders
• proceedings, tutorial notes
• evaluation forms, badges
• session signs and posters
• list of attendees, financial status
Figures 9.2 and 9.3 show examples of an invoice and a financial report. In order
to find risk areas, we might ask ourselves what could go wrong in a poorly
designed support system. Possible negative effects include low credibility

Participant: Adams Matthews Stefenson

Reference: 301
Date: January 5, 1994
INVOICE
∞∞∞∞∞∞

WaO−Oh! ’94 Registration − Regular Rate

Quantity Description Unit Price Total

1 Conference 2 days $ 495.00 $ 495.00

4 Tutorials $ 295.00 $ 1180.00
TOTAL AMOUNT $ 1675.00

Payment:
• by check to the order of Intl. Conference Management Service Inc.
• by bank wire to the following banking references: ICMS − LA County Bank, Burbank,
California, account number: 00T9486Q094

Figure 9.2 Invoice sample

 SYSTEM BORDERLINE 235

STATISTICS − January 15, 1994

Number of registered persons: 890
Number of exhibitors: 62
Number of registered persons to Tutorial T1: 55
Number of registered persons to Tutorial T2: 62
.. ..
. .
Conference fee gross sale: $ 445,000
Exhibition fee gross sale: $ 124,000
Expenditures: $ 375,000
Current Balance: $ 194,000

Figure 9.3 Report summary sample

(conflicting information sent to speakers), loss of revenue (poor handling of

registrations), incorrect estimates regarding the number of luncheons,
simultaneous translation headphones, accommodation needs, and so on.
Well-identified functionalities of the system correspond to specific user roles
with assigned responsibilities. Some of them will be captured by representative
scenarios and illustrated through dynamic diagrams. The aim of our analysis and
design is roughly to decide what concepts in the problem domain are important
enough to be modeled as classes (thus affecting the software architecture), and
what should just be mapped into object states.
A conference support system may range from a quite simple system with
limited functionality to quite ambitious software addressing a considerable
amount of things perhaps not directly related to the event itself. One of the
major benefits in using the object-oriented approach is that it often enables a
continuous evolution without invalidating anything related to the initial
requirements and the working implementation.
Therefore, we will deliberately accept incomplete specifications and will see
that despite the fuzzyness of our initial set of requirements, we can still organize
our development around a few simple abstractions. Even for a small system
used by a limited number of people, we can design and release a first self-
contained functional version, and then gradually extend it with add-ons in much
the same way as is often formalized in large-scale industrial developments
through incremental deliveries [Gilb 1988].

9.2 CANDIDATE CLASSES IN THE PROBLEM DOMAIN

When looking for our first list of classes to start the modeling, our main concern
is to review de facto classes observed in the problem space and try to capture the
236 A CONFERENCE MANAGEMENT SYSTEM

most significant ones. There is no ideal world to discover here, but a bit of
modeling practice will help us develop a feeling for what kind of problem
domain abstractions are worth considering at the highest analysis level, and
which decisions should definitely be postponed till later.
For example, although CONFERENCE_ROOM may be a tangible object in
the problem domain, it is not important enough to start working with, whereas
REGISTRATION and COMMITTEE seem to capture profound roles with major
effects on the task of organizing a conference. The first candidate classes are
shown in figure 9.4. They are ordered alphabetically by name, because in real
situations when the number of candidates tends to be larger than in this example,
grouping by affinity does not work well and only makes it more difficult to
quickly find a class in the chart.
Different user groups have different views of the system, and some problem
domain concepts will be important to some users and ignored by others. For
example, exhibitor booths with their various sizes and prices will be of interest to
the organizing committee (as they represent a considerable source of income),
while the program committee may not even worry about whether there will be an
exhibition or not.
Other concepts may be shared by several user groups, but we must not forget
that such concepts are then never viewed in exactly the same way. For example,
a tutorial may be viewed by the program committee as a piece of structured
information of a certain technical quality, which belongs to a certain educational
track.
The same tutorial may primarily be viewed by the organizing committee as a
commercial offering, whose degree of success (and corresponding revenue)
depends on the number of enlisted participants and the resulting scores of the
evaluation forms. Finally, the people in the copy shop preparing the tutorial
notes will probably view it as a stack of documents whose important properties
are things like master copy quality, time of arrival, single or double paged
printing, binding method, and relative popularity (more work for successful
tutorials).
For concepts like this, we must decide which user viewpoints to support, since
this may highly affect the corresponding class abstractions. (In our example,
class TUTORIAL will be defined from the conference management viewpoint
ignoring the documentation aspects.)

9.3 CLASS SELECTION AND CLUSTERING

From the first identified set of classes, we may now start considering candidate
partitions dividing our system into different clusters. These upper-level clusters
represent the internal model of the basic system functionalities. We choose four
 CLASS SELECTION AND CLUSTERING 237

CLUSTER CONFERENCE MANAGEMENT SYSTEM Part: 1/1

PURPOSE INDEXING
General conference administration support. keywords: conference system, first
gathered analysis classes

Class / (Cluster) Description

ATTENDEE Registered person attending the conference.

COMMITTEE Peers organizing the conference and performing a fair

evaluation of the contributions to obtain a well-balanced
program of high-quality technical presentations reflecting the
major directions of the current research and industrial
development in the field.

CONFERENCE System for managing a selected domain event of a few days’

duration, consisting of technical presentations, a choice of
short tutorial courses lectured by domain experts, and a
commercial trade show displaying industrial advances in the
field.

CONTRIBUTOR Person listed in the final conference program.

PAPER Authored technical paper submitted to the program

committee chairperson.

PROGRAM Description of conference events.

REFEREE Program committee member refereeing submitted technical

papers for possible acceptance in the technical conference
program.

REGISTRATION Record of attendee participation keeping track of affiliation,

selected sessions, and registration fees.

SEND_OUT Message sent from the committees to anyone involved in the

event, such as attendees, speakers, staff, and conference
suppliers, regardless of the physical media used (electronic
mail, postal letter, or facsimile).

TASK Elementary action performed by the management system

upon receipt of an external event, or because of some
information contained in the agenda.

TUTORIAL Short training course submitted to the program committee

chairperson addressing a specific topic in the field.

Figure 9.4 First candidate classes

clusters, as shown in figure 9.5, to capture what we view as fairly distinct

responsibilities of the system. We also provide a textual description for each
introduced cluster, which will be the basis for any further discussion between the
users of the system and its designers.
The clusters are iconized according to the BON notation, and will be expanded
later when more is known about the classes contained in each cluster. With this
238 A CONFERENCE MANAGEMENT SYSTEM

ORGANIZATION TECHNICAL_EVENTS

PRINT_OUTS REGISTRATION

CONFERENCE_MANAGEMENT_SYSTEM

SYSTEM CONFERENCE MANAGEMENT SYSTEM Part: 1/1

PURPOSE INDEXING
General conference administration support. domain: information system
functionality: conference organization,
registration follow-up
keywords: conference, course, trade show,
attendee registration

Cluster Description

ORGANIZATION Handles all major events occurring during the organization of

the conference, from the initial decision of its title, date,
duration and estimated budget, all the way to successful
completion.

TECHNICAL_EVENTS Is responsible for putting together the program, recording the

status of all contributions, checking in referee reviews, and
following a precise timetable for what has to be done and
when.

REGISTRATION Collect any registration data needed for invoicing, production

of lists, printing of badges, and automated sending of
formatted letters. Store every piece of information related to
whatever may change the cost/benefit ratio of the conference:
attendee registrations, exhibitor registrations, cancellations,
room booking, catering, etc.

PRINT_OUTS Record every possible format used to print out something of

interest to the conference: preformatted letters, badges with
registration information, final program listing selection and
filtering criteria, up-to-date final program description, and
session signs to post on conference doors.

Figure 9.5 First cluster definition sketch

cluster partitioning, we may now attempt to assign the first identified set of
classes to the chosen clusters. This will still be somewhat tentative, since no
class chart has yet been written. Only a more systematic description of the class
properties and constraints may confirm the appropriateness of our choice.
 CLASS SELECTION AND CLUSTERING 239

Since all classes so far, except CONFERENCE and PROGRAM, seem to

represent concepts that can have several variants and therefore perhaps result in
inheritance hierarchies, we mark them as deferred until we know more. The
result is shown in figure 9.6, where CONFERENCE has been selected as the root
class. A common ancestor PRESENTATION of classes PAPER and TUTORIAL
has also been introduced (the heirs are not shown in the figure).

* PROGRAM
CONFERENCE TASK

* * *
COMMITTEE PRESENTATION SESSION

ORGANIZATION TECHNICAL_EVENTS

* * *
SEND_OUT REGISTRATION ATTENDEE

* * *
LISTING CONTRIBUTOR REFEREE

PRINT_OUTS REGISTRATION

CONFERENCE_MANAGEMENT_SYSTEM

Figure 9.6 Management and registration close-up

At this point, it becomes possible to outline dependencies between the

clusters. It is worth noting that the cluster dependencies in figure 9.6 imply an
order relationship that could help in staffing and scheduling a “cluster-based”
software development. We may therefore plan to benefit from a layered
approach. In a good design, it is often possible to direct the client relations from
the application specific clusters, through intermediate clusters, down to the
utility clusters. General utilities can usually be made application independent,
whereas application-dependent clusters are more closely related to the
corresponding problem domain.
240 A CONFERENCE MANAGEMENT SYSTEM

Clearly, cluster PRINT_OUTS encapsulates capabilities that are potentially

reusable in a system that does not have anything to do with conference
management and organization. The cluster can be designed and implemented
with facilities such as:
• formatting text according to predefined format templates (used for printing
badges, poster signs, etc.)
• automatically laying out a letter selecting from a number of predefined
styles: kind invitation, acknowledgment of receipt, friendly reminder, etc.
• selecting ready-to-use text substitution elements according to letter
recipient: male or female, attendee or speaker.

Classification
Looking closer at our first general classes reveals there are numerous variants
that may have to be considered, for example:
Registrations
advance registration, discount registration, complementary registration
(press, exhibitors, VIPs)
Contributors
conference speaker, co-author, keynote speaker, tutorial speaker, panel
moderator, panel speaker, session chairperson
Committees
program committee, organizing committee
Send outs
contributor send out, attendee send out, supplier send out
Paper
rejected paper, selected paper
To model systematic variations between similar objects we basically have two
different strategies to choose from: either rely on classification by inheritance, or
translate adaptations into object states using client relations. If similar objects
will often be handled together but still require different treatment depending on
the exact variant, then inheritance is often preferable since it allows dynamic
binding of tailored operations.
An extremely important principle in systems development is to get rid of as
many case discriminations as possible, since these are very sensitive to system
changes and therefore make maintenance much more difficult. However,
inheritance is not always feasible, particularly when many variants can be
 CLASS SELECTION AND CLUSTERING 241

combined or when the roles played by an object can change dynamically. While
client relations can vary over time, inheritance relations are fixed.
If we try to use inheritance to classify the roles of people attending the
conference, as illustrated in figure 9.7, we soon witness a combinatorial
explosion. We find that even multiple inheritance is not enough to model all

PERSON

CONF_COM_
MEMBER

ORG_COM_ PROG_COM_ REVIEWER VISITOR

MEMBER MEMBER

CONF_ TUT_ SPEAKER AUTHOR

TUT_COM_ ATTENDEE ATTENDEE
MEMBER

CONF_TUT_ INVITED_ CONF_ TUT_

ATTENDEE SPEAKER SPEAKER SPEAKER

Figure 9.7 Partial classification attempt

possible role combinations, because:

• most program committee members are probably reviewers,
but maybe not all
• a program committee member can also be a speaker
• some speakers are probably reviewers, etc.
Looking back at classes such as REFEREE, CONTRIBUTOR, and ATTENDEE
introduced in the first candidate list, we realize that we cannot keep all these
roles as classes. Instead, we replace them with one simple class PERSON,
whose objects will have registration properties to model the various attendee
profiles needed.
The same situation comes up with registrations. The numerous combinations
of registration types would pollute an attempted inheritance classification with
fork and join cycles, similar to what we saw in figure 9.7. So we will once again
rely on internal properties and stick with a single class REGISTRATION.
Fortunately, these problems do not arise with simpler classification schemes,
involving for example presentations and committees.
However, before committing ourselves any further as to what classification
structures to use, it is now time to sketch a set of analysis class charts in order to
gain more insight into the properties of the corresponding abstractions.
242 A CONFERENCE MANAGEMENT SYSTEM

9.4 CLASS DEFINITION

This task, involving untyped class specifications on our way to more formal
interfaces, is usually required when a system is delivered to end-users or
customers who do not have a very clear idea of what they want. Although the
chart format suggested by BON is designed to help further translation into class
specifications based on abstract data types, it also resembles a formatted memo
used to record major aspects of the problem.
One way to start defining the classes is to go back to the first class
descriptions (candidate list in figure 9.4) and outline a more elaborate version of
each entry, taking the changes discussed so far into account. Usually each class
description raises new abstractions that may in turn become class charts.

Organization
The first two charts address the conference as a whole and the committees in
charge of steering the event technically and practically, as shown in figure 9.8.

CLASS CONFERENCE Part: 1/1

TYPE OF OBJECT INDEXING

Generic conference cluster: ORGANIZATION
keywords: conference, scientific meeting

Queries Name, Location, Conference capacity,

Organizing committee, Program committee,
Program, Budget, Attendees, Insurance policy

Commands Prepare. Close down.

Constraints Run only once a year.

Total registrations ≤ conference capacity.
Location serviced by an international airport.
Accommodation capacity ≥ conference capacity.
Insurance policy subscribed.
Organized in collaboration with sponsors.

CLASS COMMITTEE Part: 1/1

TYPE OF OBJECT INDEXING

Conference committee cluster: ORGANIZATION
keywords: reviewers, experts, managers

Queries Chairperson, Members

Commands Set up. Select chairperson.

Constraints Chairperson is committee member.

Figure 9.8 Conference and committee class charts

 CLASS DEFINITION 243

We see that the CONFERENCE class has a number of constraints attached.

Some of these may later be turned into assertions in a typed class interface (pre-
and postconditions and class invariants), but there are also facts, intentions, rules,
etc., that do not translate into formal specifications but are still interesting to
record.
For example, easy access to an international airport is an important
consideration when choosing the conference premises. But it may not be
meaningful to express this formally in a system, since what should be considered
“easy access” may require human judgment on a case-by-case basis to balance
between measurable properties such as number of international flights to this
airport, commuting distance from airport, frequency of shuttle buses, taxi
capacity at bus terminal, etc. This type of constraint may instead become part of
an on-line set of guiding instructions in an implementation to help system users
make the right decisions.
Note that there is a difference between constraints of the type just described
and constraints that are in fact absolute rules, only too complicated to be
expressed formally. The latter kind may still be part of formal assertions in the
form of comments. After all, the main purpose of assertions is specification,
while the possibility of run-time checking is a great asset but no prerequisite.
Although we need to model committees with different responsibilities (like
program and organizing committees), they will all have members and a
chairperson, so we begin by encapsulating these features in a general class
COMMITTEE to be used as ancestor.
It is common to let a separate group of people handle tutorials, since the
educational aspects (and also commercial—successful tutorials can be a
substantial help in financing the conference) may need somewhat different skills
compared to selecting which scientific reports to include. The tutorial group is
often a subset of the program committee, but could be formed separately. Since
several of its tasks are similar to handling the scientific contributions (registering
incoming proposals, accepting or rejecting, grouping into subject tracks), it
seems natural to model the group as a committee of its own.
However, handling the scientific part does require formal procedures, such as
peer reviewing and transfer of copyrights, which are not needed for tutorials, so
we cannot use the exact same class. Figure 9.9 shows a feasible static
architecture to capture what we have just discussed.
The class TECHNICAL_COMMITTEE encapsulates what needs to be added to
the general COMMITTEE abstraction in order to also handle submitted
presentations, while ORGANIZING_COMMITTEE instead adds a great deal of
administrative features. Finally, class PROGRAM_COMMITTEE adds features
for sending out call for papers and for supervising a formal reviewing process. It
will also contain a tutorial subcommittee of type TECHNICAL_COMMITTEE.
244 A CONFERENCE MANAGEMENT SYSTEM

COMMITTEE

TECHNICAL_ ORGANIZING_
{ COMMITTEE COMMITTEE

{
tutorial_committee steering_board

PROGRAM_ { scientific_board
CONFERENCE
COMMITTEE

Figure 9.9 Committee classification

Note that it would not be appropriate to name the latter class

TUTORIAL_COMMITTEE, since that would model program committees as
special kinds of tutorial committees, which is clearly wrong. On the other hand,
since the features that will be in any technical committee (of which the program
committee is a special case) seem enough to fulfill the behavior of a tutorial
subcommittee, we see no need for a separate TUTORIAL_COMMITTEE class. If
more specialized tutorial needs come up later during design, the latter class may
be added as a sibling to PROGRAM_COMMITTEE.
As a consequence of our discussion, all the committee classes are effective
classes. Even if we do not yet see any need for direct client use of the top class
COMMITTEE, it may become useful as a simple encapsulation of a set of
members with a chairperson, so there is no reason to mark it as deferred. We
therefore remove the deferred marker introduced in figure 9.6. The class charts
for the committees are shown in figure 9.10.
The size of the program committee has been limited by constraints to 40
members. To achieve a good balance between theory and practice, we have also
decided to spread the committee members evenly between industry and
academia so that the numbers differ by at most four. Moreover, publishing
limitations, the capacity of the conference premises, the number of parallel
sessions wanted, and the duration of each talk and of the whole conference will
jointly restrict the number of papers that may be accepted. As seen in the chart,
this limit has been set to 30 in our example.
Finally, the last committee chart found in figure 9.10 shows the behavior
needed for the organizational part. Although the substance of the program is
prepared by the program committee, the final mailing is channeled through the
organizing committee, so consistency may be ensured and duplication avoided.
Since committee members are just SET [PERSON], the same person may belong
to any number of committees.
 CLASS DEFINITION 245

CLASS TECHNICAL_COMMITTEE Part: 1/1

TYPE OF OBJECT INDEXING

Conference technical committee cluster: ORGANIZATION
keywords: scientific community

Inherits from COMMITTEE

Queries Received submissions

Commands Define submission guidelines. Register submission.

Accept submission. Reject submission. Group submissions.

CLASS PROGRAM_COMMITTEE Part: 1/1

TYPE OF OBJECT INDEXING

Conference program committee cluster: ORGANIZATION
keywords: scientific community,
reviewers, peers

Inherits from TECHNICAL_COMMITTEE

Queries Tutorial subcommittee, Acceptance standards,

Review meeting, Referees

Commands Prepare call for papers. Dispatch submission to reviewers.

Enter review results. Select session chairs. Prepare final program.

Constraints Members ≤ 40.

Industrials − Academics ≤ ±4.
Final selected papers ≤ 30.

CLASS ORGANIZING_COMMITTEE Part: 1/1

TYPE OF OBJECT INDEXING

Conference organization committee cluster: ORGANIZATION
keywords: non-profit organization,
steering committee

Inherits from COMMITTEE

Queries Contracted exhibition services, Contracted publisher,

Regulation policy, Sponsors

Commands Mail call for papers. Mail final program. Mail exhibition kit.
Handle registrations. Send final papers to publisher.
Print and bind final tutorial notes.
Schedule paper and tutorials sessions.
Print attendee list. Collect evaluation sheets.

Figure 9.10 Specialized committee class charts

246 A CONFERENCE MANAGEMENT SYSTEM

Conference program
With the committee responsibilities and corresponding classification sorted out,
we may now proceed to take a closer look at the conference program. A
program consists of tutorials and technical presentations selected by the program
committee members during their unique and timely review meeting ending with
heartbreaking decisions (a rejection rate of 2/3 or more is often necessary to
achieve good technical quality at a popular scientific conference). Often, for
reasons of time and travel expenses, only a small part of the program committee
will in fact attend the meeting, while the rest of the members make their
contributions as reviewers.
The definition of the conference program directly yields its constituent parts.
The chart for class PROGRAM (figure 9.11) defines two queries, Agenda and
Contributions, to access the sessions and the selected papers and tutorials. The
general PRESENTATION class encapsulates common properties of technical
papers and tutorials. We have chosen to move class PROGRAM to the
ORGANIZATION cluster (cf. figure 9.6), since putting a program together

CLASS PROGRAM Part: 1/1

TYPE OF OBJECT INDEXING

All information pertaining to the final cluster: ORGANIZATION
conference program and its related keywords: agenda, program of sessions
preparation

Queries Paper and tutorial submission deadline, Final contribution deadline,

Preliminary program, Contributions, Agenda

Commands Update. Print.

Constraints Final deadline = submission deadline + 4 months.

Contributions accepted by program committee.

CLASS PRESENTATION Part: 1/1

TYPE OF OBJECT INDEXING

Submitted paper or tutorial cluster: TECHNICAL_EVENTS
keywords: scientific result, technical
achievement

Queries Presentation code, Title, Authors, Status, Speakers

Commands Accept. Reject. Hold.

Constraints Presentation code and title must be unique.

Figure 9.11 Program and presentation class charts

 CLASS DEFINITION 247

involves keeping track of deadlines, sending reminders, and so forth, while class
PRESENTATION remains in the TECHNICAL_EVENTS cluster.
Usually, there are many types of presentation: tutorial, technical paper,
keynote address, invited talk on a special topic, panel session, poster session,
“birds of a feather” session, workshop, and so on. For the sake of simplicity we
will only detail two specific cases here: technical paper sessions and tutorial
sessions. We define the special behavior needed for papers and tutorials,
respectively, as shown in figure 9.12.

CLASS TUTORIAL Part: 1/1

TYPE OF OBJECT INDEXING

Submitted tutorial cluster: TECHNICAL_EVENTS
keywords: speech, lecture, training, public
seminar

Inherits from PRESENTATION

Queries Capacity, Number of attendees,

Technical prerequisite, Track, Duration

Constraints Number of attendees ≤ capacity.

CLASS PAPER Part: 1/1

TYPE OF OBJECT INDEXING

Submitted paper cluster: TECHNICAL_EVENTS
keywords: contribution, publication,
article

Inherits from PRESENTATION

Queries Review reports, Average score

Commands Award best paper. Transfer copyright to publisher.

Constraints Not already published when submitted.

Figure 9.12 Tutorial and paper class charts

Note how the constraints section is again used to record business rules. In this
case, class PAPER states that a submitted paper must not have been published
before (or even submitted to another conference, to be more restrictive). Even if
this cannot be automatically checked, it is an important piece of information that
may perhaps be made part of the author and/or submission guidelines.
We may now sketch the static architecture of the analysis classes associated
with the technical program: its published incarnation and its presentation-based
format (figure 9.13). Note that all classes containing dynamically interesting
248 A CONFERENCE MANAGEMENT SYSTEM

SESSION

● ●
PAPER_SESSION TUTORIAL_SESSION

presentations: SET […] lecture

● reviews: SET […] ● ●
REVIEW PAPER TUTORIAL

● *
STATUS PRESENTATION

Figure 9.13 Technical events classes

data have been marked as persistent, which means that at system shut-down the
corresponding instances must be saved on secondary storage (object or relational
database, file system, etc.) for later retrieval. In a typical information system,
such as this, many classes tend to be persistent.
Class PRESENTATION is marked as deferred because it contains commands
for accepting or rejecting a submission, which needs to be done differently
depending on type of presentation. Class SESSION, on the other hand, might be
sufficient for workshops, panel sessions, etc., and is therefore not deferred. We
do not show the charts for the session classes, nor status and review (we will
return to them when we look at the formal class interfaces).

Registration
The next cluster to examine is the registration part of the system. We can model
registrations and their connections with attendees in several ways. One
possibility is to mirror the incoming pre-registrations, which often list several
participants from the same organization, and enter only one registration per
received order (group registration). This may seem nice and simple, but a
moment’s reflection reveals that it would in fact be a very inflexible solution.
If individual attendees can only be found through the corresponding
registration numbers, we will be in trouble each time a company calls to change
or cancel the registration for one person in a group registration. Each attendee
may also choose a combination of tutorials, so one registration for each attendee
with a two-way client relation between classes PERSON and REGISTRATION
seems a smarter choice (figure 9.14).
 CLASS DEFINITION 249

● attendee ●
REGISTRATION PERSON
registration

Figure 9.14 Registration and person static coupling

The PERSON class solves the classification problem raised earlier regarding
attendee roles. Each instance represents a person who is entitled to visit all or
parts of the conference, and will be attached to a corresponding registration
recording the terms. This will automatically lead to the proper badge being
printed, and a speaker will therefore not risk being refused entry to his own
technical session because an official prepaid registration was never issued! (This
actually happened to one of the authors some years ago in San Diego.)
Figures 9.15 and 9.16 show the corresponding class charts.

CLASS REGISTRATION Part: 1/1

TYPE OF OBJECT INDEXING

Record of attendee participation, keeping cluster: REGISTRATION
track of affiliation, selected sessions, and keywords: conference registration,
registration fees permission to attend, authorization

Queries Attendee, Conference days, Selected tutorials,

Date, Amount paid, Invoice sent, Confirmed

Commands Confirm participation. Invoice referred to attendee.

Send practical information documents.

Constraints Invoice has no effect for free access attendees.

Figure 9.15 Registration class chart

CLASS PERSON Part: 1/1

TYPE OF OBJECT INDEXING

Person whose address is kept track of by cluster: REGISTRATION
the conference: committee member, keywords: conference attendee, speaker,
(potential) attendee, referee, speaker, etc. exhibitor, lecturer, visitor, participant,
authorized person

Queries Name, Title, Affiliation, Address, Country, Registration

Commands Register. Cancel registration. Substitute other person.

Constraints Anyone entitled to visit the conference is registered.

Figure 9.16 Person class chart

250 A CONFERENCE MANAGEMENT SYSTEM

Printed material
During the preparation of the conference program and incoming registrations,
different types of letters are sent out by the conference management system:
invitations, confirmations, acceptance and rejection letters, and so on. A great
deal of other material is also printed before and during a conference, such as
address labels, evaluation sheets, attendee badges, and invoices. A deferred class
PRINT_OUT encapsulates general facilities for formatting and printing
documents, which in turn use a set of predefined templates describing the general
layout of different types of printed material, as illustrated in figure 9.17.

* layout *
PRINT_OUT DOCUMENT_FORM

ADDRESS_ CONFIRMATION_ LETTER_ BADGE_

MAILING
LABEL LETTER FORM FORM

AUTHOR_ ACCEPTANCE_ REJECTION_ STICKY_ INVOICE_

GUIDELINES LETTER LETTER FORM FORM

EVALUATION_ POSTER_ POSTER_ LIST_

BADGE
SHEET SIGN FORM FORM

ATTENDEE_ EVALUATION_
STATISTICS INVOICE
LIST FORM

PRINT_OUTS TEMPLATES

Figure 9.17 Print support

Typically, the template classes will specify fonts, point sizes, logos, line
drawings, etc., used for the various blocks of information in a document of a
certain type, while the information contents of each block will be supplied by the
specific printout class.

Conference management
Arranging a conference means a great deal of interaction with other parties
(authors, attendees, exhibiting companies, conference sites, hotels, travel agents,
exhibition service contractors, publisher, reviewers, committee members).
Having a successful program with satisfied participants requires a lot of
interdependent tasks to be carried out, agreements to be checked, deadlines to be
 CLASS DEFINITION 251

met, and logistics to be planned.

Trying to go a little further in our sketch of the initial set of classes, we may
now take a design decision, which is to introduce a dedicated class to help collect
and keep track of various important time-related activities. The class, which we
call TIMETABLE, will act as a data repository that is checked by the top-level
class at regular intervals to help follow up scheduled events and produce
warnings if important deadlines are about to be missed. This results in the partial
architecture shown in figure 9.18.

●
TIMETABLE
●
PROGRAM_ { reminder
COMMITTEE
program ● agenda: SET […] *
CONFERENCE { PROGRAM SESSION
●
ORGANIZING_ { attendees: SET […]
COMMITTEE
●
REGISTRATION

Figure 9.18 Conference management

When the system is restarted each morning, say, the root class CONFERENCE
checks the timetable to see if anything pressing needs to be done first. It then
issues warnings, prints recommendations, invokes appropriate actions in the
committee classes, or whatever the level of ambition in the implementation calls
for. The normal operation would then typically be to enter some kind of
standard command input loop, and let the users input registrations, select papers,
schedule sessions, etc., at their choice.
The class chart of the timetable is shown in figure 9.19. The submission and
final contribution deadlines were moved from the PROGRAM class into
TIMETABLE in order to have a comprehensive overview of all important time-
related events that need to be followed up. No commands are specified, since
the interface to TIMETABLE is not yet decided. One reasonable solution is to
store the time information in a plain parameter file, and simply use a text editor
to dynamically change entries when circumstances call for it.
Deadlines in relation to conference sites, hotels, caterers, etc., are mostly open
to discussion and may thus change several times during the conference
preparation, and if not enough good-quality papers and tutorials have been
received, it may be necessary to extend submission deadlines. The automatic
checking will probably need to be done more frequently as certain important
dates get closer.
252 A CONFERENCE MANAGEMENT SYSTEM

CLASS TIMETABLE Part: 1/1

TYPE OF OBJECT INDEXING

Keep track of important dates and cluster: ORGANIZATION
deadlines keywords: conference events, deadlines,
reminder, control, follow-up

Queries Today’s date, Frequency of checking, Last checked date,

Advertising dates, Call for papers mailing dates,
Invitation mailing dates, Manuscript submission deadline,
Formal review deadline, Review meeting date,
Final paper deadline, Final program at latest,
Lunch orders at latest, Hotel reservations at latest,
Author notification at latest, Early payment discount deadline

Constraints Checks performed at regular intervals.

Figure 9.19 Timetable class chart

9.5 SYSTEM BEHAVIOR

Events and scenarios
Using our initial sketchy description of the system and the first set of gathered
analysis classes, we may now outline a set of significant incoming external
events that will lead to interesting system behavior. These events (incoming
stimuli) result in the creation of new objects, or the passing of information
between active objects, or the entry and propagation of external data into active
objects. Some of them will also cause significant outgoing system responses to
be produced by internal actions.
A set of incoming events are collected in the chart shown in figure 9.20. A
second chart captures outgoing events that are system responses triggered after a
chain of external incoming and internal events is activated (figure 9.21). In
practice, it is usually not worthwhile to try and provide a complete set (in any
sense) of all possible event and scenario types. The aim is to select some
examples illustrating the most typical and important usage (concentrating on
potential trouble spots), and see how they map to the static architecture. Since
there is always a cost involved in working out a scenario and scenarios represent
some of the things most volatile to system architectural changes, they should not
be overdone.
Particularly for an evolving system with fuzzy requirements, as in this case,
the scenarios should mainly be viewed as a set of throw-away elaborations that
will help us check and improve the evolving static model, and not, for example,
as the formal basis of a future system test batch. System testing and long-term
documentation have many other aspects, and should be addressed separately.
 SYSTEM BEHAVIOR 253

EVENTS CONFERENCE MANAGEMENT SYSTEM Part: 1/2

COMMENT INDEXING
Typical incoming events triggering keywords: external events, incoming
interesting system behavior. stimuli

External (incoming) Involved object types

A contribution is registered CONFERENCE, PROGRAM_COMMITTEE,

PRESENTATION*

The formal review results of CONFERENCE, PROGRAM_COMMITTEE, PAPER,

submitted paper are entered REVIEW

A contribution is accepted or CONFERENCE, PROGRAM_COMMITTEE,

rejected PRESENTATION*, STATUS

An accepted contribution is CONFERENCE, PROGRAM_COMMITTEE,

attached to a session PROGRAM, SESSION, PRESENTATION*

A session chairperson is selected CONFERENCE, PROGRAM_COMMITTEE,

PROGRAM, PAPER_SESSION, PERSON

An attendee is registered CONFERENCE, ORGANIZING_COMMITTEE,

REGISTRATION, PERSON

Figure 9.20 Incoming events

We will limit ourselves to four scenarios in this example, while in a real

development we would probably select some more. The chosen scenarios are
listed in the scenario chart shown in figure 9.22, which acts as the front page of
the dynamic model, summarizing the dynamic object diagrams to follow.

What relations to emphasize

Whenever a message is expected to be exchanged between two objects, it is
necessary to verify that the calling object can reach the receiver object. This
means there must be transitive closure on client/supplier links between the
classes of the sender and receiver objects, respectively.
However, this does not necessarily mean it will be possible to follow public
query links from the sender to the receiver class. Several client dependencies
will arise through private features and through local variables used in the
implementation of public features. Whether such implicit client relations are
shown in the corresponding static diagrams or not, is a question of what we want
to emphasize and must be left to the designer’s good judgment.
Several considerations come in here. First, even when client relations between
clusters may be derived from the types of public features, we do not necessarily
254 A CONFERENCE MANAGEMENT SYSTEM

EVENTS CONFERENCE MANAGEMENT SYSTEM Part: 2/2

COMMENT INDEXING
Typical outgoing events. keywords: internal events, outgoing
responses

Internal (outgoing) Involved object types

Call for papers is sent CONFERENCE, ORGANIZING_COMMITTEE,

PERSON, MAILING

Invitations are sent CONFERENCE, ORGANIZING_COMMITTEE,

PERSON, MAILING

A paper is sent to referees CONFERENCE, PROGRAM_COMMITTEE, PAPER,

STATUS, REVIEW, PERSON

Warning issued for exceeding CONFERENCE, REGISTRATION, TUTORIAL

tutorial session capacity

An author notification is sent CONFERENCE, PROGRAM_COMMITTEE,

PERSON, PRINT_OUT*, LETTER_FORM

An invoice is sent to a prepaying CONFERENCE, ORGANIZING_COMMITTEE,

attendee REGISTRATION, PERSON, INVOICE,
INVOICE_FORM

A badge is printed CONFERENCE, ORGANIZING_COMMITTEE,

REGISTRATION, PERSON, BADGE, BADGE_FORM

The conference attendee list is CONFERENCE, ORGANIZING_COMMITTEE,

printed REGISTRATION, PERSON, ATTENDEE_LIST,
LIST_FORM

Figure 9.21 Outgoing events

want to show them in high-level views of the system. This is because different
client relations resulting from analysis and design choices are not necessarily
equally profound or stable.
Some relations may mirror very deep model decisions, while others are there
simply because we have been forced to make a choice between equally feasible
alternatives. The latter types are more likely to change over time, and in order to
reflect the more fundamental properties of a system architecture, they may be left
out of some static diagrams.
On the other hand, there may be a very profound client dependency between
two problem domain classes, even if the exact client relations are not reflected
by return types of public features. Such client dependencies may be an
important part of the high-level views, but cannot be labeled, since we do not
want to decide yet how the necessary connection will actually be implemented.
 SYSTEM BEHAVIOR 255

SCENARIOS CONFERENCE SUPPORT Part: 1/1

COMMENT INDEXING
Scenarios selected for study. keywords: paper evaluation, notification,
registration, badge printing

Input review results and compute score:

The scores found in the formal review reports for a submitted paper are entered into the
system, and an average score is computed for later use at the program committee review
meeting.

Accept or reject a paper and notify authors:

A submitted paper is selected and an acceptance or rejection date is entered; a notification
letter is sent to the first author.

Register attendee:
An attendee is registered with his/her address, and selected tutorials are recorded.

Print attendee badge:

An attendee is selected, and the corresponding badge is printed in the appropriate format.

Figure 9.22 Scenario chart

Since the dynamic connection between two objects will also imply some sort
of static chain between them (either before or behind the curtains in an eventual
implementation), working with dynamic scenarios often helps clarify modeling
questions of the kind just raised.

First scenario
The first object scenario deals with the input of review results, and is illustrated
in figure 9.23. The scenario describes the main steps involved in finding a paper

Scenario 1: Input review results and compute score

1−2 One of the submitted papers is selected
3−5 Formal review scores are input
6−7 Score average is computed and stored

5: User input
REVIEW

4, 6

1 PROGRAM_ 2, 3, 7
CONFERENCE PAPER
COMMITTEE

Figure 9.23 First object scenario

256 A CONFERENCE MANAGEMENT SYSTEM

and entering its corresponding review results. Since an eventual implementation

will involve many details that we still want to keep open, the idea is to include
only essential object communication in an object diagram.
For example, to find a desired paper the user might first click on “show
submitted papers” in the main menu of class CONFERENCE, then repeatedly
select “next paper” from a local menu in class PROGRAM_COMMITTEE until
the sought paper appears on the screen. However, in a different implementation
the user may instead click directly on “find paper” in CONFERENCE and input a
pattern matching string for the desired title.
Since there are so many possible ways to let a user navigate in an information
system, there is no point in trying to specify what classes will take care of the
successive user control input. Instead, we concentrate only on the information
content entering the system at various steps in a scenario.
So the initial user interaction is not included in the diagram, but we assume
that the CONFERENCE object is somehow told to select a paper. This, in turn,
leads to a message being sent to the PROGRAM_COMMITTEE object, which is
in charge of the list of submissions (1). PROGRAM_COMMITTEE then consults
its set of papers until the right paper is found (2).
This will typically mean iterative calls to a generic SET class to obtain one
PAPER object after the other, telling each of them to present some identification
(title, authors, etc.) and let user interaction decide when to stop. However, as we
have just argued, the details of such interactions are better left to later phases.
Moreover, we normally do not want to clutter object diagrams with messages to
standard container objects. Instead, we use stacked object icons to imply lists
and sets.
So the message labeled 2 in the diagram is actually an abstraction of a
navigation within a set of objects, which may eventually involve other objects in
a final implementation (not only design objects and general utility objects, but
also analysis objects). For example, when presenting information about each
paper in order to let the user select the proper one, there are several alternatives.
If author names are part of the information, then PERSON objects will be
involved also in this step, but this will not be the case if only titles are presented
during the search.
When the PAPER object has been selected, PROGRAM_COMMITTEE calls it
to obtain the corresponding set of REVIEW (3), whose members are then
successively called to input the referee scores (4). In the diagram, we represent
the user input required to transfer the scores to the system as an external message
sent to a corresponding REVIEW object (5). This does not necessarily mean that
the actual user interaction takes place in class REVIEW, since (again) the details
of user terminal input are typical implementation issues that should be kept open
at this stage.
 SYSTEM BEHAVIOR 257

However, we still want to be able to show the interesting fact that external
user data will be entering the system as part of the scenario. Therefore, the
convention is to depict the data entry as a user input message directed to the
receiving object, even if it may be implemented differently. The input may for
example first be accepted by a centralized user input object, whose features will
be called by PROGRAM_COMMITTEE, which will in turn send the data to
REVIEW.
Unless there are special reasons, we do not care to illustrate external
communication that only results in browsing through a system or running up and
down menu hierarchies. Therefore, external incoming messages depicted in
object diagrams will correspond to incoming data flow or selections leading to
new objects being created or read from secondary storage; that is, interactions
that imply some significant change to the system state and its information
contents.
Similarly, outgoing messages will correspond to something substantial leaving
the system, like data transferred to external media, calls to other systems, printed
reports, etc., and not just trivial things like writing a prompt or an error message
on a user’s terminal.
Finally, when all reviewer scores have been entered, the set of REVIEW is
again consulted by PROGRAM_COMMITTEE, which computes an average
score (6), and sends it to the PAPER object for storage (7).

Second scenario
The next scenario, which illustrates the acceptance procedure for an individual
paper, is in figure 9.24. The first part of the scenario shows the messages
involved in recording the acceptance status of a paper. Steps 1−2 first select the
proper paper (same as in previous scenario). PROGRAM_COMMITTEE then
sends an accept or reject message to the paper (3), which in turn tells the
corresponding STATUS object to change the paper status (4).
This will need user input for setting the date of acceptance or rejection (5).
Again, we use the convention to depict user input as directed to the object which
will hold the information (STATUS in this case), regardless of the way it is going
to be implemented.
The second part of the scenario shows the author notification procedure. First,
PROGRAM_COMMITTEE sends a message to PAPER (6) to get the authors,
which is a SET [PERSON], and then consults this list to obtain the first author
(7). Depending on the decision reached, PROGRAM_COMMITTEE then creates
the proper form of notification letter (8), which involves a conditional choice
between two alternatives: rejection or acceptance. Both are covered by passing
the message to an object group rather than to individual objects.
258 A CONFERENCE MANAGEMENT SYSTEM

Scenario 2: Accept or reject a paper and notify authors

1−2 A paper is selected
3−5 Acceptance or rejection date is entered
6−7 The first author of the paper is selected
8 A notification letter is created
9−11 The letter is sent to first author

5: User input
LETTER_FORM
PERSON STATUS
10
7 4
ACCEPTANCE_ 8, 9 PROGRAM_ 2, 3, 6
LETTER PAPER
11: Letter printed COMMITTEE

REJECTION_ 1
LETTER
CONFERENCE

Figure 9.24 Second scenario

Finally, the notification letter is told to print itself (9), which makes it consult
the proper format template (10), and then the formatted letter leaves the system,
probably directed to some external print spooling system (11). The actual
mailing is assumed to be done manually. The inclusion in the final program is
postponed until the final version of the article is received for inclusion in the
proceedings.
Note that the reviewer scores do not enter the picture here, because these have
already been scrutinized at the program committee’s formal review meeting.
Scenario 2 only effects the decisions taken at that meeting.

Third scenario
This scenario shows the registration of a conference attendee (figure 9.25). The
user directs the system to enter registration information, which transfers control
to the ORGANIZING_COMMITTEE (1), which in turn creates a new
REGISTRATION object (2). The registration data is input (3), which is marked
as a message from the external world to the REGISTRATION object, again
regardless of how the data entry will actually be implemented.
Based on the input data, a PERSON object is either created or reused from the
existing set of persons in the conference database and the corresponding address
 SYSTEM BEHAVIOR 259

Scenario 3: Register attendee

1−3 A new registration is created
4−5 Attendee references are recorded
6−7 Selected tutorials, if any, are registered

PERSON

4
3: User input
6 ORGANIZING_ 2, 5, 7
TUTORIAL REGISTRATION
COMMITTEE

CONFERENCE

Figure 9.25 Third scenario

information is entered or updated (4). (In accordance with our earlier design
decision, each listed person in a group registration from an organization will be
registered separately to enable individual changes for each participant.) The
PERSON object is then recorded in the REGISTRATION (5). Then, again based
on the received input, possibly selected tutorials are recorded in the same
REGISTRATION (6−7).

Fourth scenario
The last scenario (figure 9.26) shows an attendee badge being printed for a
person who is entitled to visit some part of the conference. The attendee is
selected from the registration database (any person with permission to visit the
conference in any capacity must already have been registered), and a badge is
printed according to the information stored in the corresponding registration,
using the proper format description.
Besides printing the attendee name tag, the badge object may also output
instructions requesting some manual attachments to the badge, such as ribbons of
various colors stating the title of attendees with special functions: conference
chair, speaker, session chair, staff, program committee, etc. A registration
category code could be used to discriminate between the roles of different
attendees, and prevent, for example, a keynote speaker being invoiced for
visiting the conference.
260 A CONFERENCE MANAGEMENT SYSTEM

Scenario 4: Print attendee badge

1−3 Attendee is selected
4−6 A badge is created and printed

REGISTRATION BADGE_FORM

2 5

3 ORGANIZING_ 4 6: Badge printed

PERSON BADGE
COMMITTEE

CONFERENCE

Figure 9.26 Fourth scenario

9.6 FORMAL CLASS DESCRIPTION

Each class can now be described in a more thorough and formal manner,
deciding the signature of each query and command, starting from the first
definitions in the class charts. For each feature in a formal class interface, the
type of each input argument as well as the return type (in the case of a query) is
specified as either a user-defined class type, or a predefined analysis class type.
Besides signature specifications, which give the abstract syntax of the
corresponding feature calls, formal assertions are used to specify semantic
properties through pre- and postconditions and a class invariant. The bulk of
these semantic specification elements will usually be created during detailed
design and implementation.

Predefined types for analysis and design

Predefined types are very general abstractions that can be reused in a wide
variety of analysis models, and each development organization working with
BON should define a standardized extensible set of such types with well-defined
semantics. In this example, we assume the basic predefined types BOOLEAN,
VALUE, and SET.
The BOOLEAN type has value semantics, which means it is always attached to
an object and can therefore never be void (the attached value is either true or
false). The same is true for very basic numeric types, such as INTEGER and
REAL, but these are normally considered too special to be used at the highest
 FORMAL CLASS DESCRIPTION 261

analysis level. Since we want to keep the representation of various values open
as long as we can, we only use the familiar low-level types in case they are
actually part of the requirements (such as when re-enginering existing software
where already specified interfaces must be kept unchanged).
However, BOOLEAN is a notable exception, since although it will probably
map directly to some type in most implementation languages, it represents a
fundamental concept in logical reasoning and is therefore in fact extremely
general (in a sense, much more general and high level than typical top-level
abstractions such as airplanes and hospitals).
The SET type is a general generic container class that can be used to hold any
specified type of elements. Note that we are referring to very general data
abstractions without committing ourselves to any implementation. Therefore,
abstract data containers also represent high-level concepts that may be used to
reason about any kind of analysis objects.
While BOOLEAN and SET may often map directly to specific implementation
classes or to basic types, VALUE has more a flavor of TBD (To Be Decided)
about it. At the early stages of modeling, we often come across a large number
of queries returning various pieces of interesting information, but whose exact
format we do not yet want to fix.
For example, the capacity query in the CONFERENCE class (figure 9.27) may
return just an INTEGER stating the maximum number of attendants. However,
we may also find more specific information useful: how much latitude before the
crowding level becomes unacceptable vs. absolute limit (local fire regulations),
all available space already included or possibility to increase capacity (last
minute rental of adjacent annexes), etc.
Using a STRING would mean freedom to express more such details to a
human user, but instead prevent easy automatic comparison. Using a dedicated
analysis class CAPACITY could solve both problems, but also means increased
complexity. Since the capacity considerations are not part of the fundamental
system structuring questions, we would like to leave the corresponding decisions
for later.
One possibility would be to just leave capacity untyped (which is the standard
approach in many other methods), but the BON spirit is to use a very general
type instead. By specifying VALUE, we express that the information returned
will not be any of the high-level analysis classes modeled so far, that it could
(and often will) be a very basic type, but also leaves open later refinement into a
new type defined by a design or implementation class. The idea is to encourage
the analyst not to make premature decisions, but still maximize the possibility to
express and reason about things that should be decided.
Therefore, use of the VALUE type is very common, particularly in typical
administrative information systems like our conference example. If we know we
262 A CONFERENCE MANAGEMENT SYSTEM

are dealing with a piece of numeric information, we could use the basic type
NUMERIC, which guarantees that the standard arithmetic operations will be
defined for the corresponding object (but still does not preclude a string
representation, if so desired).
We will now look at the translation from charts to formal interfaces, one
system area at a time.

Organization cluster
We begin with the organizational part of the conference, and look first at the root
class with its associated program and list of important dates. The interface
specifications for these classes are shown in figure 9.27. The organizing and
program committees as well as the conference program are seen as integral parts
of the conference. This is emphasized by using the aggregation variant of the
corresponding client relations.

CONFERENCE ● PROGRAM ●

name: VALUE preliminary_program: SET [SESSION]

location: VALUE
{ agenda: SET [SESSION]
capacity: VALUE update_agenda
budget: VALUE print
insurance_policy: VALUE
reminder: TIMETABLE
program:{ PROGRAM ●
TIMETABLE
attendees: SET [REGISTRATION]
steering_board:{ ORGANIZING_COMMITTEE today: DATE
technical_board:{ PROGRAM_COMMITTEE
{ last_checked: DATE
run check_every: DURATION
Invariant ad_runs: SET [DATE]
steering_board ≠ ∅ ; call_for_papers: SET [DATE]
technical_board ≠ ∅ ; manuscript_deadline: DATE
program ≠ ∅ ; final_deadline: DATE
insurance_policy ≠ ∅ ;
attendees.count ≤ capacity; lunch_orders_before: DATE
¬ steering_board.sponsors.empty hotel_reservations_before: DATE
notification_within: DURATION
early_payment_before: DATE
check

Figure 9.27 Conference and program

 FORMAL CLASS DESCRIPTION 263

In the invariant of class CONFERENCE, this is emphasized by specifying that

the references to these parts must not be void, and so must be set up already at
object creation time. Since the committee members may not all have been
selected when the system is first started (and the program is usually empty or just
contains a preliminary sketch), the parts may of course be gradually filled with
content. The invariant just specifies that the conference object must always be
attached to the corresponding objects, empty or not.
Two features of class SET, the queries empty (returning true if there are no
elements in the set) and count (returning the number of elements), are also used
to express assertions about sponsors and number of attendees. We have chosen
to delegate the responsibility of maintaining a set of sponsors to the
ORGANIZING_COMMITTEE, but the constraint stating that the conference
should be organized in collaboration with sponsors (see the class chart in
figure 9.8) belongs to the conference proper rather than the committee class.
This is the reason why the assertion ¬ steering_board.sponsors.empty is
placed in the conference class, and not in the organizing committee class.
Generally, if every object of a class must obey certain rules under all
circumstances, these rules belong in the invariant of the class. But if the rules
are just imposed because the class is being used by other classes in a restrictive
context, then the rules belong in assertions about the client classes.
Looking at the class chart (figure 9.8) again shows that of the three commands
defined there (prepare, run, close down) only run was kept. This is not unusual,
and again shows the hybrid character of the class charts serving both as initial
class definitions and as generalized memos. They often contain operations
which mirror actions in the problem domain, but which may later turn out to
have no counterpart in a computerized system.
As a conference organizer, we may have a clear view of three distinct phases
regarding the tasks to be carried out. First we prepare everything during a period
of several months (marketing, program preparation, pre-registration, etc.), then
we actually run the conference during some usually extremely hectic days, and
finally we close it down and take care of the aftermaths.
However, for a support system this division hardly makes any sense, since the
same system should be run on a day-to-day basis for as long as it may help its
users. So it becomes natural to have a single run command that will execute at
startup time each day, read in whatever persistent objects were saved in previous
system sessions, check important deadlines, output reminders, and then accept
general user commands.
In the reminder class TIMETABLE, we have assumed the existence of two
basic temporal classes, which may be part of a general reusable time utility
cluster, as shown in figure 9.28. Class DATE represents an absolute point in
time, while DURATION represents a relative time period. The timetable is
264 A CONFERENCE MANAGEMENT SYSTEM

TIME_UTILITIES

DATE DURATION

Figure 9.28 Temporal utilities

supposed to be checked at regular intervals (expressed by the query check_every)

and author notifications must be sent within a certain time limit after the formal
decisions have been reached.
The class interfaces of the committees with their classification are shown in
figure 9.29. The reader may have noted that there are no input arguments in the
feature signatures of our high-level analysis classes specified so far. This is no
coincidence, since the input information required at the upper system levels often
needs user interaction whose details it would be premature to specify already in
the analysis model.
Therefore, when modeling interactive systems, input arguments usually start
to appear in BON formal class descriptions only during design. In the case when
we model a system that mainly interacts with other systems, or develop reusable
class libraries, things may be quite different.
In figure 9.29 we see that the class ORGANIZING_COMMITTEE has some
queries whose return types are SUPPLIER objects or sets thereof. However, we
are not going to define the supplier class in our analysis model, so one might
question its use. Why not just specify VALUE at this point, and then refine the
type later in the process, when more is known?
The reason is that although it may not be possible (or desirable) at an early
stage to decide any interface details for a certain group of values, we may still
suspect that they will have enough things in common to warrant separation by
choosing a special ancestor type for them. Simply assigning VALUE as the type
would not capture that similarity.
In the PROGRAM_COMMITTEE interface (figure 9.29), we have an example
of feature redefinition. Any TECHNICAL_COMMITTEE object has a set of
sessions attached, since what makes the corresponding class different from a
general COMMITTEE is the handling and selection of presentations to be
included in a program. In class PROGRAM_COMMITTEE, we add the formal
reviewing process and a subcommittee feature as a further specialization. Since
we know that the elements in the sessions set must be of type PAPER_SESSION,
we can redefine the feature to return this type of set. The redefinition is
specified by two plus signs, signifying that the feature is “twice implemented”.
 FORMAL CLASS DESCRIPTION 265

TECHNICAL_COMMITTEE ● COMMITTEE

Inherits: COMMITTEE chairperson: PERSON

members: SET [PERSON]
submissions: SET [PRESENTATION]
set_up
sessions: SET [SESSION]
! ¬ members.empty
register_submission
select_chair
select_presentations
! ¬ chairperson.empty
! ¬ sessions.empty
group_submissions Invariant
define_guidelines chairperson ∈ members

PROGRAM_COMMITTEE ● ORGANIZING_COMMITTEE ●

Inherits: TECHNICAL_COMMITTEE Inherits: COMMITTEE

referees: SET [PERSON] publisher: SUPPLIER

tutorial_committee: TECHNICAL_COMMITTEE exhibit_service: SUPPLIER
acceptance_criteria: VALUE regulation_policy: VALUE
sessions++ : SET [PAPER_SESSION] sponsors: SET [SUPPLIER]
select_papers send_call_for_papers
{ ^TECHNICAL_COMMITTEE.select_presentations } send_final_program
select_session_chairs send_exhibition_kit
! ∀ s ∈ sessions • (s.chair ∈ members) handle_registrations
prepare_call_for_papers prepare_proceedings
send_to_reviewers prepare_tutorial_notes
enter_review_scores schedule_sessions
prepare_final_program print_attendee_list
Invariant collect_evaluation_sheets
members.count ≤ 40

Figure 9.29 Committees

Since tutorial_committee uses the general class for technical committees, there
is no corresponding specialization in which to redefine the sessions feature for
the tutorial committee. Instead, PROGRAM_COMMITTEE becomes responsible
for ensuring that the elements in the set tutorial_committee.sessions will always
266 A CONFERENCE MANAGEMENT SYSTEM

have type TUTORIAL_SESSION. In the first case, redefinition permits static

checking by a compiler that the session elements will have the correct subtype,
while in the second case only dynamic checking is possible.
Finally, we note that feature select_presentations has been renamed
select_papers to better reflect its usage in this subclass. (See section 3.8 for a
description of the BON graphical class interface symbols.)

Registration cluster and technical events cluster

The coupling between the persistent classes REGISTRATION and PERSON is
detailed in figure 9.30.

REGISTRATION ● PERSON ●

attendee: PERSON registration: REGISTRATION

registered_at: DATE name: VALUE
amount_paid: VALUE affiliation: VALUE
invoice_sent: BOOLEAN address: VALUE
confirmed: BOOLEAN postal_mail: VALUE
paper_sessions: BOOLEAN email: VALUE
selected_tutorials: SET [TUTORIAL] phone: VALUE
Invariant fax: VALUE
¬ paper_sessions → tutorials.count > 0;
attendee ≠ ∅ ; registered_at ≠ ∅

Figure 9.30 Registration

Each registration is attached to one, and only one, person. This permits us to
easily specify a different set of tutorials and to make individual changes also for
attendees that were collectively enlisted by one organization. Applicable
company discounts and the like will be distributed on each attendee.
Conversely, each person entitled to visit all or parts of the conference will
automatically have a registration created and a registration attribute may be
added, if desired, to record the reason for complimentary non-paid access rights
(invited speaker, sponsor representative, committee member). The invariant
states that a registration must imply access to either the scientific program or
some tutorial (or both).
The interfaces for the classes which are part of the technical program are
shown in figure 9.31, and the final static architecture of the conference system in
figure 9.32. The invariants of the session classes state that all presentations
 FORMAL CLASS DESCRIPTION 267

REVIEW ● STATUS ●

reviewer: PERSON received: DATE

score: VALUE review_started: DATE
comments: TEXT accepted: DATE
Invariant rejected: DATE
score ∈ { ’A’ .. ’D’} final_received: DATE
Invariant
received ≤ review_started;
review_started ≤ final_received;
accepted = ∅ or rejected = ∅

PAPER ● PRESENTATION * TUTORIAL ●

Inherits: PRESENTATION code: VALUE Inherits: PRESENTATION

title: VALUE
copyright_transferred: BOOLEAN capacity: VALUE
authors: SET [PERSON]
reviews: SET [REVIEW] attendee_count: VALUE
status: STATUS
final_score: VALUE prerequisite_level: VALUE
speakers: SET [PERSON]
award_best_paper track: VALUE
accept *
transfer_copyright duration: DURATION
! status.accepted ≠ ∅
? status.accepted ≠ ∅ accept+
reject *
! copyright_transferred reject+
! status.rejected ≠ ∅
accept+
Invariant
reject+
∀ p, q: PRESENTATION |
p ≠ q • p.code ≠ q.code
and p.title ≠ q.title

PAPER_SESSION ● SESSION TUTORIAL_SESSION ●

Inherits: SESSION chair: PERSON Inherits: SESSION

code: VALUE
presentations: SET [PAPER] lecture: TUTORIAL
track: VALUE
Invariant start, end: DATE Invariant
∀ p ∈ presentations • conference_room: VALUE lecture.status.accepted ≠ ∅
p.status.accepted ≠ ∅
Invariant
start < end

Figure 9.31 Technical events

268 A CONFERENCE MANAGEMENT SYSTEM

ORGANIZATION OUTPUT

COMMITTEES TEMPLATES

●
ORGANIZING_ LETTER_ BADGE_
COMMITTEE FORM FORM
COMMITTEE

STICKY_ INVOICE_
●
TECHNICAL_ { FORM FORM
COMMITTEE

tutorial_committee POSTER_ LIST_

FORM FORM
●
PROGRAM_
COMMITTEE
EVALUATION_
FORM
{

scientific_board, steering_board

● program ●
CONFERENCE { PROGRAM

* layout *
reminder PRINT_OUT DOCUMENT_FORM

●
TIMETABLE

ADDRESS_
MAILING
REGISTRATION LABEL

● attendee ● CONFIRMATION_
REGISTRATION INVOICE
PERSON LETTER
registration

ACCEPTANCE_ REJECTION_
LETTER LETTER

TECHNICAL_EVENTS
AUTHOR_ POSTER_
GUIDELINES SIGN
● ●
PAPER_ SESSION TUTORIAL_
SESSION SESSION
EVALUATION_
BADGE
SHEET
presentations: SET […] lecture
● * ●
PAPER PRESENTATION TUTORIAL ATTENDEE_
STATISTICS
LIST
reviews: SET […] status
PRINT_OUTS
● ●
REVIEW STATUS

Figure 9.32 Complete static architecture

 FORMAL CLASS DESCRIPTION 269

attached must be accepted, while the more complicated invariant of the

presentation class says that the code and title of each submission must be unique.
To express the latter, we let p and q range over all pairs of PRESENTATION
objects, and assert that the corresponding attributes are different. The invariant
of class REVIEW asserts that the score is always one of the four characters ’A’
through ’D’ expressed as membership of an enumerated set. (See section 3.12
for a full description of the BON predicate logic.)
10 A video recorder

In this case study, we will model the software control system for a simple video
tape recorder. The purpose is to show the danger of concentrating too much on
the underlying hardware to be modeled. Classes that reflect tangible hardware
objects may occur at some point during detailed design and implementation, but
they may not be the ones to use at the topmost level of a system if we want to
achieve reusability.

10.1 SYSTEM BORDERLINE

Most standard hardware components in a modern video cassette recorder (VCR)
can be controlled by micro computer logic, thus enabling easily tailorable
operation for different needs. Our task is to produce a high-level design of the
control program to be installed in the Micro Computer Unit (MCU) of the new,
simple, but reliable and aggressively priced model Multronic 2001, currently
being developed at the Bonasonic company.
The technical manager of the Home Video Division has just returned from a
three-day seminar on object-oriented abstraction and software contracting, and is
very enthusiastic about its potential for reuse of knowledge and increased
product quality. Although the hardware and main functions have already been
completed, there is still time to apply the new ideas to the controlling software,
whose implementation is just about to start.
Knowing the dangers of rushing into new technology without adequate
background, the manager has engaged a consultant for in-house training in
object-oriented analysis and design, and to support the software engineers during
the initial system modeling.

Internal operation
The principle mechanical components of the video recorder are sketched in
figure 10.1. The model has four heads: two rotating video heads, a fixed

270
 SYSTEM BORDERLINE 271

Video drum

Guide Guide
Head 1 Audio/sync head
Eraser head
Pinch roller
Capstan
Head 2 shaft
Guide

Cassette
Supply reel Take-up reel

Brake Brake
solenoid solenoid

Figure 10.1 Video recorder: tape in read/write position

audio / sync head, and a fixed eraser head. The audio head handles recording and
playback of sound, and also reads sync pulses telling the current location of a
moving tape. The identical video heads are mounted on a rotating drum and are
used for playback or recording of a video signal, depending on current mode.
The rotating heads scan the tape, receiving or recording one picture frame per
revolution. Since the two heads are mounted at 180°, they can take alternate
turns: when one head is about to leave the tape, the other enters and takes care of
the next frame. The rotation makes it possible to scan and show a still picture
also when the tape is stopped in load position (pause function). The fixed eraser
head is used to demagnetize the tape just before recording.
The machine has five separate motors: a cassette motor, a load motor, a drum
motor, a reel motor, and a capstan motor. When a cassette is inserted in the
machine, it is pulled in by the cassette motor in two steps: first in, then down.
Sensors report cassette-in and cassette-down respectively, and a time-out is used
to eject a tape that has not reached the down position within 5 seconds.
When a cassette is down, the tape position is controlled by the load motor.
The tape has three positions reported by sensors: released, sync only, or loaded.
To eject the cassette, the tape needs to be in the released position. For fast
forward or rewind, the load motor brings the tape into contact with the audio
head so recorded sync pulses can be read. For playback or record, it brings the
272 A VIDEO RECORDER

tape to the loaded position shown in figure 10.1 by moving the two guides
inwards on each side of the video drum.
On playback or recording, the pinch roller is pressed against the capstan shaft
by a pinch roller solenoid, and the capstan motor moves the tape. The winding
and unwinding of the tape reels in the cassette is controlled by the reel motor. A
reel idler is pressed against either reel depending on direction, and the reel motor
thus rotates the corresponding take-up reel.
The tape tension at each reel is reported by sensors and used to control the reel
speeds through the reel motor on the take-up side, and through the reel brake
solenoid on the supply side. Sensors also detect write-protected mode and
beginning and end of tape, so the reel motor can be safely stopped and full reel
brakes applied in time.
The Multronic 2001 includes a data bus that is used by the MCU to address
the ICs controlling various hardware components.

External operation
The front panel of the VCR is shown in figure 10.2. Besides the standard
playback and recording functions, up to 32 channels may be preset for easy
reception of satellite stations, and up to eight recordings can be programmed in
advance.

CHANNEL
VHS MON
FRONT LOADING SYSTEM 13:05
REC 10
0237
1 2 3 4 5
EJECT STOP REC PLAY
6 7 8 9 0
MENU NEXT PREV
POWER PAUSE REW FF
SELECT ACCEPT CANCEL

Bonasonic Multronic 2001

Figure 10.2 Video recorder: front panel

The channel is selected by the UP and DOWN channel buttons or by the

numeric keyboard. The current time is shown in the left upper part of the display
window, and the current channel in the right upper part. When a cassette is
loaded, a tape-in indicator lights up in the middle part of the display, and the
current mode of operation and tape position is shown. In figure 10.2, the VCR is
recording from channel 10 and the position counter has reached 237. The
buttons below the window are used to control tuning of channels, setting the
clock, and programming recordings. Pressing the MENU button displays a
 SYSTEM BORDERLINE 273

selection list in the lower part of the window, as shown in figure 10.3. In this
case there was no cassette in the machine, so the upper middle part of the display
is off.

MON
13:06 10
tune 1 2 3 4 5 6 7 8

Figure 10.3 Video recorder: selection menu

Setting the clock, tuning a station, or programming the timer at position 1−8
can now be done by first positioning the rectangular cursor using the NEXT and
PREV buttons and then pressing the SELECT button. Programmed recordings are
marked by a small dot below the corresponding position.
The left part of figure 10.4 shows timer programming and the right part shows
how to set the clock. Both dialogs can be carried out while the tape is operating,
for example rewinding as in the right display. Input is entered from the numeric
keyboard into the current field, and the rectangular cursor is automatically
advanced upon valid completion of a field. The cursor can also be moved using
NEXT and PREV to correct individual entries. When a record is completed it is
stored by pressing the ACCEPT button, while pressing CANCEL leaves the menu
dialog and returns to normal operation.

MON MON
13:06 STOP
0312
10 13:06 0123
10
PROG CHAN DATE START END MONTH DAY HOUR MIN
04 03 0222 14:10 15:45 08 25

Figure 10.4 Video recorder: timer and clock dialogs

If tuning is selected, the VCR starts searching for the next available station at
the video input, and the tune entry flashes until a station is found. Pressing
ACCEPT then stores the tuned frequency at the current channel.
Playback, recording, and tape positioning can also be controlled by the
battery-powered infrared remote controller shown in figure 10.5. In the back of
the chassis there are some connection sockets to external equipment and a switch
to select input from either a video camera or a tuner.

System viewpoint
The enclosing hardware and specified functionality defines the system borderline
exactly, so we do not need to spend initial time figuring out what parts of the
274 A VIDEO RECORDER

CHANNEL
REC

PAUSE REW FF

STOP PLAY

Bonasonic
Remote Control Unit

Figure 10.5 Video recorder: remote controller

problem domain should be modeled or identifying external information flow,

metaphors, and typical use cases. In this case study, understanding the
requirements is enough to start looking for initial classes.

10.2 CANDIDATE CLASSES

An often recommended method to find classes is to look for de facto problem
domain structure. In a piece of tangible equipment like a video recorder, the
separate physical parts which need to be dealt with then become obvious
candidates for abstraction. If we can identify the parts that have program
interfaces, this will give us an initial structure to work with.

A first attempt
From the requirements in the previous sections we can identify several groups of
program-interfaced components: motors, solenoids, heads, sensors, buttons,
display window, and remote controller. We can collect this information in a first
cluster chart, as shown in figure 10.6.
The first five candidates in the chart represent deferred classes, since the
corresponding hardware components will have different interfaces and therefore
need to modify the available operations. The button and keyboard classes, on the
other hand, will probably only need to return an input code. The cassette slot
also works somewhat like a load button, since when a cassette is inserted far
enough, this is detected by a sensor which will trigger the rest of the cassette
transport.
 CLASS SELECTION AND CLUSTERING 275

CLUSTER VCR_UNIT Part: 1/1

PURPOSE INDEXING
Controlling software for the Bonasonic keywords: video recorder, first
Multronic 2001 video recorder. candidate classes

Class / (Cluster) Description

MOTOR Device to control various movable parts in the VCR: cassette

transport, tape path, tape movement, drum rotation, capstan
shaft rotation.

SOLENOID Magnetic coil to attract or repel certain movable parts: reel

brakes, pinch roller.

HEAD Electromagnet for reading, writing, or erasing information on

the tape: video heads, audio head, eraser head.

SENSOR Detector reporting certain physical conditions: start or end of

tape, tape path position, cassette inserted, cassette down, tape
protected, signal sent from remote controller.

DISPLAY_WINDOW Device displaying visible information to user: current time

and channel, tape-in, mode of operation, winding position,
menu selection, timer dialog, clock dialog.

OPERATION_BUTTON Button pushed by user to operate the VCR: playback, record,

stop, pause, rewind, fast forward, eject, channel up or down.

CONTROL_BUTTON Button for selecting options: enter menu, program recording,

set clock.

KEYBOARD Set of buttons for entering numeric input: 0−9.

LOAD_SLOT Slot for entering tape cassette.

TIMER Hardware clock that can deliver interrupts at certain preset

points in time.

TUNER AFC device that can search for tunable stations and switch
between a number of frequencies stored in memory.

FRONT_PANEL VCR front containing buttons, display, and cassette slot.

REMOTE_CONTROLLER Device transmitting an infra-red light signal when one of its

buttons is pushed by the user.

MCU Micro computer unit housing the control program.

Figure 10.6 First candidate classes

10.3 CLASS SELECTION AND CLUSTERING

In the next task, we proceed to select analysis classes from the candidates, then
classify and possibly group them into a cluster structure. The deferred
candidates, except DISPLAY_WINDOW, each have a distinct set of candidate
276 A VIDEO RECORDER

subclasses identifiable from the requirements. The corresponding set of classes

can be grouped into clusters, as shown in figure 10.7.
The subclasses to DISPLAY_WINDOW cannot be inferred from the text, since
we do not know what distinct LED matrixes or other display elements the front
panel display window consists of. Several types of displayed information may

* * *
MOTOR SENSOR HEAD

LOAD_MOTOR TAPE_EXTREME VIDEO_HEAD

REEL_MOTOR TAPE_LOADED AUDIO_HEAD

DRUM_MOTOR TAPE_SYNCPOS ERASER_HEAD

CASSETTE_ HEADS
MOTOR TAPE_TENSION

CAPSTAN_
MOTOR TAPE_PROTECTED *
SOLENOID

MOTORS
CASSETTE_IN

REEL_BRAKE
CASSETTE_DOWN

PINCH_ROLLER
IR_SIGNAL

SENSORS SOLENOIDS

Figure 10.7 Similarity grouping

share the same physical components, while more complex information (for
example the menu dialog) probably needs to use several physical components.
The LOAD_SLOT class is not needed, since the sensor CASSETTE_IN will
detect the user inserting a tape. The remote controller has a set of operational
buttons, while the front panel also has a set of control buttons to set options.
This can be expressed with aggregation relations. Grouping the button classes
into a BUTTONS cluster and then enclosing what we have plus the TIMER and
TUNER classes in a VCR_UNIT cluster produces the architecture shown in
 CLASS SELECTION AND CLUSTERING 277

figure 10.8. The VCR_UNIT comprises the hardware components which are
used by the MCU to control the home video system.

Assessing the architecture

Now it is time to take a look at the analysis classes arrived at so far, and try to
assess their potential reusability with respect to future changes. We seem to
have encapsulated the current hardware components fairly well, so that low-level
interface details may be hidden from the upper layers of the MCU logic. Instead
of deciding directly what particular IC pins should go high or low to start motors,
set video signal directions, apply brakes, etc., higher-level operations can be
applied to abstractions of the components.
However, if we start thinking of what operations the selected classes should
have, we realize that we are in trouble. The class MOTOR should probably have
a stop operation, but that is just about all we can say at this level of generality.
Some motors have only two modes, constant speed in one direction or off, others
have several speeds and may reverse the direction. The speed might be fine
tuned by varying the input voltage, or by applying brakes to increase the
resistance. Motors may be mechanically combined to perform several functions
as one unit. There are even recorders that use only one motor to replace all five
in our example.
Moreover, motors that need exact speed control, like drum motors, reel
motors, and capstan motors, are usually controlled by servo systems in separate

VCR_UNIT

REMOTE_ MOTORS
CONTROLLER
*
MOTOR
SET […]
{

HEADS
OPERATION_ TUNER *
BUTTON HEAD

MCU
CONTROL_ { FRONT_ SOLENOIDS
BUTTON PANEL
*
SOLENOID

KEYBOARD TIMER
SENSORS
*
BUTTONS SENSOR

Figure 10.8 First static architecture

278 A VIDEO RECORDER

ICs. Similar variability holds for tape heads, sensors, and solenoids. Also, the
BUTTONS cluster, REMOTE_CONTROLLER, and FRONT_PANEL classes
mirror the physical structure of the recorder, but since buttons represent
extremely simple abstractions, the chosen classification will probably not help
system maintenance very much.
Although our classes represent general components that are widely used in
many video recorders, they are (with the exception of TIMER and TUNER) either
not very interesting or else too special to use at the earliest analysis level. We
should look for something more general.

A second attempt
So we forget about specific hardware components for the moment, and look
instead at the main services offered by a video home system (VHS). Our aim is
to avoid the conceptual straitjacket resulting from choosing too special
abstractions at the highest level, but still impose some useful structure that can
be reused in future versions of the Multronic model and perhaps also in other
Bonasonic VCR products.
It seems reasonable to separate two parts: one for magnetic signal handling
and one for the mechanical transportation of tape and heads. We call these
classes VIDEO_CONTROLLER and TAPE_CONTROLLER respectively. The
TIMER and TUNER classes of our previous attempt also seem general enough
for most video systems.
Another general service, which will be more and more important for future
models, is user control of options. All modern video systems offer special
functions like programmable recordings, programmable tape editing, simulated
stereo on playback, audio dubbing, child locking, and so on.
Using these functions requires an interface that is somewhat more complicated
than just pressing one button. On the other hand, since we have no direct
pointing device or fully fledged alphanumeric keyboard (too space consuming to
be useful), we need some simple consistent metaphors to let the user input all
required information without too much difficulty.
The Multronic interface has simple menus with some navigational buttons and
a numeric keyboard as the physical interface for option control. We can use this
to create a general concept of sequential menu containing a set of entries and
some predefined commands to shift the focus between entries, select a submenu,
input a value to an entry, leave the menu, etc. This is captured by the deferred
class MENU, and the effective subclasses in the Multronic system are called
OPTIONS, PROGRAM, and CLOCK respectively. The new static architecture is
shown in figure 10.9. As we can see, it is entirely different from our first attempt
in figure 10.8. The corresponding cluster charts are found in figure 10.10.
 CLASS SELECTION AND CLUSTERING 279

VCR_UNIT

*
MENU

VIDEO_
TUNER
CONTROLLER
OPTIONS

OPERATION TIMER
PROGRAM

TAPE_ CLOCK PRESET_

CONTROLLER RECORDING

MENUS

Figure 10.9 Second static architecture

10.4 CLASS DEFINITION

Having selected and grouped the analysis classes, we now proceed to look at
their operations. The class chart for OPERATION is shown in figure 10.11; it
has one command for each operational button on the front panel.
The corresponding buttons at the remote controller lead to codes transmitted to
the IR sensor. We do not model any button classes, but instead let the
EVENT_HANDLER class take care of all panel button events and signals from
the timer, tuner, and sensors, and direct them to the proper objects. This scheme
therefore includes the remote controller, whose trivial program is not modeled.
Each button may yield a unique code, or perhaps a group code coupled with a
position number within the group, depending on the hardware component used.
The event handler may use tables or other means to trigger corresponding
operations in the VCR_UNIT classes (FSM modeling is a typical candidate here).
We do not care about the details, but prefer to leave things open not to preclude
different strategies.
The interface of the class OPERATION is very close to the user metaphor
about what goes on in the VCR. No particular hardware components are
assumed at this level. There will probably be more operations to take care of
various sensor events, but since we do not know how much of this will be
handled by local ICs and what will have to be dealt with by the micro processor,
they are left out until future detailed design.
280 A VIDEO RECORDER

CLUSTER VCR_UNIT Part: 1/1

PURPOSE INDEXING
Controlling software for the Bonasonic keywords: video controller classes
Multronic 2001 video recorder.

Class / (Cluster) Description

OPERATION General operation of the video recorder.

VIDEO_CONTROLLER Operates the magnetic signal handling part of the recorder,

such as directing the heads to amplify the signal from tape or
from input channel depending on recording or playback.

TAPE_CONTROLLER Operates all movable mechanical parts to position the tape,

move the cassette, rotate the drum and capstan, etc.

TIMER Keeps track of programmed recordings and returns a signal

when the hardware clock reaches preset times.

PRESET_RECORDING Stores one programmed recording with channel number and

start / stop time.

TUNER Searches for receivable stations and stores their frequencies

for easy selection.

MENU Deferred class containing a sequence of entries and a number

of standard commands to move between entries, select
alternatives, input values, etc.

EVENT_HANDLER Handles all external events, like interrupts from sensors,

clock interrupts, buttons pressed on front panel, or signal
received from remote controller. Directs each event to the
object responsible for taking care of it.

(MENUS) Subcluster containing classes inheriting from class MENU.

CLUSTER MENUS Part: 1/1

PURPOSE INDEXING
Different types of menu classes keywords: video recorder, user menus
encapsulating user selections.

Class / (Cluster) Description

OPTIONS Menu for tuning stations and selecting submenus to set clock
or program recordings.

PROGRAM Menu interface for programming.

CLOCK Menu interface to reset system clock.

Figure 10.10 Analysis clusters, second attempt

 CLASS DEFINITION 281

CLASS OPERATION Part: 1/1

TYPE OF OBJECT INDEXING

Main VCR operation. cluster: VCR_UNIT
keywords: video recorder, main operation

Commands Load cassette. Eject cassette.

Playback. Record. Stop. Pause.
Fast forward. Rewind. Forward cue. Reverse cue.
Channel_up. Channel_down. Input a digit.

Figure 10.11 Main operation

The display details in the requirements text are also considered too special for
the analysis model, and are left until we know more about the hardware display
elements to use. The cueing operations forward cue and reverse cue stand for
medium speed tape transport, allowing a user to skip over sections of a recording
while still viewing its contents (rapidly moving frames, usually with significant
distortion). The next charts are the controller classes in figure 10.12. These

CLASS VIDEO_CONTROLLER Part: 1/1

TYPE OF OBJECT INDEXING

Handler of the magnetic part of a VCR. cluster: VCR_UNIT
keywords: video recorder, magnetic signal

Queries Is VCR in playback mode? Is VCR in recording mode? Current channel

Commands Playback. Record. Set channel.

Constraints Memory for 32 channels.

CLASS TAPE_CONTROLLER Part: 1/1

TYPE OF OBJECT INDEXING

Handler of the mechanical part of a VCR. cluster: VCR_UNIT
keywords: video recorder, mechanical part

Queries Tape in read/write position? Tape in sync position? Tape released?

Cassette loaded? Tape stopped? Tape normal forward? Tape fast forward?
Tape rewinding? Tape cueing forward? Tape cueing reversed?

Commands Move tape to read/write position. Move tape to sync position.

Release tape. Load cassette. Stop tape. Eject cassette.
Run tape normal forward. Run tape fast forward. Rewind tape.
Cue tape forward. Cue tape reversed.

Constraints Move to read/write or sync position requires loaded cassette.

Inserted cassette which has not reached bottom in 5 seconds is ejected.

Figure 10.12 Subsystem controllers

282 A VIDEO RECORDER

classes have boolean queries reporting on the various modes of the mechanical
and magnetic subsystems of the VCR. Since the operational buttons all
correspond to potential mode changes, the commands and queries come in pairs:
to enter a mode and to check whether a mode is in effect or not.
Next, we turn to the menu classes. We choose a number of standard
operations encapsulated in the abstract class MENU. The names of the
operations are fixed, but they need to cover a fair amount of variable behavior,
so we keep their semantics flexible. Each of them is defined for each specific
menu type to produce the desired behavior. The class chart for MENU is shown
in figure 10.13.

CLASS MENU Part: 1/1

TYPE OF OBJECT INDEXING

Menu navigation and data entry. cluster: VCR_UNIT
keywords: video recorder, abstract menu

Queries Is this menu open?

Commands Open this menu. Shift focus to next entry. Shift focus to previous entry.
Select. Accept. Cancel. Delete. Input numeric value.

Figure 10.13 Menu class with standard operations

The abstract menu keeps a list of entries, and the commands to open the menu
and shift focus between its entries should be defined to display whatever user
feedback is suitable. The class structure is independent of what type of hardware
display is used, if any. Select, accept, cancel, and delete are standard commands
that often make sense in menus, but do not always have to be used.
For example, in the options menu of the Multronic, select means tuning a
station or opening the clock or one of the program positions, depending on
current entry, accept means storing the last tuned frequency if the current entry is
tune, cancel always means leaving the menu and returning to normal operation,
while delete means clearing the current programmed recording or last stored
frequency.
In the clock setting and programming submenus of the Multronic, the entries
are fields in the user input record, and accept means accepting a completed value
and resetting the clock or program position. To facilitate class definition, we
could implement the commands as no-operations rather than deferred in MENU,
so subclasses only need to define those actually used. The free implementation
of the commands combined with the mapping from the event handler into
desired class operations can cope with quite a few variations of user input in
future Bonasonic models. The remaining classes, PRESET_RECORDING,
 CLASS DEFINITION 283

TIMER, and TUNER, are straightforward, and the corresponding class charts are
not shown. We will return to them briefly in task 6 on formal definition.

10.5 SYSTEM BEHAVIOR

With the classes defined, the next task is to describe system behavior in terms of
communicating objects. The primitive events are easy to identify, since they are
all related to very distinct buttons. We group them into categories, as is shown
in the event chart in figure 10.14. We then select some scenarios to describe
with object diagrams. There are not that many interesting variants of behavior
among the top-level analysis classes in our model, so three scenarios seems
enough. The corresponding scenario chart can be found in figure 10.15.

EVENTS VCR_SYSTEM Part: 1/1

COMMENT INDEXING
External events are either from the user or keywords: video recorder, multronic 2001
from the timer.

External (incoming) Involved object types

User inserts or ejects a cassette. EVENT_HANDLER, OPERATION,

TAPE_CONTROLLER

User presses a tape operation EVENT_HANDLER, OPERATION,

button: play, record, stop, pause, VIDEO_CONTROLLER, TAPE_CONTROLLER
forward cue, reverse cue, rewind,
or fast forward.

User switches channel: channel EVENT_HANDLER, OPERATION,

up, channel down, or numeric VIDEO_CONTROLLER
input in normal mode.

User presses menu in normal EVENT_HANDLER, OPERATION, OPTIONS

mode.

User presses a control button in EVENT_HANDLER, OPTIONS, CLOCK, PROGRAM

menu mode: menu, next, prev,
select, accept, cancel.

User presses a numeric button in EVENT_HANDLER, OPTIONS, CLOCK, PROGRAM

menu mode.

User presses a button on the EVENT_HANDLER, OPERATION,

remote controller. VIDEO_CONTROLLER, TAPE_CONTROLLER

A programmed start or stop time is TIMER, PRESET_PROGRAMMING, OPERATION,

reached. VIDEO_CONTROLLER, TAPE_CONTROLLER

Figure 10.14 Significant event types

284 A VIDEO RECORDER

SCENARIOS VCR_SYSTEM Part: 1/1

COMMENT INDEXING
Typical behavior triggered by user and/or keywords: video recorder, multronic 2001
built-in timer.

Program a recording:
User enters the options menu, selects a program position, and sets time interval and station
to record.

Execute programmed recording:

A preset time is reached and a station is recorded for the chosen time interval.

Find recording, play, and rewind:

User inserts a cassette, finds a program and plays it, then rewinds the tape and restores the
cassette.

Figure 10.15 Three scenarios

Any button pressed is caught by the EVENT_HANDLER class, which will

direct the event to the proper object depending on button and VCR mode. Some
user actions have no effect in certain modes, such as pressing ACCEPT in normal
(non-menu) mode or pressing REC when the recorder is already recording.
Whether this is detected in the event handler or in the corresponding controller
object is kept open. The object diagram for the first scenario is shown in
figure 10.16. We do not model exactly how numeric values are input, since
various shortcuts may be applied to reduce the number of keystrokes.
The second scenario could be thought of as triggered by an internal event, but
we prefer to view the event as external (coming from the hardware clock). We
assume the clock interrupt is caught directly by the TIMER class without passing

OPERATION
1
Scenario 1: Program a recording EVENT_ 5, 6
1 User presses MENU button HANDLER
2 Options menu is opened 3
3 User selects a program position 2 4
4 A program record is opened OPTIONS PROGRAM
5 User inputs numeric values
7
6 User presses ACCEPT
7 Completed record is passed to timer TIMER
8 Timer stores program data
8
PRESET_
RECORDING

Figure 10.16 First scenario

 SYSTEM BEHAVIOR 285

PRESET_
RECORDING
Scenario 2: Execute a programmed recording
1−3 The programmed channel is selected 1
4 Timer tells controller to start recording
5 Tape controller told to move to read/write TIMER
6 Tape controller told to run tape normal forward
2, 4, 8
7 Video controller told to record
8−9 Recording is stopped after programmed interval OPERATION
9

5, 6 3, 7

TAPE_ VIDEO_
CONTROLLER CONTROLLER

Figure 10.17 Second scenario

through EVENT_HANDLER. The resulting object diagram is in figure 10.17.

Finally, the third scenario is shown in figure 10.18. When the tape is stopped,
both controllers may or may not be involved depending on whether the magnetic
signal should be affected. Therefore, we put message label 11 before the fork
point in the diagram. This concludes the dynamic modeling, and we turn again
to the static architecture to look in more detail at the class features.

Scenario 3: Find recording, play, and rewind

1−2 User inserts cassette
3 User presses FF
4 Tape controller told to move to sync
5 Tape controller told to run tape fast forward EVENT_
6 User presses PLAY HANDLER
7 Tape controller told to run normal forward
8 Tape controller told to move to read/write 1, 3, 6, 10, 13
9 Video controller told to record
10 User presses STOP OPERATION
11 Controllers told to stop
11
12 Tape controller told to release tape
13−14 User ejects cassette
2, 4, 5, 7, 8, 12, 14 9

TAPE_ VIDEO_
CONTROLLER CONTROLLER

Figure 10.18 Third scenario

286 A VIDEO RECORDER

10.6 FORMAL CLASS DESCRIPTION

What remains is the last analysis task: to translate the class charts into more
formal class interface descriptions. Since we do not want to specify too many
details for the VCR at the analysis level, there are not many contracting clauses
to fill in. They will come later during detailed design and implementation. The
classes for programming and tuning are shown in figure 10.19.

PRESET_RECORDING

channel: INTEGER
CLOCK + start, stop: TIME
Invariant
Inherits: MENU
channel ∈ {1..32}
MENU *

is_open: BOOLEAN
− − Is menu open?
TIMER
open
forth PROGRAM + store_program
back – PRESET_RECORDING
Inherits: MENU
select delete_program
– INTEGER
accept
? n ∈ {1..8}
cancel
delete
input
– INTEGER
OPTIONS + TUNER

Inherits: MENU tune

store

Figure 10.19 Menus, timer, and tuner

Figure 10.20 shows the interfaces of the top-level operational class and the
two controller classes. We see that at this general level, there is not (yet) much
more information in the formal class interfaces, as compared to the
corresponding class charts. However, they are much more compact, allowing
more comprehensive views of groups of classes with some of their relations.
We show in figure 10.20 that class OPERATION will be a client of the two
controller classes, but we do not know at this stage what features will cause the
corresponding relations.
 FORMAL CLASS DESCRIPTION 287

OPERATION TAPE_CONTROLLER

load_cassette −− Queries
eject is_rw: BOOLEAN
play − − Is tape in read/write position?

record is_sync: BOOLEAN

− − Is tape in audio sync position?
stop
pause is_released: BOOLEAN
− − Is tape ejectable?
fast_forward
is_loaded: BOOLEAN
rewind − − Is cassette loaded?
forward_cue is_stop: BOOLEAN
reverse_cue − − Is tape in stop position?

channel_up is_fwd: BOOLEAN

channel_down − − Is tape running normal forward?

input is_fast_fwd: BOOLEAN

– INTEGER − − Is tape running fast forward?

is_rewind: BOOLEAN
− − Is tape rewinding?

is_fwd_cue: BOOLEAN
− − Is tape cueing forward?

is_rev_cue: BOOLEAN
− − Is tape cueing reversed?
VIDEO_CONTROLLER
−− Commands
is_play: BOOLEAN move_to_rw
− − Is video in playback mode? move_to_sync
is_rec: BOOLEAN release_tape
− − Is video in recording mode?
load
channel: VALUE eject
− − Current channel
fwd
play
fast_fwd
record
rewind
set_channel
– INTEGER fwd_cue
Invariant rev_cue
¬ (is_play and is_rec) Invariant
channel ∈ {1..32} is_rw xor is_sync xor is_released

Figure 10.20 Controller classes

288 A VIDEO RECORDER

10.7 FINAL STATIC ARCHITECTURE

This concludes the analysis part of the BON process. To proceed further with
this case study we would need more information on the hardware used for the
Multronic model. Therefore, we will stop here, and the final static structure of
our VCR system is shown in figure 10.21.

VCR_UNIT

* TIME
MENU

PRESET_ VIDEO_
RECORDING CONTROLLER
PROGRAM

TUNER OPTIONS TIMER OPERATION

CLOCK
TAPE_
CONTROLLER

MENUS

EVENT_HANDLER

Figure 10.21 Final static architecture

Comparing this architecture with our first analysis model in figure 10.8 gives
us a clear illustration of the important insight that there is no objective reality.
Models only exist in relation to what you want to do with them. Useful as
tangible objects may be for providing a starting point for building a system
model, they must always be treated with suspicion precisely because they are
tangible and therefore in most cases somewhat special. What tend to survive in
the long term are often the more abstract underlying ideas.
11 Relational and object-
oriented coexistence

How should one model persistent objects and relationships between them in an
object-oriented system? This question often arises in application domains where
information systems play a central role. As argued in chapter 2, a strong case
can be made against entity−relationship modeling and its variations in object-
oriented contexts, since it breaks the inherent seamlessness of the approach.
However, even if object-oriented technology is now rapidly moving towards
commercial acceptance on a broad scale, relational databases will most likely
continue to play an important role as data repositories for a long time yet,
including many object-oriented applications.
There are several reasons for this. First of all, statistics have proven that the
average lifetime of stored data is far greater than the average lifetime of
applications handling the data. Thus while applications are being modified and
replaced, corporate data, although extended and updated, tends to remain where
it sits. Therefore, many information systems have grown extremely large and the
cost of a complete data conversion may not always be justifiable.
Moreover, databases are often accessed and manipulated by many different
applications in heterogeneous environments (often geographically distributed),
and it may not be worthwhile to rewrite all of these applications to comply with a
different database organization. Other reasons may have to do with company
policies, previous investment in database software and expertise, performance
requirements (transaction processing, concurrent updates, average uptime), and
data security (consistency controls, recovery/rollback, authorization).
The conclusion is that bridges are often needed between the relational and
object-oriented worlds. The purpose of this last case study is to discuss how
object models and relational models can be made to coexist in a system. The
approaches illustrated are drawn from actual working implementations, but since
a full discussion could easily fill a book of its own, they have been considerably
simplified.

289
290 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

11.1 FROM DATA STORAGE TO OBJECT PERSISTENCE

Let us first recall what is needed by an object-oriented execution model. Object-
oriented applications handling massive quantities of data often end up using and
creating large numbers of objects. Most existing object-oriented environments
still run on top of operating systems that do not support the basic run-time
requirements of an object-oriented approach: built-in object allocation, automatic
reclamation of unreachable objects (garbage collection) regardless of physical
location, and transparent paging at the object level of both transient and
persistent objects. Until basic facilities like these become widely available as
part of standardized families of operating systems, each object-oriented
environment needs to implement all or part of them as a separate virtual machine
(run-time system).
Not to burden applications with implementation details, object-oriented
environments must offer powerful means to handle both transient and persistent
objects. Large quantities of transient objects can often be taken care of by
traditional garbage collection in combination with virtual memory management
at the operating system level. Small amounts of persistent objects, in turn, can
be encapsulated by the run-time system using database library classes interacting
with an ordinary file system. However, to handle a potentially very large
number of persistent objects, more complete database capabilities are usually
required. A mapping is needed whenever object-oriented applications interface
with legacy systems or with databases that do not interoperate with an object-
oriented environment.
Interoperability in this context means more than just a basic coupling. Owing
to the high level of object integration required (object type and format,
polymorphism) distribution transparency of objects with implicit access via the
development environment is rapidly becoming an important issue. Some
technical approaches addressing the problems involved are beginning to emerge,
whose aim is to support the notion of an “object bus” and connect heterogeneous
object-oriented applications directly at the object level [OMG 1991].
Seamlessness between execution model and persistent data means that values
and types directly map the class instances used by the execution model; there is
no “impedance mismatch” between primary memory and disk-resident data. An
ideal solution completely frees client applications from storage details and
permits virtual addressing of an infinite object space. In such cases, the object-
oriented run-time system transparently pages in or out clusters of objects
according to their status and behavior in the application: frequency of access,
reachability, expected lifetime, and so forth.
Consider the following program fragment from a developer’s standpoint. The
declaration and qualified call translate at execution time into: “apply routine
 FROM DATA STORAGE TO OBJECT PERSISTENCE 291

register defined and exported by class CUSTOMER to an instance of

CUSTOMER (or one of its descendant classes) attached to an attribute called
attendee.”
attendee: CUSTOMER
..
.
attendee.register
In a fully transparent persistent environment, the above is complemented by:
“regardless of the effective location of the object referred to by attendee at the
time the call is executed.”
An object is persistent if its existence is independent of the system session in
which it was created. A persistent object continues to exist until it either
becomes unreachable or is explicitly deleted. Various techniques can be
employed in object-oriented environments to make an object persistent.
Figure 11.1 depicts some possibilities.

1 customer := new persistent CLIENT

2 customer: persistent CUSTOMER

persistent class CUSTOMER

..
3 .
end
customer: CUSTOMER

class CUSTOMER
inherit PERSISTENT
..
4 .
end
customer: CUSTOMER

persistent_collection: PERSISTENT_UNIVERSE
customer: CUSTOMER
5 ..
.
persistent_collection.put (customer)

Figure 11.1 Various persistency schemes

The first three examples in figure 11.1 introduce specific language constructs:
(1) extended object creation mechanism, (2) extended type declaration of entities
referring to objects, and (3) extended class declaration mechanism. The last two
examples use predefined classes to achieve persistency: (4) all children to a
common ancestor become persistent, and (5) a persistent object container accepts
any object reference, and all objects inserted into the container automatically
become persistent.
292 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

Regardless of the specific mechanism used, we may adopt the following deep
persistency principle: all objects reachable through successive references from a
persistent object also become persistent. This ensures consistency of the system
state (class invariants). Unless the transitive closure of objects referred to by a
persistent object is also stored, some objects may become invalid.
Objects explicitly made persistent through some scheme like the ones in
figure 11.1 are sometimes called persistent roots (not to be confused with root
objects starting up system executions). All other objects may dynamically be or
not be persistent, depending on whether they can be reached from a persistent
root or not.
Persistency in BON is defined as a class property and persistent classes can be
marked as such by a special class header annotation (bullet). This is often of
interest during analysis, since figuring out what objects need to survive system
sessions may be a good way to increase problem understanding. However, it
would be too restrictive to require that only persistent objects can be instantiated
from a class marked as persistent.
There may be situations in a system where a temporary object needs to behave
exactly like a persistent one, and forcing the creation of two nearly identical
classes in such cases does not make much sense. Therefore, marking a class as
persistent in BON means that its objects are potentially persistent.
We conclude this section by stating two principles regarding persistency,
which are important for the seamlessness and reversibility of the BON approach.
The aim is to keep analysis and design models simple and consistent,
independently of where the objects will ultimately reside.

Principle 1
There should be no syntactic distinction between persistent and transient
data in terms of how they are defined and used.
Principle 2
The persistent object model should be designed as a seamless part of the
full static model.

With these preliminaries we are ready to take a look at the problems involved
when object persistency (or part of it) is to be based on an underlying relational
model. We will discuss an approach for achieving a high degree of transparency
with regard to object retrieval and update—in spite of the structural differences
between object models and relational databases. The focus will be on the
dynamic construction of queries to reduce as much as possible the static
dependency of applications on the actual database schema.
 OBJECT MODELS AND RELATIONAL MODELS 293

11.2 OBJECT MODELS AND RELATIONAL MODELS

Our objective is to design an integration layer that can access data in different
storage formats and automatically convert between them. This layer may also
have its own local repository, which may be used for caching to avoid constant
data transfer through a gateway. The kind of environment envisioned is
illustrated in figure 11.2.

Object
Integration Layer Repository
Application

SQL Data Objects

SQL Gateway

Relational Object
Database Database

Figure 11.2 Transparent integration of heterogeneous storage

Different forms of coupling are possible. At the highest level of integration,

persistent class instances are stored along with their features and class
descriptions. With a more pragmatic approach, only the data part (state
variables) of persistent objects is stored. In both cases, the database needs to be
closely integrated with the execution model.
However, when the logical structure of the available persistent storage is
totally unrelated to the object model (relational databases, flat files, indexed
files) a separate interface layer is needed to do the necessary transformations. In
figure 11.2 a SQL interpreter is used as backend to retrieve and store data in a
relational database.18
In either case, it should be possible to define application object models, where
the persistency decisions are kept completely free from any implementation
choice, and all application objects are accessed the same way whether
transparently constructed from external data or not.

18
The relational data language SQL was earlier named SEQUEL and is usually pronounced as
though it still were. We therefore write “a SQL…” rather than “an SQL…”.
294 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

Integrity constraints
A number of data integrity rules are usually enforced in a relational system to
prevent certain types of inconsistencies from entering the database. These rules,
commonly known as integrity constraints, address various aspects of the
semantic content of stored data. We will mention a few of them below, and see
how they translate to a BON object model.
Domain integrity refers to type checking between the values used in query
expressions and the declared types of the corresponding entities. All RDBMS
provide the necessary level of checking to avoid any violation of the type system
rules. Since BON is statically typed, it is assumed that the supporting
environment (CASE tool at the analysis and design level, and programming
system at the implementation level) will detect any type error.
Referential integrity has to do with the consistency of references between
schema elements. Whenever an entry in a relational table refers by a foreign key
value to an entry in another table, that other table must exist and have an entry
with matching primary key value. Any modification of the database content
must keep all related tables consistent and prevent the introduction of unmatched
references. These checks are usually supported at the RDBMS level. It is
assumed in BON that referential integrity is captured by assertions in class
descriptions. In the example given in figure 11.3, integrity is guaranteed by the
postconditions associated with the routines bid_farewell and retire.

COMPANY EMPLOYEE

staff: SET [EMPLOYEE] working_place: COMPANY

bid_farewell retire
– leaving_person: EMPLOYEE
! retired
! leaving_person ∉ staff;
leaving_person.retired retired: BOOLEAN

Figure 11.3 Referential integrity in class contracts

User-defined integrity is usually taken care of by stored procedures or triggers

in relational databases. With BON it is part of the object-oriented data
description. The creation routines of a class are responsible for ensuring that
each object of the class is created in a consistent state. Any operation changing
the state of the object will, if necessary, trigger other monitoring operations to
ensure that the consistency is maintained.
At execution time, depending on the supporting environment, integrity
violations may invoke exception handlers, rollback procedures, and other
recovery mechanisms.
 OBJECT MODELS AND RELATIONAL MODELS 295

We now turn to the design of an integration layer coupling an object model

with a SQL gateway. There are three aspects of such a layer that need to be
addressed:
• How to retrieve data from an existing relational database.
• How to regenerate application objects from the retrieved data.
• How to design a relational schema suitable to store and retrieve a given
persistent object model.
Our aim is to find a generic model to tackle the problems, which is as
independent as possible of both the object-oriented application and the relational
database system used.

11.3 A RELATIONAL DATABASE WRAPPER

Short overview of the relational model
In relational databases, information is modeled as relations defined on finite sets
of data values called domains. A relation is a set of ordered lists of data values
called tuples.
Each tuple is an element of the cartesian product of domains D1×D2×D3…
(set of all possible ordered lists of values with one element drawn from each
domain). Each occurrence of a domain in the definition of a relation is called an
attribute, and the same domain may occur several times. Note the difference
between domain and attribute: a domain is a basic pool of permissible values,
while an attribute represents the use of a domain within a relation.
Domains and relations are implemented as tables with m rows and n columns.
Each row corresponds to a tuple (an element in the set), and each column to a
relational attribute. For this reason, RDBMS vendors usually use the terms
table, column, and row instead of relation, attribute, and tuple.
The great majority of relational systems are normalized, which means that all
domain elements must be atomic values. All values in a given domain have the
same type chosen from a small set of predefined basic types, such as: INTEGER,
FLOAT, DOUBLE, DATE, CHAR, STRING, MONEY. Each attribute (or
column) of a relation has a name, so it can be referred to without using its
relative position, and a type, which is the basic type of the corresponding
domain.
Relations are usually defined with constraints imposed on the tuples to avoid
data duplication or cross-dependency. A common constraint on a relation is to
require that each tuple be uniquely identifiable by a subset of the attribute values.
Such a subset is called a primary key. Often one attribute is enough to identify
296 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

tuples, in which case we have a single-attribute primary key.

Access to relational data is achieved through general set operations: selection,
projection, product, join, union, intersection, and difference. In addition to
tables predefined in the relational schema, new tables may be created
dynamically through such operations. The data access set operations are
expressed in a relational database language called SQL (Structured Query
Language) used to store, retrieve, and modify information in the database. (SQL
has become the de facto standard in the relational database world, and was
accepted as an international standard by ISO in 1987. The latest ISO version,
SQL/92, became ratified in 1992 [Date 1993].)
In figure 11.4, three example tables are shown: CUSTOMER, INVOICE, and
PRODUCT. The header of each table shows the table name and the names of
each attribute. Below the double line are the tuples, whose values conform to the
basic type of each attribute (these types are not shown in the table).
The most frequently exercised operations in relational database applications
are usually simple selections of tuples whose attribute values satisfy certain
conditions, insertion or deletion of tuples, and change of attribute values in

CUSTOMER

Client_Id Name Address Zip_Code

A45 Jack’s Snack 899 Ventura Blvd, La Cienaga CA 92340

L20 Red Lobster 9B Nathaniel Hawthorne, Tauton MA 02780
Primary key: (Client_id)

INVOICE

Purchase_Order Product_number Qty Client_Id

940120-010 1022 500 A45

940322-093 1024 80 Y89

Primary key: (Purchase_Order, Product_number)

PRODUCT

Product_number Description Unit_Price

1022 Corned Beef 1.54 oz can 0.99

1023 Snails in garlic butter 0.8 oz bag 4.99
1024 Peeled tomatoes 1.9 oz bottle 1.99
Primary key: Product_number

Figure 11.4 Tables from a relational schema

 A RELATIONAL DATABASE WRAPPER 297

tuples. Selection is mostly combined with projection, which means that only a
subset of the attribute values are retrieved.
The join operation is important for more complex retrieval. For example,
assume we want a list of all customers who ordered products with a unit price of
at least five dollars. The result should be presented as a table with the following
attributes: client id, client name, client address, product description.
The combination of these attributes does not exist as a table per se in our
schema, but it is possible to join our three tables to obtain the requested
information. Using SQL syntax, the selection can be expressed as follows:
select CLIENT.client_id, name, address, description
from CLIENT, INVOICE, PRODUCT
where CLIENT.client_id = INVOICE.client_id and
INVOICE.product_number = PRODUCT.product_number and
PRODUCT.unit_price >= 5.0

The result of a selection query is generally a set. Therefore, SQL provides

facilities to iteratively fetch each matching row. A cursor maintained by the
database server points to the currently retrievable row, and standard operations
can be used to move the cursor from one row to another within the result.
The relational model for database management, originated in the late 1960s by
E. F. Codd, has a strong mathematical foundation [Codd 1970]. It has been
thoroughly researched and a large number of rules and criteria for relational data
organization and manipulation have been proposed [Codd 1985a, Codd 1985b,
Codd 1990]. For good comprehensive overviews of the area, see [Date 1990,
Date 1983, Date 1993].

Designing a database interface cluster

Any cluster layer interfacing a relational database and an object-oriented system
would be responsible for managing server sessions, maintaining the relational
schema, performing queries and updates, and doing the mapping between rows
and objects. To summarize this, let us define the major abstractions of such a
layer and group them as shown in the cluster chart of figure 11.5.
We can also display the classes in a first static architecture sketch as shown in
figure 11.6. The implementation of each model class will encapsulate a set of
external calls to the database server. The DATABASE_INTERFACE cluster is a
client of class ANY because any type of object may become persistent and thus
need to receive external data.
Our first aim is to outline the interface of a general reusable cluster for
accessing a relational database. The specification of this cluster is at a rather
technical level, so we will skip the class chart definitions and move directly to
the formal class descriptions.
298 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

CLUSTER DATABASE_INTERFACE Part: 1/1

PURPOSE INDEXING
Layer to make relational database keywords: object and relational
manipulations transparent. coexistence, rdbms interface

Class / (Cluster) Description

DB_SESSION Session manager responsible for handling the connection to

the database server and for tracking the proper or non-proper
completion of all transactions.

DB_QUERY SQL wrapper sending selection commands to the database

server.

DB_CHANGE SQL wrapper sending store, update, and delete commands to

the database server.

DB_RESULT Representation of one matching row returned by the database

server in response to a SQL selection.

Figure 11.5 First candidate classes in interface layer

▲ ▲
DB_SESSION DB_CHANGE

ANY
▲ cursor ▲
DB_QUERY DB_RESULT

DATABASE_INTERFACE

Figure 11.6 First cluster sketch

Class DB_SESSION encapsulates the most important primitives to handle

transactions between an application and the database server. Its interface is
outlined in figure 11.7.
Class DB_QUERY sends SQL selection queries to the database server and
stores the resulting table rows. Clients can then iterate on the result supplying a
callback routine to process the table rows, one by one. The interface description
of class DB_QUERY is shown in figure 11.8.
Typically, the callback object will be the client object itself, inheriting from
ACTION and defining the execute feature. Since the client already has a
reference to the DB_QUERY object on which the iteration was invoked, the
execute routine will be able to access the database cursor of the corresponding
selection.
 A RELATIONAL DATABASE WRAPPER 299

DB_SESSION ▲

connect
−− Connect application to database server.
! is_connected
disconnect
− − Disconnect application from database server.

! ¬ is_connected
commit
−− Update database with last modifications.
? is_connected
rollback
−− Backup to previous state.
? is_connected
is_connected: BOOLEAN
− − Is application connected to the database server?

transaction_status: VALUE
− − Status of last performed transaction

Figure 11.7 Encapsulation of a database session

Class DB_RESULT represents the database cursor pointing to the current table
row returned by the database server. It is responsible for the conversion of data
fields from the SQL structure on the server side into corresponding basic object
attributes that may be accessed in a normal way by the object model. Any of the
fields can thus be inspected, which gives clients full control to do whatever
processing is needed.
However, in many cases the main part of the action for each returned row will
be to transfer some or all of the data fields into the corresponding attributes of
some result object. Therefore, the load_object command of DB_RESULT (see
figure 11.8) will automatically convert and load data from the fields of the
current row into an object supplied as an argument.
Each basic attribute of the argument object that has a name and a type which
matches an attribute of the table row will be set to the corresponding value. An
object attribute is considered basic if its type corresponds directly to a type
defined in the relational database. The mapping from one type system to another
can be preset in a utility class and accessed when needed.
It is the client’s responsibility to ensure that each basic object attribute which
is to receive a value corresponds exactly (by name and type) to one of the table
300 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

DB_QUERY ▲ DB_RESULT ▲

query count: VALUE

−− Select table rows from database. − − Number of columns in row
– sql_selection: STRING attribute_value: ANY
iterate_on_result − − Value at index
− − Start an iteration on resulting rows. – index: INTEGER
over: BOOLEAN cursor attribute_name: STRING
− − More rows to iterate on? − − Name at index

! Result ↔ (row_action = ∅ or – index: INTEGER

cursor = ∅ or row_action.over) attribute_index: VALUE
− − Index of attribute name
cursor: DB_RESULT – name: STRING
− − Current resulting table row
load_object
count: VALUE − − Load object from current row.
− − Number of rows in result
– receiver: ANY
row_action: ACTION
− − Object implementing an execute
− − routine for each iteration

set_action
DB_CHANGE ▲
– action: ACTION
! row_action = action
modify
Insert, update, or delete
−−
row_action information in database.
−−
– sql_statement: STRING
ACTION *

over: BOOLEAN
− − Stop calling execute?

execute *

Figure 11.8 Encapsulation of relational storage and retrieval

row attributes. Furthermore, the names and types of the attributes of an object
must be dynamically accessible at execution time for the automatic data transfer
to work. In object-oriented environments where this information is not available,
a corresponding table (preferably generated from the corresponding class
description by some tool) may have to be attached to each persistent class.
Note that the automatic loading does not require all basic attributes of the
receiving object to match columns in the table row, nor all columns to
correspond to an object attribute. Only basic attributes with matching name and
type will be transferred. This convention has two advantages:
 A RELATIONAL DATABASE WRAPPER 301

• Some basic object attributes may be left out of a query if they are
considered uninteresting in some context (perhaps given default values).
• Several objects of different type may be loaded, one at a time, from the
same query result. This will be important for our design of the higher-
level layers.
Class DB_CHANGE, finally, is simply used to pass a SQL statement requesting
an update, deletion, or insertion in the relational database.

Mapping non-basic object attributes

So far, we have only discussed storage and retrieval of basic object attributes.
Usually, however, the attributes of most objects in an object model will be a
mixture of basic and non-basic types. If we look at the class REGISTRATION
from the conference case study, whose interface is repeated in figure 11.9, we
find two attributes that relate to other model classes, namely PERSON and
TUTORIAL.

REGISTRATION ●

attendee: PERSON
registered_at: DATE
amount_paid: VALUE
invoice_sent: BOOLEAN
confirmed: BOOLEAN
paper_sessions: BOOLEAN
selected_tutorials: SET [TUTORIAL]

Figure 11.9 A persistent class description

These attributes represent object references which cannot be mapped directly

to relational attributes. (In this case study, we will assume that all typed features
of persistent model objects represent class attributes rather than functions, unless
otherwise stated.)
However, we can use the automatic data transfer previously described to have
the basic attributes of a REGISTRATION object (registered_at, amount_paid,
invoice_sent, confirmed, and paper_sessions) automatically initialized from the
database rows. In an eventual implementation, general types such as VALUE in
figure 11.9 will have been specialized to a basic type of the programming
language making the correspondence to the relational types clear. A possible
302 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

mapping could be the one in figure 11.10.

The PERSON and TUTORIAL objects corresponding to attendee and
selected_tutorials can then be initialized separately with rows coming from other
relations, and the corresponding references filled in by the application object
responsible for restoring REGISTRATION objects.
Name Type
REGISTERED_AT DATE(MM/DD/YY,HH24.MI.SS)
AMOUNT_PAID FLOAT
INVOICE_SENT CHAR(1)
CONFIRMED CHAR(1)
PAPER_SESSIONS CHAR(1)

Figure 11.10 Relational table representing class REGISTRATION

When a persistent object structure is stored in the database, we must capture

the unique identity of each object through primary keys in the database tables.
Sometimes basic attributes already exist, whose values can be used to fully
identify each object. In our case, it might be possible to use registered_at as the
primary key, provided that time information of enough granularity is included
and that registrations are not entered simultaneously.
If such an attribute (or group of attributes) does not exist, we need to add an
extra field containing a unique identifier. Such an extra field may also be useful
for efficiency purposes when the existing primary key is long and many other
persistent objects will refer to this one.
Object reference attributes may be stored as fields pointing to other tables
representing the referenced objects (so-called foreign keys). Different strategies
may be used to represent references between model objects, depending on the
corresponding instance multiplicity.
• If the reference is one-to-one (as for attendee in class REGISTRATION
above), an extra field in either one of the tables REGISTRATION or
PERSON containing a primary key to the other is enough.
• If the reference is one-to-many (as for children in a class MOTHER), an
extra field in the table representing the “many” pointing to the table
representing the “one” will do.
• If the reference is many-to-many (as for selected_tutorials above), we may
need to represent the relation as a joint table TUTORIAL_ATTENDENCE
containing two primary keys: one pointing to table REGISTRATION and
one to TUTORIAL.
However, if the maximum number of instances referred to by either side
is low, another possibility is to map the reference into a fixed number of
extra fields containing either a zero reference or a primary key value.
 A RELATIONAL DATABASE WRAPPER 303

For example, if the maximum number of tutorials each person may

attend is four, a separate table can be avoided by adding attributes
tutorial1, tutorial2, tutorial3, and tutorial4 to table REGISTRATION.
There are also various strategies for mapping inheritance relations to relational
database schemas. For overviews, see [Rahayo 1993, Premerlani 1994,
Blaha 1994].

A scenario
We conclude this section with a scenario illustrating how an application may use
the query facility with automatic data transfer. The dynamic diagram is shown
in figure 11.11.

10
REGISTRATION

4
1 9
DB_SESSION CLIENT_OBJECT DB_RESULT

7 2, 3, 5, 8 6
DB_QUERY

Scenario: Query database and load result

1−3 A client object starts a database session, creates a query
object, and invokes a SQL query on it.
4−7 The client creates a new registration and starts an iteration on
the query result, supplying itself as action object. The query
object resets the cursor to point to the next resulting row, and
invokes the client action routine.
8−10 The action routine obtains the cursor object and tells it to load
the registration object from the table row.

Figure 11.11 Scenario showing access and automatic load

11.4 INTERFACING AN EXISTING RELATIONAL SCHEMA

Another important issue for the database encapsulation is the structure of the
relational schema (what precise tables and attributes we need to access). Various
factors affect the design of such a schema.
304 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

Factors determining the schema

Basically, there are two situations regarding choice of data organization:
• The relational schema mapping the object model will not be used for any
other purpose.
• The relational schema is also used by other applications that do not
necessarily have exactly the same view of the world.
In the first case, the database can be used as an implementation vehicle with
relational tables replacing flat files, indexed files, or any other data storage
facility. This situation leaves a great deal of freedom in designing the relational
schema, and data administrators may capture the object model in a way that best
fits the application. The goal is then to find a suitable tradeoff between a
mapping giving good performance and one that is easy to understand and
maintain.
In the second case, many additional factors must be taken into account. Often
the schema and associated database (or a significant part of it) is a legacy to our
system, which we may not be able to do anything about. The way attributes are
distributed in the object model may then be very different from the organization
of the corresponding data on the relational side.
This may result in SQL queries performing complicated joins across numerous
scattered tables, which may degrade performance. Often the solution is to
perform a number of pre-joins on selected tables and store them temporarily
during an object-oriented session. The initial cost at session start is then offset
by the improved performance during data transfer between the models.
With a growing number of heterogeneous applications sharing persistent
information across networks, the ability to adapt to existing structures is
important. In fact, even if the schema can be optimally tailored to an object
model (no initial legacy), the database organization tends to become much more
rigid once it has been filled with large amounts of data.
As time passes, the object model will gradually change (perhaps even faster
than with traditional models, because of the inherent flexibility of the object-
oriented approach). The impedance mismatch may then make it too expensive to
continually modify the relational structures to keep up with the changes. This
means we may have to face the legacy situation soon enough, even in cases
where the only applications ever accessing the data are object oriented!
The rest of this case study will be devoted to a detailed discussion of how
existing schemas can be mapped to model objects. We will design a set of
clusters of reusable classes enabling applications to become independent of the
exact database organization, and show how this cluster fits in with the general
database encapsulation presented in the previous section.
 INTERFACING AN EXISTING RELATIONAL SCHEMA 305

We will see how the design can be done gradually in layers raising the level of
abstraction to overcome the structural differences between the relational and
object models.

Schema dependency
If a typed implementation language is used, making static changes in the object
model implies recompilation of the application. This is reasonable, and usually
corresponds to a new version of the software. However, updates of the relational
schema in a database shared by many applications may occur frequently (new
columns added to tables, new tables added, minor reorganizations for efficiency).
A solution which forces recompilation and reinstallation of an application each
time a schema change occurs is therefore too rigid in most cases.
For this reason, we should strive to keep our applications free from exact
knowledge of the mapping between the object model and the relational database.
Rather than placing complete information directly in the static class structure
about the names of each database column accessed and the table it resides in, the
mapping should be dynamically reconfigurable by modification of data in some
repository. But how can we obtain adequate performance without integrating the
relational structure in our object-oriented applications?

A virtual database
One solution is to define a virtual database containing a set of virtual tables, and
then make applications statically dependent only on this database. (Such virtual
tables are known as views in RDBMS terminology.) If the virtual database is
chosen reasonably close to the real database, the conversion between the two
schemas will be straightforward, and can be effected by SQL statements
dynamically maintained as stored procedures or persistent strings. This gives
freedom to rename and create new tables in the database and to rename and
move around columns between them without changing the static structure of the
applications.
Regarding the logical representation of each column in the database, the
amount of freedom depends of course on the complexity of the mapping. If the
database stores temperature in degrees Fahrenheit and the object model uses the
Celsius scale, we cannot expect SQL to hide this fact. Also, even if the SQL
dialect provided by the database server would allow expressions to retrieve two
database columns given_names and last_name and directly return a concatenated
attribute name to the object model, it will hardly be possible to do the reverse on
update.
Therefore, the logical structure of the real database must usually be mirrored
by the virtual database, and applications must be statically dependent on the
306 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

representation form chosen for each column interacted with. However,

independence of the exact tables in which the columns reside, as well as of any
table or column renaming, is still a great advantage.
Each virtual table is represented by a class encapsulating a set of basic
attributes. The virtual table classes (which collectively represent the database
visible to the application) will be named row classes. Instances of row classes
are called row objects and each row object will act as a gateway to the real
database.
Each virtual table is chosen so that there is a simple mapping between its
attributes and the attributes of the real database tables. The persistency classes
will encapsulate operations to do the conversion using a SQL database server.
However, the corresponding SQL statements will be maintained outside the
application to always reflect the current state of the real database schema. An
application can load the correct SQL mapping at startup time or, in case it needs
to run continuously for a long time, be triggered to reload any change that may
occur during execution.

An example application
To illustrate the above approach, let us select four of the persistent classes from
the conference case study (chapter 9). The corresponding class descriptions,
showing only the features which we assume will be implemented as attributes,
are repeated in figure 11.12.
We also assume there is a corporate database which is to be used for mapping
relevant parts of our object model. In the corporate database, we find four tables
containing information that can be used to represent the basic attributes of
classes PERSON and REGISTRATION. These are shown in figure 11.13. There
are no existing tables corresponding to classes TUTORIAL or PRESENTATION.
The CUSTOMER and AFFILIATION tables come from the company’s
general customer register and the INVOICE table from its accounting system.
The REGISTRATION table is assumed to have been designed as part of an older
system which handles conference registrations but not the technical program.
We also assume that at present there is nothing we can do to change the formats
of these tables. This represents a kind of legacy situation not uncommon in
practice.
Since some objects are more difficult than others to map to a relational
system, it may be an advantage to have a relational and object-oriented
persistency mix. In this case, we choose to store and retrieve PERSON and
REGISTRATION objects in the relational database, while TUTORIAL objects
will be stored using some object persistency mechanism provided in the
language environment.
 INTERFACING AN EXISTING RELATIONAL SCHEMA 307

OBJECT_MODEL

REGISTRATION ● PERSON ●

attendee: PERSON registration: REGISTRATION

registered_at: DATE name: VALUE
amount_paid: VALUE affiliation: VALUE
invoice_sent: BOOLEAN address: VALUE
confirmed: BOOLEAN postal_mail: VALUE
paper_sessions: BOOLEAN email: VALUE
selected_tutorials: SET [TUTORIAL] phone: VALUE
fax: VALUE

TUTORIAL ●

Inherits: PRESENTATION PRESENTATION *

capacity: VALUE code: VALUE

attendee_count: VALUE title: VALUE
prerequisite_level: VALUE authors: SET [PERSON]
track: VALUE status: STATUS
duration: DURATION speakers: SET [PERSON]

Figure 11.12 Simplified persistent object model

This strategy also fits well with general performance considerations. Since the
tutorial objects are relatively few and frequently accessed, they should remain in
main memory during system execution. Person and registration objects, on the
other hand, may occur in great numbers but only the ones currently being
processed need fast access.
The general access, manipulation, and update of our object model and
corresponding relational data can now be outlined as follows:
• Creation of new tutorial objects.
These objects must be present before any registrations can be accepted,
since choice of tutorials is part of the registration data. Before the
reference to speakers and authors is filled in, the CUSTOMER and
AFFILIATION tables are searched to check whether some of the persons
308 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

CUSTOMER REGISTRATION
Name Type Name Type
CUSTOMER_CODE INT PERSON_CODE INT
SALUTATION CHAR(4) ENTRY_DATE DATE
LAST_NAME CHAR(32) DISCOUNT_RATE FLOAT
FIRST_NAME CHAR(32) CONFERENCE CHAR(1)
MIDDLE_INITIAL CHAR(1) TUTORIAL1 CHAR(8)
COMPANY INT TUTORIAL2 CHAR(8)
EXTENSION CHAR(16) TUTORIAL3 CHAR(8)
POSITION INT TUTORIAL4 CHAR(8)
DEPARTMENT CHAR(32) CONFIRMATION DATE
INVOICE INT

AFFILIATION INVOICE
Name Type Name Type
COMPANY_CODE INT INVOICE_CODE INT
COMPANY_NAME CHAR(32) CUSTOMER INT
ACTIVITY INT ISSUED_DATE DATE
COMPANY_SIZE INT PAYMENT_DATE DATE
STREET CHAR(32) PAYMENT_TYPE CHAR(8)
BUILDING CHAR(32) AMOUNT_PAID FLOAT
ZIP_CODE CHAR(8) AMOUNT_RECEIVED FLOAT
CITY CHAR(32) VAT FLOAT
COUNTRY_CODE INT
COUNTRY_NAME CHAR(32)
FAX CHAR(16)
PHONE CHAR(16)
EMAIL CHAR(16)

Figure 11.13 Relational tables in an existing database

are already present in the corporate database. If this is the case, all
attribute values of the PERSON objects are initialized with the
corresponding values from the database. Persons not found in the database
will be created and initialized from the input data on the object model side.
• Creation of new registration objects.
These objects are created from registration input data and will refer to the
already defined tutorial objects. As above, assigning the reference
attendee will either retrieve an old PERSON object from the database, or
create a new object.
• Update relational database.
Database updates may be performed at regular intervals, or when
requested by an operator. Unless some personal data needs to be
 INTERFACING AN EXISTING RELATIONAL SCHEMA 309

corrected, the CUSTOMER and AFFILIATION tables will only be

updated if new persons have been entered in the object model. The
REGISTRATION table will be updated for each new registration and
when existing ones are modified. The latter occurs, for example, when a
letter of confirmation has been sent or a tutorial selection is changed.

A virtual database interface

We are now in a position to start putting things together and sketch a general
design for mapping persistent objects to relational systems whose tables are not
in direct accordance with the object structure. We will use the architecture
depicted in figure 11.14.

PERSISTENT_OBJECTS

● ● *
REGISTRATION PERSON PERSISTENT

model_object

STORAGE_MANAGEMENT
manager
*
REGISTRATION_ PERSON_ STORAGE_
MANAGER MANAGER MANAGER

VIRTUAL_RDB

REGISTRATION_ PERSON_ *
ROW ACTION
ROW ROW

RDB_INTERFACE

Figure 11.14 Persistent object management (outline)

The two row classes of the virtual database are defined in figure 11.15. They
are the virtual relational representation of the corresponding persistent objects.
A row class encapsulates the interface of the RDB cluster. It may be given a
SQL selection, in which case it will create a DB_QUERY object, attach itself to
it, and forward the query.
310 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

ROW REGISTRATION_ROW

session: DB_SESSION Inherits: ROW

db_query: DB_QUERY
person_code: INTEGER
cursor: DB_RESULT
entry_date: STRING
? db_query ≠ ∅
conference: STRING
! Result = db_query.cursor
tutorial1, tutorial2,
query tutorial3, tutorial4: STRING
– sql_selection: STRING confirmation: STRING
update issued_date: STRING
– sql_statement: STRING
amount_paid: REAL
insert
invoice: INTEGER
– sql_statement: STRING
iterate_on_result
over: BOOLEAN PERSON_ROW
! Result ↔
(db_query = ∅ or db_query.over) Inherits: ROW

load_from_cursor
customer_code: INTEGER
? db_query ≠ ∅ and cursor ≠ ∅
last_name: STRING
attach_to_query first_name: STRING
– holder: ROW
middle_initial: STRING
? holder ≠ ∅
company: INTEGER
! db_query = holder.db_query
company_name: STRING
set_action street: STRING
– action: ACTION
zip_code, city: STRING
? db_query ≠ ∅
country_name: STRING
! db_query.row_action = action fax, phone, email: STRING

Figure 11.15 Table row classes

The set_action and iterate_on_result commands can then be used to scan

through the resulting table rows and load_from_cursor will load the current row
object, so it can be further processed. The resulting rows must of course contain
all columns corresponding to the attributes of the row object, so it can be loaded.
This is the responsibility of the client passing the SQL query.
Each persistent class will have two other classes corresponding to it: a row
class representing the object in the virtual database, and a manager class to do
the conversion between the representations. Each manager will only know the
 INTERFACING AN EXISTING RELATIONAL SCHEMA 311

internal format of the row objects it is responsible for.

Comparing the columns of figure 11.13 with the row classes shows that the
attributes of PERSON_ROW correspond directly to a subset of the columns of
CUSTOMER and AFFILIATION, which can easily be obtained by join and
projection using SQL. The same is true for REGISTRATION_ROW with respect
to tables REGISTRATION and INVOICE. Note, however, that the clients of the
virtual database know only about the virtual tables, and not even the row class
knows about the scattering in the real database. There is no static trace of tables
CUSTOMER, AFFILIATION, and INVOICE.
Since we are using a corporate customer base to store personal data about all
conference attendees, one may ask what to do with participants who have no
company affiliation. Such questions are typical for legacy situations where
structures are reused for purposes slightly different from the initial intention. In
this case we assume that we can invent and store a special COMPANY_CODE
representing private participants and simply leave the corresponding
COMPANY_NAME column blank for these entries.
We will return in the next section to the issue of how queries can be formed in
the object model to retrieve persistent object structures. For now, let us just
assume that the proper SQL requests will at some point be supplied as an
argument to the query feature of the row object.
As already mentioned, it is perfectly legal for a query to return more columns
than what is needed to load a given row object. First, there may be a need for the
client (in its row action) to check certain data in the cursor structure that will not
be transferred to the row object. Second, if a suitable naming convention is used
so that the destination of each attribute can be inferred dynamically from its
name, a client can use the result of a single SQL query to fill several row objects,
one at a time, without the need for unique attribute names across the row classes.
Since non-matching attributes are ignored in each transfer, no conflicts will
occur. A standard solution would be to use the SQL renaming facility (SELECT
… AS) to give each resulting column a name related to the proper row object
attribute, regardless of what names are used in the database tables.
For example, if the data for both a REGISTRATION_ROW and a
PERSON_ROW were to be returned by the same SQL query, we could use
names “REGISTRATION_ROW$ENTRY_DATE” etc. for columns whose
destination is the former object and “PERSON_ROW$FIRST_NAME” etc. for
those aimed for the second object.
If the object to be loaded is of type NAME and has an attribute attr the
conversion routines would look for columns named “NAME$ATTR” and
transfer the corresponding value. (A SQL statement of type “SQL NAMES ARE
...$...” introducing some character not used for other purposes would ensure
unambiguous interpretation.)
312 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

11.5 QUERYING A PERSISTENT OBJECT MODEL

Besides a basic mechanism for the retrieval and storage of persistent objects, we
need a way to express what objects we want. The general issues of object-
oriented query languages are still at the research stage with different directions
favoring procedural or declarative approaches. Most concrete proposals from the
latter school, so far, have been based on relational algebra (Object SQL); see for
example [Kim 1990, Loomis 1991].
However, many commercial applications (probably the majority) do not need
the full power of relational algebra to fulfill their functionality, since the types of
retrieval performed are pretty much known in advance. Therefore, simplicity
and flexibility is often more important than complete generality.
In this section, we will look at a simple approach that can be incorporated with
our relational database encapsulation to express a fairly broad class of queries in
a very natural way. It can also be used as a basis for automatic translation into
SQL statements, provided that the queries are not too complex.

Query frames
The idea is to transpose the technique of Query-by-Example [Zloof 1977] to the
object-oriented world. Rather than passing a query as a string expressed in some
query language, we may simply supply a template describing the retrieval
criteria for each attribute of a persistent object. The storage manager responsible
for retrieving the corresponding type of object may then inspect the template and
return the objects matching the criteria.
A possible scheme would be the following: the client creates a new object of
the required persistent type, fills in the attributes that will serve as retrieval
criteria, and calls a retrieve operation on the object. The supplier side will then
fill in the missing attributes by returning all matching objects, one by one, using
the iteration facilities described earlier.
However, there are some disadvantages with this approach. First, basic
attributes that are not of reference type (like INTEGER or REAL) always have
values. Therefore, there is no obvious way to signal whether an attribute of this
type has been set or not.
If a query result contains a real attribute temperature, a value 0 in the template
could mean either null (all objects wanted), or zero (only objects of temperature
zero wanted). This can be circumvented by defining special values (usually the
largest representable negative numbers) and letting clients use these to signify
null values for reals and integers.
However, a more severe drawback is that the selection criteria are limited to
exact equality. If this is all we need, the approach is nice and simple, but more
expressiveness is usually required. So we are going to use a more general
 QUERYING A PERSISTENT OBJECT MODEL 313

approach, which is to define for each persistent class a corresponding query

frame.
The query frame of a class is a class containing attributes with the same
names, but where all basic types have been replaced by ANY (a predefined type
to which all other types conform), and all class types have been replaced by the
type of its corresponding query frame. The query frame class corresponding to
class REGISTRATION is shown in figure 11.16.
Using objects of this kind to set up a query frame structure, rather than the
objects themselves, opens up new possibilities for expressing criteria.

REGISTRATION_FRAME

attendee: PERSON_FRAME
registered_at: ANY
amount_paid: ANY
invoice_sent: ANY
confirmed: ANY
paper_sessions: ANY
selected_tutorials: SET [TUTORIAL_FRAME]

Figure 11.16 Query frame for class REGISTRATION

Retrieval by example
For each basic attribute of the query frame, there are two choices:
1. The frame attribute is set to a value of the same type as that of the
corresponding attribute in the model object, in which case the selection
criterion becomes exact equality on this value. This is an important
option, since we may want to compute the corresponding value
dynamically without being forced to convert the result into a string.
2. The frame attribute is set to a string, in which case the criterion may be an
expression in any language chosen. For attributes in the model object of
string type, we are then faced with a small ambiguity: string values will
always be interpreted as criteria expressions rather than as literal values.
So if the expression "> ’Johnson’" normally means “all values sorted after
’Johnson’”, some escape conventions are needed to express a literal match
of the same string. However, this is not much of a problem, since even
very simple string matching languages will need facilities for resolving
such situations anyway.
314 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

The two types of attribute initialization may be freely mixed in a query. As an

illustration, consider the query frames set up in figure 11.17 for the selection of a
set of registrations. The query asks for all registrations entered after March 15,
1994, where an invoice was sent, less than $500 has been paid, and the attendee
lives in the USA.

REGISTRATION_FRAME PERSON_FRAME

attendee registration: Void

registered_at: "> ’1994-03-15’ " name: Void
amount_paid: "< 500 " affiliation: Void
invoice_sent: true address: "LIKE ’%USA%’"
confirmed: Void postal_mail: Void
paper_sessions: Void email: Void
selected_tutorials: Void phone: Void
fax: Void

Figure 11.17 Selection of registration objects using query frames

It is also possible to allow lists of query frame structures, each representing a

possible selection on the attribute values of the object and its supplier objects.
Such a list would then represent logical or of the selection criteria set up by each
frame structure.

Retrieval by key
It is important for a client to be able to cut off the retrieval of deep structures, so
that not everything needs to be transferred at once. Particularly, there may be
recursive object structures that simply cannot be retrieved in just one SQL
statement. To this end, we employ the convention that whenever an attribute of
class type (representing a reference to another object) is set to Void in a query
frame, the corresponding object is not retrieved. This is the case for
selected_tutorials in figure 11.17. If the attribute had been initialized with an
empty SET [TUTORIAL], the tutorial objects would have been retrieved too for
each registration.
When an object reference is cut off in a query frame by initializing an attribute
of class type to Void in a retrieval by example, it does not necessarily mean that
the client will not be interested in the corresponding object. It may be wanted
after some inspection of the retrieved data.
 QUERYING A PERSISTENT OBJECT MODEL 315

Assume a terminal operator scans through a large number of registration

objects without retrieving the corresponding attendee fields until a particular
registration is reached, at which point the personal data suddenly becomes
interesting. It would then be awkward if the application had to have a provision
for reinitializing a parallel REGISTRATION_FRAME object (this time with the
attendee reference initialized to an empty PERSON object rather than Void), and
then retrieve the same registration once more in order to get the personal data.
However, this will not be needed, since even if the PERSON object was not
retrieved the first time, the REGISTRATION_MANAGER has the key
information in the corresponding REGISTRATION_ROW object to get it directly
from the database.

11.6 PERSISTENT OBJECT MANAGEMENT

We will now take our design one step further, and establish enough detail to
outline a full scenario from start to end of a simple persistent object retrieval. To
this end, we extend the upper level of the preliminary sketch in figure 11.14 and
introduce a few more classes as shown in figure 11.18 to capture the general
principles involved.

PERSISTENT_OBJECT_MODEL
* MANAGER_
PERSISTENT
TABLE

model_object
APPLICATION

QUERY_FRAMES
*
QUERY_
FRAME

manager
STORAGE_MANAGEMENT
*
STORAGE_
MANAGER

Figure 11.18 Persistent object management

Our aim is to keep as much persistency detail as possible out of the class
definitions of the model objects. Therefore, the only static differences between a
class whose objects are potentially persistent and one whose objects are just
transient are the following two.
316 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

First, a persistent class must inherit from the class PERSISTENT (see
figure 11.18). This will enable clients to invoke retrieval operations on the
objects and iterate through sets of matching instances. Second, it will need to
redefine the signature of retrieve_by_example. The argument supplied as
retrieval criteria for a persistent REGISTRATION object, for example, must be
defined as REGISTRATION_FRAME.
Each type of persistent object is retrieved and stored in the underlying
database by a corresponding manager class, and all manager classes inherit from
STORAGE_MANAGER. The idea is not to build static knowledge into the
persistent classes by specifying the exact type of manager needed to take care of
the corresponding objects. Instead, there will be a dynamic mapping available,
so that persistent objects can invoke their proper manager by simply stating their
own type. Since the class name is already a unique type identification, a
mapping from class name strings to the corresponding manager will be enough
(to keep the discussion simple, we assume that the persistent classes are non-
generic).
The class name of an object can often be obtained dynamically from
predefined system classes in many object-oriented environments. One of two
standard techniques may then often be used for manager routing:
• If there are facilities in the environment to create a new instance of a class
directly from the class name, we only need a mapping to the class name of
the manager.
• If this is not possible but there is an “object cloning” facility available, we
may instead use object templates. At system initialization, one instance of
each persistent manager class is created to serve as a cloning template, and
a table of type TABLE [STORAGE_MANAGER, STRING] is set up to map
each persistent class name into a reference to one of the template objects.
The returned reference is then forwarded to a cloning facility, which will
instantiate a new copy of the object,
The class MANAGER_TABLE in figure 11.18 is assumed to take care of the
mapping, using some suitable technique. When called upon to access persistent
data, the features of PERSISTENT will thus look up the proper manager and
establish the bidirectional client link between the object and its manager.
It is important to note that although the two classes PERSISTENT and
STORAGE_MANAGER depend on each other, they are independent of which
subtype the other party will have. The specific manager class that will do the
actual conversion work must of course have full access to the attributes of the
persistent object, so REGISTRATION_MANAGER will statically depend on
REGISTRATION, but not the reverse (see figure 11.18).
 PERSISTENT OBJECT MANAGEMENT 317

Also, in this design we have assumed one manager for each persistent object
type. However, this is not necessary when dynamic routing is used. If there are
a large number of persistent classes in a system, their management will probably
tend to repeat typical patterns, and it may then be desirable to have fewer, more
general, managers to take care of groups of persistent object types.
We now proceed to look at the collaborating features of the two common
ancestors of persistent classes and storage managers respectively.

Persistent objects
The interface of class PERSISTENT is shown in figure 11.19. The first time a
persistency operation is called on a persistent object, the appropriate manager
template will be located through a routing table shared by all persistent objects.
A new storage manager will then be created and attached to the manager
attribute of the persistent model object, and a back reference assigned in the
manager object.
Three forms of retrieval, retrieve_by_example, retrieve_by_command, and
load_from_cursor, are available for persistent objects. All three commands will
be transparently forwarded to the appropriate storage manager without any
processing. Note that the only thing that needs to be changed when the feature
retrieve_by_example is redefined in a persistent class is the type of the query
frame argument. All implementation logic will reside in the corresponding
manager.
The first retrieval form implements the high-level selection criteria suitable for
application clients, which should be independent of any lower-level access
details. However, even the storage managers should know the low-level details
only of the objects they manage. Note that this includes what is defined by the
corresponding persistent class, but does not include what is defined by any of its
supplier classes.
For example, to retrieve a REGISTRATION object, the registration manager
will (in most cases) need to retrieve a corresponding PERSON object referred to
by attendee. However, it would be most unfortunate if the mapping of the
attributes of class PERSON into attributes of the virtual relational database (or
even worse, to the real database) had to be statically known by class
REGISTRATION_MANAGER.
If this were the case, we would need to create and maintain manager
implementations not only for each persistent class, but also for each combination
of a persistent class using another one as client. In a system with a large number
of persistent classes, the situation would soon become unmanageable.
One improvement would be to let the registration manager call a
PERSON_MANAGER to have the attendee part retrieved and translated.
318 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

PERSISTENT * STORAGE_MANAGER *

retrieve_by_example * Inherits: ACTION

– frame: QUERY_FRAME
retrieve_by_command retrieve_by_example *
– STRING – frame: QUERY_FRAME

load_from_cursor retrieve_by_command *
– holder: STORAGE_MANAGER – STRING

? holder ≠ ∅ load_from_cursor *
– holder: STORAGE_MANAGER
store
? holder ≠ ∅
iterate_on_result
− − Process each matching instance. store *
over: BOOLEAN iterate_on_result *
− − More instances left? − − Process each matching instance.

! Result ↔ (instance_action = ∅ or over: BOOLEAN

− − More instances left?
instance_action.over or
manager = ∅ or manager.over) ! Result ↔ (instance_action = ∅ or
instance_action.over or
instance_action: ACTION
row_object = ∅ or row_object.over)
− − Object to process each instance

! manager = ∅ → Result = ∅ ; instance_action: ACTION

− − Object to process each instance
manager ≠ ∅ →
Result = manager.instance_action set_action
– action: ACTION
set_action
– action: ACTION ! instance_action = action
! instance_action = action model_object * : PERSISTENT
routing_table:(1) MANAGER_TABLE row_object * : ROW
manager: STORAGE_MANAGER Invariant
Invariant model_object ≠ ∅ ↔
model_object.manager = @
manager ≠ ∅ ↔
manager.model_object = @

Figure 11.19 Collaborations for persistency management

However, this would create a lot of static dependencies between different

managers (somehow mirroring the dependencies between the corresponding
model objects, but with enough differences to create maintenance problems). So
a better solution is to go a step further, and always channel any persistency
operation through the persistent objects themselves.
This is where the alternative retrieval forms retrieve_by_command and
load_from_cursor come in. They are both meant for storage managers rather
 PERSISTENT OBJECT MANAGEMENT 319

than application clients. The arguments supplied when invoking these operations
can be used for communication between managers, but the routing will be done
by the corresponding model object, so that no unwanted static dependencies are
created.
The iteration features are similar to the ones already discussed for the lower-
level clusters. An application can attach an action object (usually itself) and then
receive a callback for each retrieved object instance matching the selection
criteria.

Storage managers
A storage manager translates persistent data between a model object and a
corresponding virtual relation (in case the instances are stored in a relational
database, as for REGISTRATION and PERSON in our example) or some other
storage (in case the instances are stored elsewhere, as for TUTORIAL). We will
only discuss the relational aspect in this case study.
The three forms of retrieval are different. The first, retrieve_by_example, will
cause the manager to read the supplied query frame object (or object structure, if
“inner” frame objects are also included) and use the attribute information to find
a suitable SQL query that will return the data required to set up the matching
objects.
As was argued earlier, it is desirable to minimize the static dependencies on
the exact organization of the real database, which is why we introduced a virtual
relational database represented by the row classes. However, the SQL
statements certainly need to be phrased in terms of the current database schema,
so how can we avoid becoming statically dependent on that schema when putting
the queries together?
We will return to the issue of automatic generation of SQL queries in the
concluding section, but for now we will only assume that whatever steering
information needed to dynamically construct the SQL statements that may occur
in our system (not always that many different types) is somehow maintained
outside the compiled executables. Applications will thus not need recompilation
when schema changes occur that do not affect the logical organization of the
persistent objects, which is our goal.
Unless we come up with a good automatic translation for a broad class of
queries, the stored tables may have to be structured ad hoc and perhaps not be so
trivial to maintain. However, even with a low level of automatic support, we
should be better off than if we are forced to change our compiled classes for each
minor schema change.
We assume that a class SQL_MAPPINGS will encapsulate a set of mapping
primitives, which will be used by the managers to dynamically build the required
320 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

SQL statements. In an ambitious approach, the mapping data required would

probably be stored with the objects in the relational database.
The retrieve_by_command feature is mainly used when a manager needs to
obtain an object through its primary key. As was explained in an earlier section,
a REGISTRATION_FRAME query object may be set up to cut off the retrieval of
personal data by initializing the attendee reference to Void. The registration
manager will then only retrieve the basic attributes of REGISTRATION, but will
keep the primary key to the corresponding PERSON object in case it is requested
later.
A common scenario may be a terminal operator quickly browsing through a
large number of registrations with retrieval of the personal data turned off.
When certain field values appear on the screen, the operator becomes interested
and orders the personal data to be filled in. The registration manager will then
typically be in the middle of its callback execute routine processing the current
instance of an iterate_on_result, and have a truncated REGISTRATION as
model_object.
The application object (which receives the order while waiting for input in its
callback action) then creates an empty PERSON_FRAME object, assigns it to the
attendee attribute of the REGISTRATION_FRAME object, and issues a new
retrieve_by_example on the same registration object.
The registration manager then detects that a new retrieve has been issued in
the middle of an iteration, which leads to a different action. Rather than as a
request for a new retrieval of model objects, the query is now understood as a
request to retrieve more of the deep structure of the REGISTRATION instance
already available. The degree of depth in such a new retrieval is again controlled
by the values (void or non-void) of the non-basic attributes of the frame object
and its suppliers, recursively. In the case of PERSON, there is no further
structure to retrieve.
So the manager rescans the query frame and detects that the attendee attribute
is no longer void and should be filled in. The registration manager then
generates a suitable SQL query to obtain the missing data and invokes a
retrieve_by_command on an empty PERSON object with the SQL string as
argument. It then attaches itself as action object and starts a separate
iterate_on_result, which will only return one PERSON object since it was
retrieved by its primary key. The translated person object is then assigned to the
registration object and control is returned to the application execute routine,
which displays the missing data on the screen and the operator can continue
browsing.
(The conventions just described represent of course a rather special design
decision, but the idea is to convince the reader that reasonably clean solutions are
indeed possible.)
 PERSISTENT OBJECT MANAGEMENT 321

Finally, the third form of retrieve, load_from_cursor, directs the manager to

load its model_object with relational data obtained by another manager.
When composite objects like REGISTRATION are to be retrieved from the
database, the most efficient way is usually to let the database server do the
required joins to obtain both the registrational and personal data at once. This
means that a registration manager will construct the full query and supply it to a
REGISTRATION_ROW object, which will call the database server and get a set
of table rows in return.
The registration manager then invokes a load_from_cursor on the row object
to load all basic attributes of the registration, and the manager can translate these
into the registration object. However, the person attributes for the attendee
supplier object remain, and they cannot be loaded and translated by this
manager, since we do not want cross-dependencies on internal formats.
Instead, the registration manager simply creates an empty PERSON object and
invokes load_from_cursor on it, supplying itself as cursor holder. The person
object does not know anything about cursors, but it can propagate the request to
a PERSON_MANAGER, which will then access the supplied manager argument,
extract the corresponding REGISTRATION_ROW, and supply it to a
PERSON_ROW as argument of an attach_to_query.
The person manager invokes load_from_cursor on the PERSON_ROW object,
which (since it has been reattached) will then transfer data from the cursor held
by the registration row, and then translates the resulting row attributes into the
person object. The registration manager can now use the retrieved person object
to complete its registration object.

A full retrieval scenario

We are now ready to present a complete scenario describing how a persistent
registration object is retrieved from the database. A dynamic object diagram
with its accompanying scenario box can be found in figure 11.20. However, for
the interested reader, we will also go through each step in more detail below and
mention the operations involved.
A typical scenario would proceed as follows.
1. A client initializes a REGISTRATION_FRAME with attribute attendee
attached to a PERSON_FRAME object to signify that the personal data of
each registration should also be retrieved. The attributes (of both frame
objects) whose values are to be part of the selection criteria are set to
mirror the conditions. The client then invokes retrieve_by_example on a
REGISTRATION object, supplying the query frame as argument.
The registration object invokes a corresponding retrieve_by_example on
its REGISTRATION_MANAGER passing the query frame. If the
322 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

1
APPLICATION_CLIENT REGISTRATION_FRAME

31 2, 8
19
PERSON REGISTRATION
29 20 18, 30 3, 9
21 4
PERSON_MANAGER REGISTRATION_MANAGER
22, 24, 28 5, 10, 13, 17
23
PERSON_ROW REGISTRATION_ROW
12
26 16 15 6, 11, 14
27
25
DB_RESULT 7
DB_QUERY

Scenario: Retrieval of persistent registration objects

1−2 The client initializes a query frame with selection criteria, and asks a
registration object to retrieve matching instances.
3−7 The registration object calls a registration manager, which inspects the query
frame and sends the appropriate SQL statements to a registration row object.
The row object forwards the query to the database server, which returns a
cursor structure.
8−10 The client tells the registration object to iterate on the result, that is load the
matching objects one by one, each time giving the client a callback to process
the instance. The request is forwarded to the registration manager, which
starts an iteration on the registration row, supplying itself as action object.
11−12 The registration row starts an iteration on the cursor structure, forwarding
the manager as action object. The manager action is called with the cursor
pointing to the data of the first matching object.
13−19 The registration manager tells the registration row to load itself from the
cursor, and sets the basic attributes of the registration object by translating
from the corresponding row attributes. It then creates a person object and
tells it to load itself from an existing cursor, supplying itself as cursor holder.
20−23 The person object forwards the load request to a person manager, which
obtains the registration row from the registration manager argument. The
person manager tells a person row to attach itself to an existing cursor
structure held by the registration row.
24−31 The person manager tells the person row to load itself from the current cursor
and sets the attributes of the person object. The registration manager
completes the registration by inserting a reference to the person object, and
calls the client action routine for application processing.

Figure 11.20 Typical retrieval scenario

 PERSISTENT OBJECT MANAGEMENT 323

registration object has no manager, a new one is created and attached using
the shared routing table.
2. The registration manager then translates the query frame attribute values to
appropriate SQL statements, and calls query on a REGISTRATION_ROW
with the query string as argument (see figure 11.15 for the interface of row
classes). The REGISTRATION_ROW attaches itself to a DB_QUERY
object and calls its query operation passing the SQL string. A set of table
rows is then returned from the database server (see figure 11.8 for the
interface of the database encapsulation).
3. The client uses set_action on a REGISTRATION to attach an action object
for processing (usually itself) and then calls iterate_on_result on the
registration, which is passed to the iterate_on_result of the manager. The
registration manager calls set_action on the REGISTRATION_ROW
supplying itself as action object, followed by an iterate_on_result on the
row.
4. The REGISTRATION_ROW transfers the manager as action object to the
DB_QUERY and calls its iterate_on_result. The DB_QUERY creates a
DB_RESULT representing the first matching table row and invokes the
execute callback in the registration manager. The manager then calls
load_from_cursor on the REGISTRATION_ROW, which then calls
load_object through the cursor feature of DB_QUERY, supplying itself as
receiving object.
5. The DB_RESULT object loads matching attributes of the first row of the
result into the REGISTRATION_ROW object. The execute routine of the
registration manager then proceeds to translate the row object attributes
into the REGISTRATION object.
6. All basic attributes of the registration have now been loaded, and the
registration manager proceeds to retrieve person data while still
performing its execute callback routine. The registration manager creates
a new PERSON object and invokes load_from_cursor on the empty object.
The cursor holder passed as argument to load_from_cursor is a reference
to the REGISTRATION_MANAGER itself.
The PERSON object then invokes load_from_cursor on its
PERSON_MANAGER (attached to the person object via feature manager).
The REGISTRATION_MANAGER reference just received by the PERSON
object is again passed as cursor holder in this second call.
7. The load_from_cursor command in the PERSON_MANAGER starts by
getting a reference to a REGISTRATION_ROW through the row_object
324 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

feature of the registration manager passed in as argument. It then invokes

attach_to_query on the PERSON_ROW referred to by feature row_object,
passing as argument the registration row. The effect is to connect the
person row to the cursor structure held by the registration row.
When this has been done, the person manager issues a
load_from_cursor on the PERSON_ROW. The person row calls
load_object on the cursor of its newly connected DB_QUERY and
supplies itself as receiver argument. The DB_RESULT object now loads
matching attributes of the first row of the result into the PERSON_ROW
object.
8. The person manager translates the person row object attributes into the
PERSON object. Control is then returned to the execute routine of the
registration manager, which finishes its work by attaching the person
object to the registration, which then becomes complete. It then invokes
the callback routine of the client.
9. The execute of the client then processes the returned persistent object.
When the routine exits, control is returned to the execute of the registration
manager, which returns it to the DB_QUERY, which moves the cursor to
the next resulting table row and again invokes a callback in the manager.
10. The above scenario continues until no more resulting rows remain, or the
application client has signaled termination by setting the over feature of its
action object to true. When this happens, the manager will propagate the
decision by setting its own over feature, and this will make the
iterate_on_result of DB_QUERY return control instead of issuing a
callback. The corresponding iterate_on_result of the manager also
returns, and the client may proceed at the point of its initial iteration call to
the registration object.
We will conclude the case study with a discussion on how much automatic
support is feasible for the generation of SQL queries.

11.7 AUTOMATIC GENERATION OF SQL STATEMENTS

Since the SQL queries must be expressed in terms of the real database schema,
we do not want to have them hard-wired into our software applications. So we
are faced with two problems:
• We need to obtain dynamically the mapping between our virtual database
and the real database in order to generate a SQL selection statement that
will return the proper columns for our row classes.
 AUTOMATIC GENERATION OF SQL STATEMENTS 325

• We need to find a way for application clients to express selection criteria

solely in terms of model object attributes, which can be dynamically
translated into SQL statements on the real database.
These are challenging requirements which are not easy to meet. If we were to
invent an object-oriented query language completely unrelated to SQL, the
dynamic translation would most likely become intractable. However, if we
impose some suitable restrictions on the permitted queries and on the relational
mappings used, new possibilities open up.

Model queries
The idea is to use SQL syntax transposed to the model object attributes. If the
strings assigned to the attributes of a query frame object are valid SQL
expressions restricting the corresponding model object attributes, a simple
translation scheme is possible. With this approach, the query frames of
figure 11.17 are equivalent to the query shown in figure 11.21, expressed in an
object-oriented SQL notation.

select
registered_at, amount_paid, invoice_sent, confirmed,
paper_sessions, attendee.name, attendee.affiliation,
attendee.address, attendee.postal_mail, attendee.email,
attendee.phone, attendee.fax
where
registered_at > ’1994-03-15’ and
amount_paid < 500 and
invoice_sent = true and
attendee.address LIKE ’%USA%’

Figure 11.21 An object-oriented SQL query

Such a notation may also be used by application clients as an alternative to the

query frames of retrieve_by_example. A selection is then passed as a string to
retrieve_by_command. If all attributes that are part of the selection criteria in the
model object are representable as SQL expressions built from columns in the real
database, the translation is reduced to simple substitution.
We may still allow some criteria involving object attributes that cannot be
transformed to SQL expressions using the database schema, but then these
criteria cannot be part of the selection sent to the database server. A
correspondingly larger number of table rows will thus be retrieved, and the ones
not matching the non-translatable criteria filtered out a posteriori by the manager
of the persistent object type.
326 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

Attribute mappings
The mapping facilities that are needed for storage managers to make automatic
conversions are encapsulated in the class SQL_MAPPINGS, whose interface is
outlined in figure 11.22. All mappings are between strings and can be
maintained as simple editable tables and, for example, be stored as persistent
strings in the database.

SQL_MAPPINGS

db_select: STRING
− − Table columns of real database
− − corresponding to persistent_class
– persistent_class: STRING
db_from: STRING
− − Tables corresponding to persistent_class
– persistent_class: STRING
db_join_condition: STRING
− − Table join condition in real database
− − corresponding to a non-basic attribute
− − of a persistent class. Argument format:
−− "CLASS_NAME.attribute_name"
– attribute: STRING
db_attribute_expr: STRING
− − Table column expression in real database
− − corresponding to a non-basic attribute
− − of a persistent class. Argument format:
−− "CLASS_NAME.attribute_name"
– attribute: STRING
db_update_pattern: STRING
− − SQL template for generation of real database
− − update from attributes of row_class
– row_class: STRING

Figure 11.22 Query generation primitives

The first four features return information needed to build queries that can be
sent to the database server. A generated query consists of three parts:
• The table columns that need to be selected to provide enough information
for building each persistent instance.
• The join conditions for the natural joins to be performed between tables
representing a client and tables representing a supplier. For example, class
REGISTRATION needs to join its tables with those of class PERSON to
obtain the data corresponding to the attribute attendee.
 AUTOMATIC GENERATION OF SQL STATEMENTS 327

• The selection criteria expressed as SQL conditions on the retrieved

database table columns.
The target columns of the selection are precisely the attributes of the
corresponding row classes. The call db_select ("REGISTRATION " ) would
return:
"REGISTRATION.PERSON_CODE as REGISTRATION_ROW$PERSON_CODE,
REGISTRATION.ENTRY_DATE as REGISTRATION_ROW$ENTRY_DATE,
REGISTRATION.PAPER_SESSIONS as REGISTRATION_ROW$CONFERENCE,
REGISTRATION.TUTORIAL1 as REGISTRATION_ROW$TUTORIAL1,
REGISTRATION.TUTORIAL2 as REGISTRATION_ROW$TUTORIAL2,
REGISTRATION.TUTORIAL3 as REGISTRATION_ROW$TUTORIAL3,
REGISTRATION.TUTORIAL4 as REGISTRATION_ROW$TUTORIAL4,
REGISTRATION.CONFIRMATION_DATE as
REGISTRATION_ROW$CONFIRMATION_DATE,
REGISTRATION.INVOICE as REGISTRATION_ROW$INVOICE,
INVOICE.ISSUED_DATE as REGISTRATION_ROW$ISSUED_DATE,
INVOICE.AMOUNT_PAID as REGISTRATION_ROW$AMOUNT_PAID"

If the selection includes personal data, a db_select ("PERSON " ) will also be
needed to add the corresponding columns to the format of the retrieved rows.
Besides the selected columns, we need to accumulate all tables that participate
in the resulting selection. These could in principle be extracted from the former
strings, since we store all column selections qualified by table name to avoid any
name clashes. However, sometimes table alias names must be used for
unambiguous reference (see the next section).
We therefore store the table reference strings separately, and the calls db_from
("REGISTRATION " ) and db_from ("PERSON " ) yield, respectively:
"REGISTRATION, INVOICE"
"CUSTOMER, AFFILIATION"

Next, we need to build the restriction clause on the rows initially selected. This
must include the corresponding join condition for each “inner” object. Calling
db_join_condition ("REGISTRATION.attendee " ) would return:
"REGISTRATION.PERSON_CODE = CUSTOMER.CUSTOMER_CODE"

In fact, when the basic attributes of a class are stored in more than one table,
there will be join conditions involved besides the ones introduced by “inner”
objects. Therefore, entries may also have to be stored for the class names
themselves, not just for each of their non-basic attributes. This is the case for
both our classes and the calls db_join_condition ("REGISTRATION " ) and
db_join_condition ("PERSON " ) will return the following strings:
"REGISTRATION.INVOICE = INVOICE.INVOICE_CODE"
328 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

"CUSTOMER.COMPANY = AFFILIATION.COMPANY_CODE"

which must also be part of the restriction clause. Finally, before completing the
generated query by appending the selection criteria supplied by the client object,
each model object attribute occurring in these criteria needs to be substituted by
a corresponding SQL expression in terms of the real database columns. To this
end, the calls:
db_attribute_expr ("REGISTRATION.registered_at " )
db_attribute_expr ("REGISTRATION.amount_paid " )
db_attribute_expr ("REGISTRATION.invoice_sent " )
db_attribute_expr ("PERSON.address " )
will return the following strings, respectively:
"REGISTRATION.ENTRY_DATE"
"INVOICE.AMOUNT_PAID"
"INVOICE.ISSUED_DATE is not null"
"PERSON.STREET || ’, ’ PERSON.ZIP_CODE || ’, ’ PERSON.CITY
|| ’, ’ COUNTRY_NAME"

Self-joins
Assume a conference system which also keeps track of alternative arrangements
(often called spouse programs) for people accompanying attendees. A possible
model is shown in figure 11.23.

ATTENDEE ● PERSON COMPANION ●

Inherits: PERSON name: VALUE Inherits: PERSON

address: VALUE
registration: REGISTRATION selected_tour: VALUE
companion: COMPANION attendee: ATTENDEE

Figure 11.23 Simplified model for ATTENDEE and COMPANION

An ancestor class PERSON contains common attributes, while ATTENDEE

and COMPANION add the needed extensions.
If for some legacy reasons we want to use only one relational table to store
both types of object, letting each row represent either a conference attendee or
someone accompanying an attendee, such a table containing personal data could
look like the one in figure 11.24. A row representing an attendee would leave
the selected tour column blank, while a companion row would have companion
code zero.
 AUTOMATIC GENERATION OF SQL STATEMENTS 329

PARTICIPATOR
Name Type
PARTICIPATOR_CODE INT
NAME VARCHAR(50)
ADDRESS VARCHAR(50)
REGISTRATION_CODE INT
COMPANION_CODE INT
SELECTED_TOUR CHAR(12)

Figure 11.24 Table covering both ATTENDEE and COMPANION

At first glance, it may seem as though an ATTENDEE object would need to be

built from two SQL queries: the first selecting the attendee row and the second
selecting a possible companion based on the retrieved value of the companion
code column. However, this is not necessary.
We could enter a companion row with blank name and address fields, and let
it represent the “empty” companion. In each row representing an
unaccompanied attendee, the corresponding key value would be used as
companion code, thus ensuring referential integrity.
The SQL selection below could then be generated (column renaming omitted
for brevity), and the resulting rows would contain all data needed to regenerate
the complete ATTENDEE objects, one by one.
select
ATTENDEE.NAME, ATTENDEE.ADDRESS,
COMPANION.NAME, COMPANION.ADDRESS, COMPANION.SELECTED_TOUR
REGISTRATION.REGISTERED_AT, ...
from
PARTICIPANT as ATTENDEE, PARTICIPANT as COMPANION,
REGISTRATION
where
ATTENDEE.name = ’John Hopkins’ and
ATTENDEE.COMPANION_CODE = COMPANION.PARTICIPANT_CODE and
ATTENDEE.REGISTRATION_CODE = REGISTRATION.REGISTRATION_CODE

The generation scheme

To summarize, the generation of a SQL query using the features of
SQL_MAPPINGS would proceed as follows:
1. Scan the query frames to find out how many objects must be instantiated
from each table row returned by the SQL query. This is recursively
controlled by the client, which can either initialize a reference to a non-
basic object in the query frame, or leave it as Void.
330 RELATIONAL AND OBJECT-ORIENTED COEXISTENCE

2. Construct the select-part by appending the database attribute list of the

corresponding row class for each object that is to be loaded from a
resulting table row. This is done through repeated calls to db_select.
(Here, as in the following steps, we implicitly assume that any needed
keywords and comma separators are also inserted.)
3. Construct the from-part by appending the database tables involved for
each object (repeated calls to db_from).
4. Start constructing the where-part by appending the join condition for each
model object, as well as for each attribute referring to a participating
“inner” object (like attendee in our example). This is done through
repeated calls to db_join_condition.
5. Translate all selection criteria supplied by the client query frames that are
expressed in terms of model object attributes into equivalent criteria
understandable by the database server. This is accomplished by
substituting the corresponding SQL expressions of the real table columns
using calls to db_attribute_expr.
6. Complete the where-part by appending the resulting criteria strings.

Updating the database

The generation of SQL statements for updates and insertions is much simpler.
Since all data, including primary keys, is available in the row objects, we only
need to store SQL templates to be filled in with current attribute values. The call
db_update_pattern ("REGISTRATION " ) would return:
update REGISTRATION set
"ENTRY_DATE = :entry_date, CONFERENCE = :conference,
TUTORIAL1 = :tutorial1, TUTORIAL2 = :tutorial2,
TUTORIAL3 = :tutorial3, TUTORIAL4 = :tutorial4,
CONFIRMATION_DATE = :confirmation_date, INVOICE = :invoice
where PERSON_CODE = :person_code"

Preceding a name by a colon is the usual SQL placeholder notation to signify

that a value will be dynamically substituted. The manager could either use the
pattern to fill in a complete SQL statement for immediate execution, or put it in a
stored procedure to be furnished by attribute values using the dynamic SQL
conventions for argument passing.
The issued_date and amount_paid attributes are not part of the pattern, since
in our legacy example we assume that these values will only be read by the
conference system, while the corresponding updates are done by a separate
invoicing system.
 AUTOMATIC GENERATION OF SQL STATEMENTS 331

Moreover, we have omitted insert and delete from our design for simplicity,
but the mappings for these features can be handled similarly.

11.8 FULL STATIC ARCHITECTURE

Finally, we present the complete architecture of our persistency layers in
figure 11.25, showing how it all fits together. As can be seen, the different
layers are well separated and the static dependencies between them have been
reduced to a minimum.
MANAGER_
PERSISTENT_OBJECT_MODEL TABLE

* 1 routing_table
PRESENTATION
*
PERSISTENT
● ● ●
TUTORIAL REGISTRATION PERSON model_object

QUERY_FRAMES
*
APPLICATION QUERY_
FRAME
SQL_
MAPPINGS

STORAGE_MANAGEMENT
manager

*
TUTORIAL_ REGISTRATION_ PERSON_ STORAGE_
MANAGER MANAGER MANAGER MANAGER

row_object instance_action
VIRTUAL_RDB

REGISTRATION PERSON_ *
ROW ACTION
ROW ROW

row_action
RDB_INTERFACE

▲ ▲ ▲ cursor ▲
DB_SESSION DB_CHANGE DB_RESULT DB_QUERY

Figure 11.25 Persistent object mapping: full static architecture

12 Exercises

12.1 CLUSTERING A PROBLEM DOMAIN

(This exercise requires the preliminary reading of chapters 3−4.)
Analysis classes relate directly to a problem domain. However, depending on
the selected viewpoint, different logical groupings may be defined to satisfy
criteria such as high cohesion, low coupling, and a (possibly nested)
classification structure emphasizing important aspects of system usage.
1. Partition the cluster TRANSPORTATION given in figure 12.1 according to
the three alternate viewpoints: Crew Member, Passenger, Manufacturer.

TRANSPORTATION

FLIGHT_
AIRCRAFT CAPTAIN ATTENDANT CABIN

LUNCH_
SEAT RESERVATION AIRLINER
SERVICE

SPECIAL_
SEGMENT FLIGHT CAPACITY
MEAL

EXIT_ TRAFFIC_
ALTITUDE CONTROLLER SPEED
DOOR

Figure 12.1 Cluster TRANSPORTATION

2. The set of analysis classes given in figure 12.2 was inspired by

[Rumbaugh 1991]. Suggest appropriate viewpoints corresponding to the
two different clusterings, and replace the temporary symbols C1, C2, …
by self-explanatory names.

332
 CLUSTERING A PROBLEM DOMAIN 333

Viewpoint 1:
C1

GHOST LADY

LORD COOK

CASTLE

C3

C4 C2

STAIRS FLOOR TOWER DUNGEON

{
CORRIDOR ROOM BOILING_OIL DUNGEON

WINDOW STONE DRAWBRIDGE

Viewpoint 2:
C1
C2

GHOST LADY
STAIRS FLOOR

WINDOW COOK
CORRIDOR ROOM
CASTLE

BOILING_OIL LORD
STONE TOWER

DRAWBRIDGE
MOAT DUNGEON

Figure 12.2 Two alternative cluster partitionings based on different viewpoints

3. The initial analysis of a vacation cruise system has led to a set of classes
given in figure 12.3. For each of the three different viewpoints Travel
Agent, Vacationer, and Captain, draw the system borderline that leaves out
the classes considered irrelevant in a system based on this viewpoint.
334 EXERCISES

VACATION_CRUISE

TWIN_BED_ SINGLE_
LIFE_BOAT CAPTAIN DEPOSIT
CABIN CABIN

SWIMMING_ MIDSHIP_ SAFETY_

CLUB_MED TRIP
POOL OFFICER BOX

UPPER_ LOWER_
PRICE LANDMARK LATITUDE
DECK DECK

SOCIAL_ CRUISE_
PROGRAM WEEK FLAG BOOKING
LINER

Figure 12.3 Cluster VACATION_CRUISE

12.2 DEFINING CLASS RELATIONSHIPS

(This exercise requires the preliminary reading of chapter 4.)
For each numbered requirement, identify candidate classes and define the type of
relationship involved between the candidate classes. Argue your choice.
1. All factory rooms have a buzzer triggered by a central clock to signal the
end of the working day.
2. Architects are often involved in different concurrent projects.
3. A country has a capital city.
4. Political parties, such as the Republican party, the Democratic party, and
the Libertarian party, have voters and supporters.
5. A road connects two cities.
6. Teaching assistants are Ph.D. students and faculty members. They get a
monthly payroll check from the university administration.
7. All files are defined as a sequence of bytes. A file can be a text file, a
directory, or a special file. A text file can be replaced by a symbolic link
to a plain file. A plain file can be either a binary file or a sequential data
file. A file belongs to exactly one file system.
8. An undirected graph consists of a set of vertices and a set of edges. Edges
connect pairs of vertices.
9. Project planning starts with a list of tasks. Each task is recursively
decomposed into subtasks until elementary tasks are reached. Resources
 DEFINING CLASS RELATIONSHIPS 335

are then allocated for each elementary task. One resource is the task
schedule, defined by a starting time and a duration. The resources
available for a non-elementary task are determined by looking at the
resources for each of its subtasks.

12.3 ASSERTIONS AND CLASSIFICATION

(This exercise requires the preliminary reading of chapters 3−4.)
1. Given the interface of class PERSON in figure 12.4, add a precondition to
the command marry expressing that one can only marry someone who is
not already married. Assuming two distinct person references john and
helen, sketch one or more restricted features that may be used by an
implementation of marry to ensure that
john.marry (helen)
is strictly equivalent to
helen.marry (john)
2. Write the class interfaces of BRIDE and GROOM in a way that ensures
that each person can only be married to persons of the opposite sex, and
only to one at a time.

PERSON

spouse: PERSON
marry
BRIDE GROOM
– other: PERSON
! spouse = other
MARRIAGE
Invariant
spouse ≠ ∅ → spouse.spouse = @

Figure 12.4 Class PERSON and cluster MARRIAGE

3. Given the utility class FAMILY in figure 12.5, make it an ancestor of the
class PERSON in figure 12.4 and extend the contracts in classes PERSON,
BRIDE, and GROOM to keep the parent−child relationship consistent and
prevent brothers and sisters from marrying each other.
336 EXERCISES

FAMILY

father, mother: PERSON

children: SET [PERSON]

Figure 12.5 Class FAMILY

12.4 DYNAMIC BEHAVIOR

(This exercise requires the preliminary reading of chapter 5.)
The requirements for claim processing within a product selling organization
leads to the scenario description and dynamic diagram given in figure 12.6. The
diagram captures the main control flow of a claim processing system.
A customer’s complaint is received by a salesperson and initiates a claim
procedure. A claim form is sent to the product manager and to other persons. A
copy of the form is routed to the technical services responsible for the defect as
indicated by the salesperson.
The technical department makes a report with diagnosis and corrective action
taken. The report is passed back to the product manager, who accepts or turns
down the claim. The salesperson eventually authorizes reimbursement or
replacement products as settlement.
Extend the dynamic model as indicated below. (Recall that message relations
signify the flow of control, not of data. Therefore, passive information
containers have incoming arrows when accessed by active client objects, but
outgoing arrows only when calling in turn some other object.)
1. Define a “quick settlement” scenario whereby the customer and the
salesperson immediately agree on a settlement over the phone for low-
value products.
2. Update the scenario (dynamic diagram and textual description) so that
somebody (salesperson, sales manager, or receptionist) phones the
customer to explain the settlement when a decision has been reached.
3. Update the scenario to include persons in charge of following each claim
through the claim handling procedure. If decisions are delayed, customers
should be notified of current status to avoid repeated soliciting.
4. Update the scenario to show shipment of replacement products, or
reimbursement checks, respectively.
 DYNAMIC BEHAVIOR 337

1: Incoming external event (phone call)

TECH_DEP
5 2
SALESPERSON DELIVERY
LOGISTICS
3
PRODUCT_ 6
CLAIM_FORM 4
DEVELOPMENT

QUALITY 8 11
PRODUCT_MANAGER SALES_MANAGER

7 10 12
9 13
CLAIM_REPORT SETTLEMENT ACCOUNTANT

14: Outgoing internal event (letter)

Scenario 1: Process a claim

1 A phone call is received by a receptionist and directed to a salesperson.
2−3 Salesperson checks delivery, and fills out a claim form according to
customer’s complaint.
4−5 Salesperson sends claim form to product manager with copies to persons in
technical department.
6−8 Technical department considers the claim, makes a claim report, and
forwards result to product manager.
9−11 Product manager consults claim report, makes settlement decision, and tells
sales manager.
12 Sales manager forwards claim report with authorization to accounting
department.
13−14 Accounting department sends letter with customer settlement: reimbursement
or replacement of product.

Figure 12.6 Dynamic scenario of claim processing system

12.5 PRESCRIPTION AND DESCRIPTION

(This exercise and the following ones require the preliminary reading of
chapters 3−5.)
Consider the following problem solving requirement:
338 EXERCISES

A shepherd wishes to move a cabbage, a lamb, and a wolf across a bridge in a

carriage. The carriage can only transport one item at a time. If the lamb is left
alone with the cabbage, the cabbage will be eaten. If the lamb is left alone with
the wolf, the lamb will be eaten.
Assume a scenario where everything is safely transported across the bridge
with a minimum number of trips, and infer the following static modeling
information.
1. List the analysis classes.
2. Group the analysis classes into clusters and draw the static architecture.
3. Outline class features.
4. Define assertions to ensure that no item will be eaten.
5. Define a class invariant to ensure that the transportation will terminate
(finite number of trips).
6. Express the successful strategy using a scenario and a dynamic diagram.
7. Referring to your resulting design, look for a will_eat feature that you may
have introduced and implement it in your favorite object-oriented
language.

12.6 TRAFFIC-CONTROL SYSTEM

The following specification is taken from [Korson 1990].
A software system manipulates the hardware for traffic control located at an
intersection. The hardware includes a set of sensors, the traffic lights, and a
control box. The software reads the state of the sensors, determines the new
state for the intersection, and signals the lights to change. In addition, the system
is to have other capabilities, such as a test option to cycle the lights through the
set of configurations. The system can also be set to a default state, which might
be flashing yellow in one direction and red in the other. The sensor indicates the
presence (or absence) of a vehicle in a particular lane.
There are several kinds of sensors with different internal workings, all of them
interrupt−driven with a bit set whenever the sensor is tripped. After the decision
to change the state of the intersection, every sensor is reset. This particular
intersection uses three-color lights for go-straight lanes. Each light has a current
state, a set of valid next states, and a default state. The controller physically
contains the switches for the lights, the data stores for the sensors, and a clock
for timing the state of the poll/decision cycle. The controller software reads the
clock, polls the sensors, determines the next state, and sends the signals to
change the lights.
 TRAFFIC-CONTROL SYSTEM 339

1. Delineate the software system borderline and the interface with the
hardware.
2. List implementation details that can be excluded from the analysis model.
3. List a first set of analysis classes mapping the problem domain.
4. Define the scenario and associated dynamic diagram describing two
vehicles arriving concurrently at the intersection from different directions.
5. Estimate the number of objects (class instances) referred to in your
dynamic model.
6. Outline the static architecture and the class descriptions.

12.7 DICE GAME

At the OOPSLA ’89 Conference in New Orleans (USA), a contest was organized
by Tom Love and presented to different representatives of the object-oriented
language community. The aim was to compare designs and solutions resulting
from tools, methods, and languages available at that time. The following is an
outline of the problem requirements.
A dice game called “Greed” is played between two or more players. Five dice
are rolled from a cup, and the winner is the first player to score 5000 points. To
enter the game, 300 points are needed on the first roll. If the first roll is greater
than 300, the player has the option to stop and keep the initial score or to
continue. On continuing, only the dice that have not yet scored may be rolled. If
all the dice score (in one or multiple rolls) the player may start a new turn by
rolling all five dice. A player may continue rolling as long as a score is made on
each roll. An individual roll producing no points is called a “bust”. When a bust
occurs, all points accumulated during the turn are lost and the turn ends. Three
of a kind score 100 times the face value of one die, unless the face value is 1 in
which case the score is 1000. Single values of 1 and 5 score 100 and 50,
respectively.
1. List candidate classes and define the analysis charts of a system that would
monitor the “Greed” game: players, scores, rules, display of results.
2. Outline the class interfaces with public features.
3. Outline a first static diagram.
4. Outline a scenario and dynamic diagram corresponding to a “bust.”
5. Look for analysis classes that may be generalized for use in any dice
game.
340 EXERCISES

6. Consider the different alternatives for adding a graphical interface to the

initial set of model classes.

12.8 CAR RENTAL COMPANY

A car rental company wishes to computerize its car reservation and invoicing
system. All vehicles are delivered from and returned to the same location. The
system will help keep track of vehicles, making single or block reservations, and
issuing invoices based on the pricing structure and running promotions.
The company has several different car models from different manufacturers in
its fleet of vehicles. The models are grouped in price categories. Different rental
plans are available: daily time and mileage rate, daily rate including unlimited
mileage, weekly rates, and a special weekend rate for vacationers.
The price charged by the car rental company is established in advance for a
given period. The price for a given model depends on the price category, the
type of customer, and the rental plan chosen. Corporate customers get a special
discount on normal rates, but not on special weekend rates.
The company finds it important to have information available on the options
which may be fitted to certain car models, such as two or four doors, automatic
or manual transmission, anti-lock brakes, cruise control, and so forth. The
information on what is actually fitted to the cars in the rental fleet must
correspond to the information on options provided by the car suppliers. There is
no charge for these options, but customers will often request some of them when
reserving cars, and the company wishes to try to meet such requests.
In addition to the fitted options there are optional non-fitted extras, such as
roof-rack, trailer, snow-chains, and child seats. These extras may be requested at
a special additional charge. Different optional insurance plans are also available
at extra charge.
A customer may handle a car reservation directly, through a travel agent, or
through another company. The reservation may request certain options and
extras. The means of payment is usually indicated at reservation time, but may
be changed when the car is returned. Reservations are accepted for a given
advance period, and cars are tracked daily according to their status: available,
booked, rented, under maintenance, being repaired, etc.
1. A first analysis resulted in the static architecture depicted in figure 12.7.
Assign possible labels to client/supplier relationships. Look for missing
classes and clusters.
2. Draw the associated dynamic model corresponding to a block booking of
two cars: a two-door Volkswagen with manual gears and a Chevy with
automatic transmission and roof-rack. The selected rental plan specifies
 CAR RENTAL COMPANY 341

CAR_RENTAL_SYSTEM

CONTRACT

SUPPORT

*
* MEANS_OF_
DATE CLIENT RENTAL
PAYMENT

CONTRACT_ELEMENTS

AVAILABILITY LOCATION * *
VEHICLE RATE

MODEL * *
EXTRA INSURANCE

OPTION_TYPE
* DRIVER
OPTION

RENTAL_ITEMS
VEHICLE_PROPERTIES

Figure 12.7 Static outline of the car rental system

unlimited mileage, one week rental, corporate rate, and collision damage
insurance policy.
3. A smart reuse manager has noticed that an existing corporate component
library already includes the classes PRODUCT, STORAGE_AREA,
PRICING, CONSUMER, DELIVERY, and AVAILABILITY, resulting from
a previous project dealing with stock management for a grocery store.
Adapt the car rental classification to take advantage of these classes.

12.9 TRUCK FREIGHT

A system is designed to automate the preparation of freight transportation on
trucks. Truck carriers must comply with strict transportation regulations. In
addition, they may become liable for any damage caused by an improper
delivery in case the regulations were not faithfully observed. The system should
342 EXERCISES

help assess the nature of the transported shipment and produce the legal
documents that truck drivers must handle.
The different laws and ordinances regulating the traffic of trucks on roads
keep changing: new ones are issued, others become obsolete. A product may be
prohibited for road transportation in some country, or in a geographic area.
Regulations are based on the type of freight: type of substance, oversized or
overweight loads, perishable products.
The regulations constrain the selection of the itinerary and the delivery time
frame. For instance, some toxic substances may not be transported across
populated areas, or at night; fresh fish must be transported from the harbor to the
retailers in less than two days; and livestock may not be transported for a
continuous duration of more than 5 hours.
Therefore, the truck carrier company is responsible for assessing the nature of
transported products and delivering the freight in good condition, on time, and
without violating the regulations.
Each truck load is defined by the properties of the transported goods. The
shipping company constantly queries and updates a freight database, and
manages the translation between the technical descriptions given by the customer
for each freight and their legal counterparts in terms of product classes,
restrictions on the itinerary, transportation periods, and the level of qualification
required for the truck driver. The system should support the following tasks:
• Perform a formal evaluation of a product to be transported, based on
information provided by the manufacturer. For chemical substances, the
evaluation is based on physical properties.
• Given the product identification code thus established, define the best
possible itinerary from a starting point to a destination with possible
restrictions on geographic areas and roads.
• Assign a truck driver with suitable qualifications for the type, weight, and
size of the freight.
Products are categorized hierarchically in three levels: the class of the substance,
the group within the class, and the identification code within the group. A
product may belong to several classes, groups, and codes, but only one is said to
be the major criterion while the others are considered minor criteria.
Special attention is paid to hazardous products, which are evaluated according
to their physical characteristics such as: flash point, boiling point, solubility, and
chemical components. Below are some examples:
• The flash point classifies a product within class 1 (inflammable liquids).
• The toxic effect classifies a product within class 2 (toxic substances).
 TRUCK FREIGHT 343

• The corrosive effect classifies a product within class 3 (corrosive

substances).
• The perishability classifies a product within class 4 (rapidly degradable
substances).
A substance may be of class 1, 2, and 3, but only one class is retained as the
major criterion. Within each class, groups are established and identified by one
alphabetic character. Below are some groups defined for class 2 (toxic
substances).

• 2.A: Very poisonous products with a flash point below 21 °C and a boiling
point below 200 °C, which do not belong to class 1.
• 2.B: Organic substances with a flash point ≥ 21 °C and non-inflammable
organic substances.
• 2.C: Metal−organic substances and carbonyls.
• 2.D: Non-organic substances that react with water, water soluble liquids,
and acids of poisonous gases.
• 2.E: Other non-organic substances.
Finally, for each group, identification codes are assigned to the product or
substance. Below are some codes within 2.C (metal−organic substances and
carbonyls):

• 2.C.10: Refrigerated perishable fruits or vegetables.

• 2.C.28: Live stock.
• 2.C.31: Organic lead compounds.
A first analysis produced the cluster sketch in figure 12.8 and one class chart
shown in figure 12.9. Continue the analysis by doing the following:

1. List the analysis classes and produce a class dictionary.

2. Define a first static model by completing the initial set of clusters and
classes.
3. Define a dynamic scenario describing the assignment of a truck driver to a
shipment.
4. Suggest a design that would make it possible to adapt the system to most
future regulation changes without the need to modify any source code.
344 EXERCISES

SHIPMENT

● *
● SET […] BASIC_ PHYSICAL_
GOODS { PRODUCT PROPERTY

FREIGHT COMPOUNDS

NON_
LIVE_STOCK FURNITURE ORGANIC ORGANIC

CHEMICALS GROCERIES

METAL_
CARBONYL
ORGANIC

CHARACTERISTICS

* *
S_GROUP S_CLASS CLASSES GROUPS

*
S_CODE TOXICITY

Figure 12.8 Partial cluster partitioning of truck freight system

CLASS BASIC_PRODUCT Part: 1/1

TYPE OF OBJECT INDEXING

Element defined in security data and stored cluster: PRODUCT
in database. created: 1994-04-06 jmn

Queries Eatable? Live? Flammable?

Corrosive? Water_soluble?
Flash_point, Boiling_point, Solubility,
Description, Hazardousness,
Substance class, Substance group, Substance code

Commands Assign substance class.

Assign substance group.
Assign substance code.

Figure 12.9 BASIC_PRODUCT class chart

 TRUCK FREIGHT 345

5. Assume there are less than 50 class categories and groups, but thousands
of identification codes. Investigate how this may or may not impact your
design choices.

12.10 REAL-TIME PROCESS CONTROL

In a fragrance company, the industrial production of perfumes employs a
chemical reaction controlled by some real-time software. A liquid mixture made
from raw vegetal essences is poured into a tank, and the heat is increased until it
reaches a certain peak controlled by external sensors.
There is also a feedback of elements flowing out of the boiler. The outgoing
steam is kept until it complies with some trade secret combination of essences
that are input as control parameters by the company chief executives and by
technicians (nicknamed “noses”) in charge of maintaining the accuracy of
existing or newly created fragrances. When perfection has been reached, the
steamed combination is forwarded to a cooling system.
A first analysis of the software resulted in the cluster chart shown in
figure 12.10.
The system has several control points, three of which are listed below:
• C1: control point to regulate the liquid valve

CLUSTER Fragrance_production Part: 1/1

PURPOSE INDEXING
Classes controlling the chemical reactions created: 1993-05-23 jmn
involved in producing various exact
fragrances.

Class / (Cluster) Description

CONTROLLER Actor responsible for monitoring the boiling process under

normal operational conditions.

HEATING Device enabling change of temperature accessed through the

controller.

PIPE Incoming and outgoing parts where the mixture is flowing.

SENSOR Actor retrieving external values.

TANK Container of mixture (liquid and steam) flowing through

pipes and heated with heating system.

STEAM_REACTION Actor used to record and check the production of steam from
the mixed reaction liquid.

Figure 12.10 FRAGRANCE_PRODUCTION cluster chart

346 EXERCISES

• C2: control point to regulate the steam valve

• C3: control point to regulate the heating process
Each control point has a corresponding sensor:
• S1: checking the liquid level in the tank
• S2: checking the steam pressure in the tank
• S3: checking the temperature level in the tank
In addition, the system must comply with some physical constraints: the water
level, pressure, and temperature in the tank must range between certain lower
and upper bounds. The corresponding values (which vary depending on the type
of reaction currently in process) are dynamically entered by the system operator.
To optimize the production process, the system must also enforce the
following monitoring strategy:
• If the tank becomes too hot, the system should first try to add more cold
liquid within a preset time frame. If after that the temperature is still too
high, the pressure is decreased during another time frame. If this also
proves insufficient, the heat is reduced as a last resort.
• If the tank becomes too cold, the reverse procedure is employed: first
increase the heat, then increase the pressure, and then reduce the amount
of incoming liquid.
What should be considered too hot and too cold, as well as the lengths of the
time frames, are input system parameters entered by the operator or retrieved
from a database.
1. Define the first architectural draft of the system.
2. Describe the different classes listed in the cluster chart.
3. How would you express the constraints at the design level? What will be
left to the programming phase?
4. Draw a dynamic diagram representing a complete production phase:
incoming of fragrance mixture, control of its compound elements,
optimization of the process according to the physical sensor values, and
transfer of the resulting steam into the cooler.
 Part V
Appendices
348 PART V
Appendix A:
BON textual grammar

A.1 INTRODUCTION
This appendix presents a formal syntax specification of the BON textual
notation, which is useful for automatic processing and for maintenance of BON
designs where no case tool is available. The specification gives a comprehensive
overview of all notational concepts in BON, facilitating the construction of
parsers for translating BON charts and diagrams into other desired formats.
Communicating BON designs from either case tools or text files to other tools
with well-defined interfaces is thus straightforward. Interesting possibilities in
this respect include configuration management tools, commercial DBMS
environments, and widely available document processors and desktop publishing
tools. For example, it is not very difficult to create templates for the informal
BON charts, using some of the more advanced word processors. With textual
BON it is then possible to mix the interactive input of charts with automatic
generation from information stored elsewhere.
The formal description also gives the reader a second chance to resolve
possible unclarities that always lurk in natural language descriptions. Nothing
can compensate for the precision of a formal notation when it comes to
communicating the difficult cases unambiguously.
On the other hand, a language grammar is much harder to read to get a general
overview of a notation than are typical examples of language usage. For this
reason, and since a fair portion of the BON textual notation has not been shown
elsewhere in the book, the next appendix will provide the interested reader with
textual equivalents to some of the graphical BON diagrams presented earlier.
The textual version of BON does not include any means for describing the
layout of diagrams. This would require an independent set of concepts largely
orthogonal to the BON structural elements. Such a language may emerge later
as a result of experience with case tools supporting BON.

349
350 BON TEXTUAL GRAMMAR

A.2 THE SYNTAX NOTATION

We will present the BON textual grammar in an extended BNF (Backus Naur
Form), where the extensions are very close to the ones used in [Meyer 1992a].
The syntax notation is based on the following concepts.
Any syntactically meaningful part of a BON textual specification, such as a
cluster, a class, or an assertion, is called a component. The structure of all
components of a certain category is described by a construct, and an individual
component conforming to this description is called a specimen of the construct.
Each construct has a unique construct name, which is a single word in roman
font starting with a capital letter. For example, Class_chart, Feature_clause, and
Object_group are construct names, and the corresponding specimens are any
individual class chart, feature clause, etc., that may be built according to the rules
specified by the grammar.
Every construct is either terminal or non-terminal. A specimen of a terminal
construct is called a lexical element or a token. The set of tokens make up the
basic vocabulary which may be used to construct sentences in the language, and
their internal structure is not described by the grammar. The set of tokens are
either considered known a priori, or else described separately (usually through
regular expressions applied to sequences of individual characters, or by informal
language).
Non-terminals, on the other hand, are described in terms of other constructs,
either terminal or non-terminal. Such a description is called a production for the
construct, and has the following form:

∆
Construct = right-hand-side

By convention, every non-terminal construct appears (through its construct

name) as the left-hand side of exactly one production. Terminal constructs, on
the other hand, may only appear on the right-hand side (by definition). The
symbol =∆ means “is defined as”.
The right-hand side of a production specifies the structure of the left-hand
construct, and since every non-terminal construct has a production attached, the
corresponding specimen can always be recursively decomposed into sequences
of tokens.
It is not always possible to tell whether a construct is terminal or non-terminal
without checking if it occurs as the left-hand side of a production or not.
However, two common token types are written using different typography for
easy identification: keywords (in lower case boldface) and fixed operators
(enclosed in double quotes). There are three basic types of production, which
have the following forms.
 THE SYNTAX NOTATION 351

Aggregate
Defines the construct as a fixed sequence of construct parts. One or more
elements in the sequence may be marked as optional by enclosing them in square
brackets. For example:

Parenthesized ∆ "(" Expression ")"

=
∆
Inheritance_relation = Child inherit [ "{" Multiplicity "}" ]
Parent [ Semantic_label ]

defines Parenthesized as a left parenthesis followed by an Expression followed

by a right parenthesis, and Inheritance_relation as a Child construct followed by
the keyword inherit, then an optional multiplicity part (Multiplicity enclosed in
braces), then a Parent construct, then an optional Semantic_label.

Choice
Defines the construct as one of a fixed number of alternative constructs. It is
written as a non-empty sequence of constructs separated by vertical bar. The
production

∆
Expression = Quantification | Call | Operator_expression | Constant

therefore means that an Expression is a Quantification, or a Call, or an

Operator_expression, or a Constant.

Repetition
Defines the construct as a variable length sequence of specimens of another
construct, possibly separated (if more than one element) by a given separator.
The separator (if any) may be either terminal or non-terminal. A repetition right-
hand side is written in one of the two forms below:
{ Element_construct Separator_construct … }

{ Element_construct Separator_construct … }+

The first form signals that the sequence may be empty, while the second requires
at least one element. Omitting the separator construct means that multiple
elements are concatenated without separators in this type of sequence. Below
are some examples. The first production defines an Index_list as a sequence of
one or more Index_clause, separated by a semicolon. The second defines
352 BON TEXTUAL GRAMMAR

Features as a sequence of one or more Feature_clause without any separator.

The third production defines a Dynamic_ref as zero or more Group_prefix
followed by a Dynamic_component_name.

Index_list ∆ { Index_clause ";" … }+

=
∆
Features = { Feature_clause … }+
∆
Dynamic_ref = { Group_prefix … } Dynamic_component_name

With these preliminaries, we are now ready to give the full syntax
specification of the BON textual notation. The grammar is defined in the
following sections and then concluded by a discussion of the lexical components,
summing up the keywords and operators used.

A.3 BON SPECIFICATION

∆
Bon_specification = { Specification_element … }+
∆ Informal_chart | Class_dictionary |
Specification_element =
Static_diagram | Dynamic_diagram |
Notational_tuning

A.4 INFORMAL CHARTS

∆
Informal_chart = System_chart | Cluster_chart | Class_chart |
Event_chart | Scenario_chart | Creation_chart
∆
Class_dictionary = dictionary System_name
{ Dictionary_entry … }+
end
∆
Dictionary_entry = class Class_name cluster Cluster_name
description Manifest_textblock

System_chart ∆ system_chart System_name

=
[ indexing Index_list ]
[ explanation Manifest_string ]
[ part Manifest_string ]
[ Cluster_entries ]
end
Cluster_entries ∆ { Cluster_entry … }+
=
Cluster_entry ∆ cluster Cluster_name description Manifest_textblock
=
System_name ∆ Identifier
=
 INFORMAL CHARTS 353

Index_list ∆ { Index_clause ";" … }+

=
Index_clause ∆ Identifier ":" Index_term_list
=
Index_term_list ∆ { Index_string "," … }+
=
Index_string ∆ Manifest_string
=

Cluster_chart ∆ cluster_chart Cluster_name

=
[ indexing Index_list ]
[ explanation Manifest_string ]
[ part Manifest_string ]
[ Class_entries ]
[ Cluster_entries ]
end
Class_entries ∆ { Class_entry … }+
=
Class_entry ∆ class Class_name description Manifest_textblock
=
Cluster_name ∆ Identifier
=

∆
Class_chart = class_chart Class_name
[ indexing Index_list ]
[ explanation Manifest_string ]
[ part Manifest_string ]
[ inherit Class_name_list ]
[ query Query_list ]
[ command Command_list ]
[ constraint Constraint_list ]
end
∆
Query_list = { Manifest_string "," … }+
∆
Command_list = { Manifest_string "," … }+
∆
Constraint_list = { Manifest_string "," … }+
∆
Class_name_list = { Class_name "," … }+
Class_name ∆ Identifier
=

∆
Event_chart = event_chart System_name
[ incoming | outgoing ]
[ indexing Index_list ]
[ explanation Manifest_string ]
[ part Manifest_string ]
[ Event_entries ]
end
∆
Event_entries = { Event_entry … }+
∆
Event_entry = event Manifest_string involves Class_name_list
354 BON TEXTUAL GRAMMAR

Scenario_chart ∆ scenario_chart System_name

=
[ indexing Index_list ]
[ explanation Manifest_string ]
[ part Manifest_string ]
[ Scenario_entries ]
end
Scenario_entries ∆ { Scenario_entry … }+
=
Scenario_entry ∆ scenario Manifest_string description Manifest_textblock
=

Creation_chart ∆ creation_chart System_name

=
[ indexing Index_list ]
[ explanation Manifest_string ]
[ part Manifest_string ]
[ Creation_entries ]
end
Creation_entries ∆ { Creation_entry … }+
=
Creation_entry ∆ creator Class_name creates Class_name_list
=

A.5 STATIC DIAGRAMS

Static_diagram ∆ static_diagram [ Extended_id ] [ Comment ]

=
component Static_block end
Extended_id ∆ Identifier | Integer
=
Comment ∆ { Line_comment New_line … }+
=
Line_comment ∆ "− −" Simple_string
=
Static_block ∆ { Static_component … }
=
Static_component ∆ Cluster |
=
Class |
Static_relation

∆
Cluster = cluster Cluster_name
[ reused ] [ Comment ]
[ Cluster_components ]
∆
Cluster_components = component Static_block end
∆
Class = [ root | deferred | effective ]
class Class_name [ Formal_generics ]
[ reused ] [ persistent ] [ interfaced ] [ Comment ]
[ Class_interface ]
∆
Static_relation = Inheritance_relation | Client_relation
 STATIC DIAGRAMS 355

Inheritance_relation ∆ Child inherit [ "{" Multiplicity "}" ]

=
Parent [ Semantic_label ]
Client_relation ∆ Client client [ Client_entities ] [ Type_mark ]
=
Supplier [ Semantic_label ]
Client_entities ∆ "{" Client_entity_expression "}"
=
Client_entity_expression ∆ Client_entity_list | Multiplicity
=
Client_entity_list ∆ { Client_entity "," … }+
=
Client_entity ∆ Feature_name | Supplier_indirection | Parent_indirection
=
Supplier_indirection ∆ [ Indirection_feature_part ":" ] Generic_indirection
=
Indirection_feature_part ∆ Feature_name | Indirection_feature_list
=
Indirection_feature_list ∆ "(" Feature_name_list ")"
=
Parent_indirection ∆ "−>" Generic_indirection
=

Generic_indirection ∆ Formal_generic_name | Named_indirection

=
Named_indirection ∆ Class_name "[" Indirection_list "]"
=
Indirection_list ∆ { Indirection_element "," … }+
=
Indirection_element ∆ "..." | Named_indirection
=
Type_mark ∆ ":" | ":{" | Shared_mark
=
Shared_mark ∆
= ":" "(" Multiplicity ")"

∆
Child = Static_ref
∆
Parent = Static_ref
∆
Client = Static_ref
∆
Supplier = Static_ref
∆
Static_ref = { Cluster_prefix … } Static_component_name
∆
Cluster_prefix = Cluster_name "."
∆
Static_component_name = Class_name | Cluster_name
∆
Multiplicity = Integer
∆
Semantic_label = Manifest_string

A.6 CLASS INTERFACE DESCRIPTION

∆
Class_interface = [ indexing Index_list ]
[ inherit Parent_class_list ]
Features
[ invariant Class_invariant ]
end
∆
Class_invariant = Assertion
∆
Parent_class_list = { Class_type ";" … }+
∆
Features = { Feature_clause … }+
356 BON TEXTUAL GRAMMAR

Feature_clause ∆ feature [ Selective_export ]

=
[ Comment ]
Feature_specifications
Feature_specifications ∆ { Feature_specification … }+
=
Feature_specification ∆ [ deferred | effective | redefined ]
=
Feature_name_list [ Type_mark Type ]
[ Rename_clause ]
[ Comment ]
[ Feature_arguments ]
[ Contract_clause ]

∆
Contract_clause = Contracting_conditions end
∆
Contracting_conditions = Precondition | Postcondition | Pre_and_post
∆
Precondition = require Assertion
∆
Postcondition = ensure Assertion
∆
Pre_and_post = Precondition Postcondition

∆
Selective_export = "{" Class_name_list "}"
∆
Feature_name_list = { Feature_name "," … }+
∆
Feature_name = Identifier | Prefix | Infix
∆
Rename_clause = "{" Renaming "}"
∆
Renaming = "^" Class_name "." Feature_name
∆
Feature_arguments = { Feature_argument … }+
∆
Feature_argument = "−>" [ Identifier_list ":" ] Type
∆
Identifier_list = { Identifier "," … }+
∆
Prefix = prefix ’"’ Prefix_operator ’"’
∆
Infix = infix ’"’ Infix_operator ’"’
∆
Prefix_operator = Unary | Free_operator
∆ Binary | Free_operator
Infix_operator =

∆
Formal_generics = "[" Formal_generic_list "]"
∆
Formal_generic_list = { Formal_generic "," … }+
∆
Formal_generic = Formal_generic_name [ "−>" Class_type ]
∆
Formal_generic_name = Identifier
∆
Class_type = Class_name [ Actual_generics ]
∆
Actual_generics = "[" Type_list "]"
∆
Type_list = { Type "," … }+
∆
Type = Class_type | Formal_generic_name
 CLASS INTERFACE DESCRIPTION 357

Unary ∆ delta | old | not | "+" | "−"

=
Binary ∆ "+" | "−" | "*" | "/" |
=
"<" | ">" | "<=" | ">=" |
"=" | "/=" | "//" | "\\" | "^" |
or | xor | and | "−>" | "<−>" | member_of | ":"

A.7 FORMAL ASSERTIONS

∆
Assertion = { Assertion_clause ";" … }+
∆
Assertion_clause = Boolean_expression | Comment
∆
Boolean_expression = Expression
∆
Expression = Quantification | Call | Operator_expression | Constant
∆
Quantification = Quantifier Range_expression [ Restriction ] Proposition
∆
Quantifier = for_all | exists
∆
Range_expression = { Variable_range ";" … }+
∆
Restriction = such_that Boolean_expression
∆
Proposition = it_holds Boolean_expression
∆
Variable_range = Member_range | Type_range
∆
Member_range = Identifier_list member_of Set_expression
Type_range ∆ Identifier_list ":" Type
=

Call ∆ [ Parenthesized_qualifier ] Call_chain

=
Parenthesized_qualifier ∆ Parenthesized "."
=
Call_chain ∆ { Unqualified_call "." … }+
=
Unqualified_call ∆ Identifier [ Actual_arguments ]
=
Actual_arguments ∆ "(" Expression_list ")"
=
Expression_list ∆ { Expression "," … }+
=
Operator_expression ∆ Parenthesized | Unary_expression | Binary_expression
=
Parenthesized ∆ "(" Expression ")"
=

∆
Unary_expression = Prefix_operator Expression
∆
Binary_expression = Expression Infix_operator Expression
∆
Set_expression = Enumerated_set | Call | Operator_expression
∆
Enumerated_set = "{" Enumeration_list "}"
∆
Enumeration_list = { Enumeration_element "," … }+
∆
Enumeration_element = Expression | Interval
∆
Interval = Integer_interval | Character_interval
∆
Integer_interval = Integer_constant ".." Integer_constant
∆
Character_interval = Character_constant ".." Character_constant
358 BON TEXTUAL GRAMMAR

Constant ∆ Manifest_constant | Current | Void

=
∆ Boolean_constant | Character_constant
Manifest_constant = |
Integer_constant | Real_constant |
Manifest_string
Sign ∆ "+" | "−"
=
Boolean_constant ∆ true | false
=
Character_constant ∆ "’" Character "’"
=
Integer_constant ∆ [ Sign ] Integer
=
Real_constant ∆ [ Sign ] Real
=
Manifest_textblock ∆ String_begin String String_end
=
String ∆ { Simple_string New_line … }+
=
Manifest_string ∆ String_begin Simple_string String_end
=

A.8 DYNAMIC DIAGRAMS

∆
Dynamic_diagram = dynamic_diagram [ Extended_id ] [ Comment ]
component Dynamic_block end
∆
Dynamic_block = { Dynamic_component … }
Dynamic_component ∆ Scenario_description |
=
Object_group |
Object_stack |
Object |
Message_relation

∆
Scenario_description = scenario Scenario_name [ Comment ]
action Labeled_actions end
∆
Labeled_actions = { Labeled_action … }+
∆
Labeled_action = Action_label Action_description
∆
Action_label = Manifest_string
∆
Action_description = Manifest_textblock
∆
Scenario_name = Manifest_string

∆
Object_group = [ nameless ] object_group Group_name [ Comment ]
[ Group_components ]
∆
Group_components = component Dynamic_block end
∆
Object_stack = object_stack Object_name [ Comment ]
∆
Object = object Object_name [ Comment ]
 DYNAMIC DIAGRAMS 359

Message_relation ∆ Caller calls Receiver [ Message_label ]

=
Caller ∆ Dynamic_ref
=
Receiver ∆ Dynamic_ref
=
Dynamic_ref ∆ { Group_prefix … } Dynamic_component_name
=
Group_prefix ∆ Group_name "."
=
Dynamic_component_name ∆ Object_name | Group_name
=
Object_name ∆ Class_name [ "." Extended_id ]
=
Group_name ∆ Extended_id
=
Message_label ∆ Manifest_string
=

A.9 NOTATIONAL TUNING

This will be explained in the next section.

∆
Notational_tuning = Change_string_marks |
Change_concatenator |
Change_prefix
∆
Change_string_marks = string_marks Manifest_string Manifest_string
∆
Change_concatenator = concatenator Manifest_string
Change_prefix ∆ keyword_prefix Manifest_string
=

A.10 LEXICAL COMPONENTS

We conclude this chapter with a discussion of the lexical components, which are
used to form BON textual descriptions in accordance with the grammar defined
in the preceding sections. These components are the terminal constructs that do
not appear as the left-hand side of any production in the grammar, and therefore
need to be described separately.

Identifiers
The Identifier construct is defined as a sequence of alphanumeric characters
including underscore. An identifier must begin with an alphanumeric character
and must not end with an underscore (whose purpose really is to mimic word
separation). Letter case is not significant, but using consistent style rules is
important.
The recommended BON standard is to use all upper case names for class and
cluster names, all lower case for feature names, and lower case beginning with a
capital for object groups and constants values. We also strongly recommed
using underscore for word separation rather than, for example, in-word
capitalization, since this greatly enhances readability.
360 BON TEXTUAL GRAMMAR

Free operators
The Free_operator construct represents feature names used as infix and prefix
operations. Such operations may be textual keywords, such as the boolean and
and or, but are more often composed of special characters, like "+", "**", "=>",
etc.
The purpose is usually to make object-oriented expressions (which are always
feature calls in the end) look very similar to the formalisms used in some
discipline thus providing a more compact and readable notation for the problem
at hand. Since it is difficult to foresee exactly what operator combinations may
be needed, BON only defines the Free_operator construct as a sequence of non-
spacing printable characters that does not conflict with any of the predefined
ones. However, in practice, more restrictions are added by each development
environment.

Comments
Major analysis and design elements, such as static diagrams, clusters, classes,
object groups, etc., often need to have comments attached to them in order to
explain overall modeling aspects that have no natural place among the
constituent parts at lower levels.
Therefore, the BON textual notation recognizes comments to major
specification elements as part of the grammar, thereby encouraging the standard
placement of them. This also provides a parser with the possibility to check and
possibly enforce certain strategic descriptions. However, besides the places
recognized by the grammar, comments may be inserted anywhere in a BON
textual description, except inside strings.

Strings
The construct Simple_string is defined as any string of characters not containing
a New_line character. The non-terminal construct Manifest_string is a
Simple_string enclosed in the pair of terminals String_begin and String_end.
Similarly, the non-terminal Manifest_textblock is a sequence of Simple_string
separated by New_line and enclosed by the same pair of terminal constructs (see
the grammar above).
These delimiters are defined by default as a string containing one double quote
character. The character sequence "show some class, don’t treat me like an
object" is then interpreted as a Manifest_string. However, to facilitate the
accommodation of double quotes inside strings without having to insert escape
characters, the delimiting strings may be changed (often to some control
characters in connection with automatic processing).
 LEXICAL COMPONENTS 361

BON also defines a lexical Concatenator construct. If a Concatenator is found

inside a Simple_string, it is removed along with all characters (including
New_line) up to and including the next Concatenator construct. This makes it
possible to embed formatting white space into strings for readability, without
making the formatting characters part of the strings.
The Concatenator construct is defined as a single backslash by default, but
may be changed by the user. It must not conflict with the string delimiters. An
example of its use is shown below.

"This is a long simple string, which has been broken into\

\ two lines for readability"

The basic constructs Integer, New_line, Character, and Real are not further
specified, since they may need different definitions depending on the
development environment.

Reserved words
Reserved words are terminal constructs which are predefined sequences of letters
only, and which cannot be used as identifiers by the user, since this might lead to
language ambiguities. The reserved words in BON consist of keywords and
predefined names. There are only three of the latter type: Current, Result, and
Void. The full list is shown in figure A.1.
In a sizable language, there is always the risk that some keywords steal
valuable name space from the user, and textual BON, being fairly expressive, is

action creator false not reused

and Current feature object root
calls deferred for_all object_group scenario
class delta incoming object_stack scenario_chart
class_chart description indexing old static_diagram
client dictionary infix or string_marks
cluster dynamic_diagram inherit outgoing such_that
cluster_chart effective interfaced part system_chart
command end invariant persistent true
component ensure involves prefix Void
concatenator event it_holds query xor
constraint event_chart keyword_prefix redefined
creates exists member_of require
creation_chart explanation nameless Result

Figure A.1 BON reserved words

362 BON TEXTUAL GRAMMAR

no exception. To counter this disadvantage, BON defines a terminal construct

Keyword_prefix, which is empty by default.
By defining Keyword_prefix as the string "$", for example, we may change
the syntax of BON so all keywords now need to be prefixed by a dollar sign, thus
freeing all the corresponding normal words for use in specification of the system
under development.

Special symbols
Finally, we collect the complete set of special symbols used in BON with an
overview of their meaning (figure A.2). Each of them has been described earlier
in the book. The ones marked as operators (except for the type operator ":") can
be viewed as class features of infix form that may be redefined by descendant
classes.

Conclusion
The BON textual notation is a full specification language for object-oriented
system designs, whose purpose is threefold:
• It can be used to communicate exact specifications between various tools
and environments, thus taking advantage of the advances in many
independent areas of presentation.
• It can be used for better understanding of the concepts underlying the
graphical notation and for settling ambiguities. With today’s widely
available parser generator utilities, the task of writing a parser for the
language becomes easy.
• It provides a means of storing and updating a specification in a simple
way, using standard text editors, which can serve as an alternative to a
dedicated case tool. It may be feasible to copy small whiteboard diagrams
with pencil on paper in connection with design sessions, but maintaining
larger specifications requires more. Anybody who has experienced the
pain of trying to keep evolving graphical figures up to date without strong
automatic support knows only too well what we are talking about.
Finally, regarding the different presentations that may be generated from a BON
textual description, we have not tried to cover graphical layout in the textual
language. The basic graphical appearance of each textual concept has been
defined earlier in this book, along with validity constraints and rules for how
relational arrows may be combined, labels be positioned, etc. But what valid
alternative to choose is left to the strategies of the individual case tool.
 LEXICAL COMPONENTS 363

SYMBOL NAME USE

−− double dash Introduces comments

’ single quote Encloses character constants
" double quote Encloses prefix and infix operator names
, comma General element separator
; semicolon Separator for parent lists, assertion
clauses, and indexing clauses
() parentheses Grouping of expressions, multiplicity
[] square brackets Encloses generic parameters
{} braces Encloses restricted export lists, renaming,
enumerated sets
+−*/ plus, minus, times, Arithmetic operators
division
// \\ double slash, double Integer division, modulo operators
backslash
^ up arrow Power operator, renaming
< > less than, greater than Relational operators
<= >= less than or equal, Relational operators
greater than or equal
= /= equal, not equal Equality and non-equality
−> right arrow, implies Feature arguments, constrained genericity,
logical implication
<−> equivalence Logical equivalence
. dot Feature calls, renaming, relational
references, object_id
.. double dot Interval marker
: colon Type mark, type operator, index separator
:{ aggregate mark Indicates aggregate supplier

Figure A.2 BON special symbols

Appendix B:
BON textual examples

B.1 BON SPECIFICATION

In this appendix, we will show the corresponding textual version of some of the
graphical BON descriptions presented earlier in the book. This will give
interested readers a feeling for the textual language without the need to decode
everything from the grammar given in the previous appendix.
A BON textual specification consists of a sequence of one or more of the
following types of specification element, which may be given in any order:
• Informal chart (system, cluster, class, event, scenario, creation)
• Class dictionary
• Static diagram
• Dynamic diagram
• Notational tuning
The last type of specification does not address the system being modeled, but
may be used to dynamically alter some lexical details regarding the terminal
symbols recognized by textual BON. Three operations are available for
language tuning:
• string_marks to change the default double quote string delimiter
• concatenator to change the default backslash string concatenator
• keyword_prefix to alter the names of all textual BON keywords into new
names, where each old name is prefixed by a given string.
The class dictionary is meant to be generated automatically from the class charts
and / or class interfaces, so no corresponding BON chart format is specified for it.

364
 BON SPECIFICATION 365

We will look at the informal charts and the static and dynamic diagrams in order,
and discuss some of the choices made in the design of textual BON.

B.2 INFORMAL CHARTS

The informal charts are enclosed in <name>_chart … end blocks, and the
keywords delimiting their respective fields correspond closely to the labels used
in the graphical forms. The class charts for CITIZEN and NOBLEPERSON
(from chapter 3) are shown in figure B.1, and two cluster charts are in figure B.2.

class_chart CITIZEN
indexing
cluster: "CIVIL_STATUS " ;
created: "1993-03-15 jmn";
revised: "1993-05-12 kw"
explanation
"Person born or living in a country"
part "1/1 "
query
"Name " , "Sex " , "Age " , "Single " , "Spouse " , "Children " , "Parents " ,
"Impediment to marriage"
command
"Marry " , "Divorce "
constraint
"Each citizen has two parents.",
"At most one spouse allowed.",
"May not marry children or parents or person of same sex.",
"Spouse’s spouse must be this person.",
"All children, if any, must have this person among their parents."
end
class_chart NOBLEPERSON
indexing
cluster: "CIVIL_STATUS " ;
created: "1993-03-15 jmn";
revised: "1993-05-12 kw", "1993-12-10 kw"
explanation
"Person of noble rank"
part "1/1 "
inherit CITIZEN
query
"Assets " , "Butler "
constraint
"Enough property for independence.",
"Can only marry other noble person.",
"Wedding celebrated with style.",
"Married nobility share their assets and must have a butler."
end

Figure B.1 Class charts: types of citizen (cf. figure 3.3)

366 BON TEXTUAL EXAMPLES

cluster_chart ORGANIZATION
indexing
author: "Kim waldén", "Jean-marc nerson";
keywords: "organization " , "staff "
explanation
"Handles all major events occurring during the\
\ organization and completion of a conference."
part "1/1 "
class CONFERENCE
description
"The root class of the conference system."
class PROGRAM
description
"Information about the final conference program\
\ and its preparation."
class TIMETABLE
description
"Repository of scheduled events."
cluster COMMITTEES
description
"The committees engaged in the conference organization\
\ to take care of the technical and administrative parts."
end

cluster_chart COMMITTEES
indexing
cluster: "ORGANIZATION " ;
author: "Kim waldén", "Jean-marc nerson";
keywords: "committee " , "scientific board", "steering board"
explanation
"Groups all general and special types of committees."
part "1/1 "
class COMMITTEE
description
"General committee abstraction."
class STEERING_COMMITTEE
description
"Committee in charge of practical arrangements."
class PROGRAM_COMMITTEE
description
"Committee in charge of selecting technical contributions."
end

Figure B.2 Cluster chart (cf. figure 3.2)

It is possible to include new line characters verbatim in most of the descriptive

text fields of the charts, since these are defined as Manifest_textblock in the
grammar. However, we have chosen not to do so, because in a case tool with
window resizing capabilities we would need to reformat text fields to fit the
current window size anyway. Instead, we use the string concatenator "\" to build
simple strings from line formatted specification input.
 INFORMAL CHARTS 367

Descriptive text fields consisting of a number of semantically separable

phrases, such as query, command, and constraint, are syntactically defined as a
sequence of Manifest_string separated by a comma.

B.3 STATIC DIAGRAMS

Cluster structure
The next examples, figures B.3 and B.4, show the nesting of static components.

static_diagram First_system_breakdown
component
cluster CONFERENCE_MANAGEMENT_SYSTEM
component
cluster ORGANIZATION
cluster TECHNICAL_EVENTS
cluster PRINT_OUTS
cluster REGISTRATION
end
end

Figure B.3 First cluster definition sketch (cf. figure 9.5, upper part)

Cluster components are enclosed in component … end blocks, unless the

cluster is empty in which case the cluster name suffices. (An empty cluster will
be displayed graphically as an icon, but whether a non-empty cluster will be
iconized or not has to do with the presentation format, which is not addressed by
textual BON.)
As in the examples, static and dynamic diagrams may have names attached for
identification and readability. Since such names are not significant for the
modeled system (unlike, for example, class and cluster names), we define them
as Extended_id (either Identifier or Integer). This may sometimes be practical
for automatically generated textual diagrams.

Class headers
Figure B.5 shows a set of annotated class headers. There is only one syntactic
Class construct, so headers are described by omitting the Class_interface part.

Inheritance relations
The next examples show inheritance relations of various types. Multiplicity may
be specified by a number enclosed in braces, as seen in figure B.6, where the
class TRIPLE_INDEX inherits three times from class INDEX. The joined
368 BON TEXTUAL EXAMPLES

static_diagram Nested_data_structures
component
cluster DATA_STRUCTURES
component
cluster SORTING reused
component
class LINEAR_SORT
class QUICKSORT
class RADIX_SORT
class INSERTION_SORT
class SHELLSORT
class HEAPSORT
class TOPOLOGICAL_SORT
class SORT_MERGE
end
cluster GRAPHS
component
deferred class GRAPH
class WEIGHTED_GRAPH
class DIRECTED_GRAPH
class UNDIRECTED_GRAPH
class BIPARTITE_GRAPH
class DENSE_GRAPH
end
end
end

Figure B.4 A nested data structure cluster (cf. figure 3.16)

static_diagram
component
root class CONTROL_PANEL
class TRANSACTION persistent
class MAILER interfaced
class HASH_TABLE [T, U]
deferred class FLYING_OBJECT
effective class HELICOPTER
class INPUT reused interfaced
class VECTOR [G] reused
deferred class SESSION interfaced
end

Figure B.5 Annotated class headers (cf. figure 3.15)

 STATIC DIAGRAMS 369

relational arrows in figure 4.1, on the other hand, have no correspondence in

textual BON. Figure B.7 shows the inheritance between classes and clusters.
We use dot notation on the cluster names involved to refer to classes residing
inside nested clusters, as seen in the figure. This makes it possible to have local
name spaces for classes in different clusters at the analysis level, and also

static_diagram
component
class COLD_STORE
class FREEZER
class REFRIGERATOR
class INDEX
class TRIPLE_INDEX
deferred class FLYING_OBJECT
effective class AIRCRAFT
effective class ROCKET
class SPACE_SHUTTLE
deferred class VEHICLE
effective class BICYCE
effective class BOAT
effective class CAR
REFRIGERATOR inherit COLD_STORE
FREEZER inherit COLD_STORE
TRIPLE_INDEX inherit {3} INDEX
AIRCRAFT inherit FLYING_OBJECT
ROCKET inherit FLYING_OBJECT
SPACE_SHUTTLE inherit AIRCRAFT
SPACE_SHUTTLE inherit ROCKET
BICYCLE inherit VEHICLE
BOAT inherit VEHICLE
CAR inherit VEHICLE
end

Figure B.6 Different type of inheritance (cf. figure 4.1)

documents more clearly what cluster interdependencies we have in our diagrams.

If the implementation language supports renaming of classes on a cluster basis,
possible name clashes may easily be resolved later.

Client relations
Labels attached to client relations are enclosed in braces, as seen in figure B.8,
which contains a number of compacted generic links. Here (as in figure B.7) we
have split the figure into multiple diagrams. The third diagram in figure B.8 had
to be separated to avoid name clashes on the classes involved, which would
make the specification ambiguous. The first and second diagrams were
separated to obtain the grouping of the static components contained in the figure.
370 BON TEXTUAL EXAMPLES

static_diagram 1
component
cluster PARENTS
component
class A
class B
end
cluster CHILDREN
component
class C
class D
class E
end
CHILDREN.C inherit PARENTS.A
CHILDREN.C inherit PARENTS.B
CHILDREN.D inherit PARENTS.A
CHILDREN.D inherit PARENTS.B
CHILDREN.E inherit PARENTS.A
CHILDREN.E inherit PARENTS.B
end
static_diagram 2
component
cluster PARENTS component … end
cluster CHILDREN component … end
CHILDREN inherit PARENTS.A
CHILDREN inherit PARENTS.B
end
static_diagram 3
component
cluster PARENTS component … end
cluster CHILDREN component … end
CHILDREN.C inherit PARENTS
CHILDREN.D inherit PARENTS
CHILDREN.E inherit PARENTS
end
static_diagram 4
component
cluster PARENTS component … end
cluster CHILDREN component … end
CHILDREN inherit PARENTS
end

Figure B.7 Recursive abstraction (cf. figure 4.3)

Although we have had no example so far in the book of nested generic

derivation, this does occur in practice not too infrequently. The BON notation
therefore supports nested generic type parameters in compacted client relations,
so that several levels of indirection through standard container classes can be
removed and diagrams concentrated on the essential concepts for the modeling.
 STATIC DIAGRAMS 371

static_diagram 1
component
class BASEBALL_CARD
class SON
class LIST [T] reused
class PARENT
class ACCOUNT
SON client {expenses} LIST
PARENT client {assets} LIST
LIST client {(first, last): T} BASEBALL_CARD
LIST client {(first, last): T} ACCOUNT
end
static_diagram 2
component
class NURSE
class PATIENT
class SURGEON
class OPERATION
class TABLE [U, V−>KEY] reused
class PATIENT_ID
class DATE
NURSE client {patients} TABLE
SURGEON client {duties} TABLE
TABLE client {item: U} PATIENT
TABLE client {item: U} OPERATION
TABLE client {key: V} PATIENT_ID
TABLE client {key: V} DATE
end
static_diagram 3
component
class SON
class BASEBALL_CARD
class PARENT
class ACCOUNT
class NURSE
class PATIENT
class SURGEON
class OPERATION
SON client {expenses: LIST […]} BASEBALL_CARD
PARENT client {assets: LIST […]} ACCOUNT
NURSE client {patients: LIST […, PATIENT_ID]} PATIENT
SURGEON client {duties: LIST […, DATE]} OPERATION
end

Figure B.8 Multiple generic derivation (cf. figure 4.14)

Each occurrence of […] in the label of a compacted link refers to the supplier
class. For example,
FIGURE client {SET [SET […]]} POINT
expresses that class FIGURE has some entity of type SET [SET [POINT]].
372 BON TEXTUAL EXAMPLES

Figure B.9 shows some indirect client dependencies resulting from the generic
derivation of parent classes. The braces enclose the client entity part, which
describes what causes the relation. The entity part may express generic
indirection through a parent class as in figure B.9, role multiplicity as in
figure B.10, and multiple feature labels as in figure B.11. Aggregation relations
are expressed by a corresponding type mark between the client entity part and
the supplier.

static_diagram 1
component
class SEQUENCE [T] reused
class BYTE
class FILE
FILE inherit SEQUENCE
SEQUENCE client {−> […]} BYTE
end
static_diagram 2
component
class FILE
class BYTE
FILE client {−> SEQUENCE […]} BYTE
end

Figure B.9 Generic client relation through inheritance (cf. figure 4.15)

static_diagram
component
class VISITING_ALIEN
class LANDING_DOCUMENT
class APARTMENT
class ROOM
class HOUSE
class ARCHITECT
VISITING_ALIEN client {2} LANDING_DOCUMENT
APARTMENT client {3} :{ ROOM
HOUSE client {1} ARCHITECT
ARCHITECT client {3} HOUSE
end

Figure B.10 Multiplicity markers (cf. figure 4.16)

Since role multiplicity maps to a fixed number of static relationships between

a client and a supplier, each playing a different role during execution, the
corresponding numbers are just a replacement for client labels. It is therefore
syntactically natural to place the corresponding multiplicity numbers in the same
 STATIC DIAGRAMS 373

static_diagram
component
class VISITING_ALIEN
class LANDING_DOCUMENT
class APARTMENT
class ROOM
class HOUSE
class ARCHITECT
VISITING_ALIEN client {immigration_form, customs_form} LANDING_DOCUMENT
APARTMENT client {kitchen, bedroom: SET […], living_room: SET […]} :{ ROOM
HOUSE client {designer} ARCHITECT
ARCHITECT client {summer_house, winter_cottage, main_residence} HOUSE
end

Figure B.11 Naming rather than enumerating roles (cf. figure 4.17)

position as the labels (enclosed in braces, and before the type mark).
Multiple sharing of instances, on the other hand, corresponds to one entity
being dynamically attached to a fixed number of possible instances during
execution. Therefore, such instance multiplicity is instead enclosed in ordinary
parentheses and put after the type mark as can be seen in figure B.12.
In diagram 6 of this figure, we have an example where both types of
multiplicity are used simultaneously to express that a PC class has two client
relations, each dynamically sharing three instances of class FILE_SERVER. The
next example shows the static relations between elements inside and between
different clusters (figure B.13).

static_diagram 1 static_diagram 4
component component
class PC class PC
class FILE_SERVER class FILE_SERVER
PC client {2} : (1) FILE_SERVER PC client : (2) FILE_SERVER
end end
static_diagram 2 static_diagram 5
component component
class PC class PC
class FILE_SERVER class FILE_SERVER
PC client {server1} : (1) FILE_SERVER PC client {server} : (2) FILE_SERVER
PC client {server2} : (1) FILE_SERVER end
end static_diagram 6
static_diagram 3 component
component class PC
class PC class FILE_SERVER
class FILE_SERVER PC client {2} : (3) FILE_SERVER
PC client {server1, server2} : (1) FILE_SERVER end
end

Figure B.12 Different ways to express sharing (cf. figure 4.19)

374 BON TEXTUAL EXAMPLES

static_diagram Graphical_editor
− − This diagram shows the basic design of a graphical editor with grouping facilities.
component
cluster DISPLAY_ELEMENTS − − Contains different display objects
component
cluster GEOMETRIC_FIGURES − − Subcluster with geometric figures
component
class LINE
class ELLIPSE
class CIRCLE
class RECTANGLE
class SQUARE
CIRCLE inherit ELLIPSE
SQUARE inherit RECTANGLE
end
deferred class FIGURE
class PIXEL_MAP
class TEXT
class GROUP
GEOMETRIC_FIGURES inherit FIGURE
end
deferred class DISPLAY_OBJECT
class POINT
DISPLAY_ELEMENTS inherit DISPLAY_OBJECT
DISPLAY_ELEMENTS.GROUP client {members: SET […]} :{ DISPLAY_OBJECT
DISPLAY_ELEMENTS.FIGURE client {points: SET […]} POINT
end

Figure B.13 Graphical objects with grouping (cf. figure 5.17)

Class interfaces
The textual notation for the technical events classes of the Conference case study
is shown in figure B.14. The textual form needs a few more delimiters to
become unambiguous than the corresponding graphical form. For example,
contract clauses must be terminated by end. Finally, we show a textual form of
the overall static architecture of the Conference study in figure B.15 and
figure B.16.

B.4 DYNAMIC DIAGRAMS

We conclude with some dynamic diagrams. Objects are represented graphically
by the respective class name optionally qualified by a parenthesized object id.
Usually only a few objects need to be qualified in graphical dynamic diagrams,
since the spatial context will be enough to identify most of them; this holds also
for multiple instances of the same class. However, since textual descriptions
have no positional information available, we always need the qualification
 DYNAMIC DIAGRAMS 375

static_diagram Technical_events invariant

component for_all p, q: PRESENTATION such_that
class REVIEW persistent p /= q it_holds p.code /= q.code and
feature p.title /= q.title
reviewer: PERSON end
score: VALUE class TUTORIAL persistent
comments: TEXT inherit PRESENTATION
invariant feature
score member_of { ’A’ .. ’D’} capacity: VALUE
end attendee_count: VALUE
class STATUS persistent prerequisite_level: VALUE
feature track: VALUE
received: DATE duration: DURATION
review_started: DATE effective accept
accepted: DATE effective reject
rejected: DATE end
final_received: DATE class PAPER_SESSION persistent
invariant inherit SESSION
received <= review_started; feature
review_started <= final_received; presentations: SET [PAPER]
accepted = Void or rejected = Void invariant
end for_all p member_of presentations it_holds
class PAPER persistent p.status.accepted /= Void
inherit end
PRESENTATION class SESSION
feature feature
copyright_transferred: BOOLEAN chair: PERSON
reviews: SET [REVIEW] code: VALUE
final_score: VALUE track: VALUE
award_best_paper start, end: DATE
transfer_copyright conference_room: VALUE
require invariant start < end
status.accepted /= Void end
ensure class TUTORIAL_SESSION persistent
copyright_transferred inherit
end SESSION
effective accept feature
effective reject lecture: TUTORIAL
end invariant
deferred class PRESENTATION lecture.status.accepted /= Void
feature end
code: VALUE PAPER inherit PRESENTATION
title: VALUE TUTORIAL inherit PRESENTATION
authors: SET [PERSON] PAPER_SESSION inherit SESSION
status: STATUS TUTORIAL_SESSION inherit SESSION
speakers: SET [PERSON] PAPER client REVIEW
deferred accept PRESENTATION client STATUS
ensure status.accepted /= Void end PAPER_SESSION client PAPER
deferred reject TUTORIAL_SESSION client TUTORIAL
ensure status.rejected /= Void end end

Figure B.14 Technical_events (cf. figure 9.31)

376 BON TEXTUAL EXAMPLES

static_diagram Conference_architecture
− − This diagram shows the overall architecture of the Conference Management system.
component
cluster ORGANIZATION
component
cluster COMMITTEES
component
class COMMITTEE
class ORGANIZING_COMMITTEE persistent
class TECHNICAL_COMMITTEE persistent
class PROGRAM_COMMITTEE persistent
ORGANIZING_COMMITTEE inherit COMMITTEE
TECHNICAL_COMMITTEE inherit COMMITTEE
PROGRAM_COMMITTEE inherit TECHNICAL_COMMITTEE
PROGRAM_COMMITTEE client {tutorial_committee} :{ TECHNICAL_COMMITTEE
end
root class CONFERENCE persistent
class PROGRAM persistent
class TIMETABLE persistent
CONFERENCE client {scientific_board, steering_board} :{ COMMITTEES
CONFERENCE client {program} :{ PROGRAM
CONFERENCE client {reminder} TIMETABLE
end
cluster REGISTRATION
component
class REGISTRATION persistent
class PERSON persistent
REGISTRATION client {attendee} PERSON
PERSON client {registration} REGISTRATION
end
cluster TECHNICAL_EVENTS
component
class SESSION
class PAPER_SESSION persistent
class TUTORIAL_SESSION persistent
deferred class PRESENTATION
class PAPER persistent
class TUTORIAL persistent
class REVIEW persistent
class STATUS persistent
PAPER_SESSION inherit SESSION
TUTORIAL_SESSION inherit SESSION
PAPER inherit PRESENTATION
TUTORIAL inherit PRESENTATION
PAPER_SESSION client {presentations: SET […]} PAPER
TUTORIAL_SESSION client {lecture} TUTORIAL
PAPER client {reviews: SET […]} REVIEW
PRESENTATION client {status} STATUS
end

Figure B.15 Complete static architecture, part 1 (cf. figure 9.32)

 DYNAMIC DIAGRAMS 377

cluster OUTPUT
component
cluster TEMPLATES
class LETTER_FORM
class BADGE_FORM
class STICKY_FORM
class INVOICE_FORM
class POSTER_FORM
class LIST_FORM
class EVALUATION_FORM
end
cluster PRINT_OUTS
component
class MAILING
class ADDRESS_LABEL
class CONFIRMATION_LETTER
class INVOICE
class ACCEPTANCE_LETTER
class REJECTION_LETTER
class AUTHOR_GUIDELINES
class POSTER_SIGN
class BADGE
class EVALUATION_SHEET
class ATTENDEE_LIST
class STATISTICS
end
deferred class PRINT_OUT
deferred class DOCUMENT_FORM
TEMPLATES inherit DOCUMENT_FORM
PRINT_OUTS inherit PRINT_OUT
PRINT_OUT client {layout} DOCUMENT_FORM
end
ORGANIZATION client OUTPUT
ORGANIZATION client TECHNICAL_EVENTS
ORGANIZATION client REGISTRATION
REGISTRATION client OUTPUT
REGISTRATION client TECHNICAL_EVENTS
TECHNICAL_EVENTS client REGISTRATION
TECHNICAL_EVENTS client OUTPUT
end

Figure B.16 Complete static architecture, part 2 (cf. figure 9.32)

whenever more than one object of the same class occur in a textual diagram.
Again we use dot notation for separation, and require such object names to be
suffixed by an Extended_id (either Identifier or Integer) for unique reference.
Figure B.17 shows an example where such qualification is needed.
We note that the objects GROUP.outer and GROUP.inner were qualified
already in figure 5.21, while the other object identities could all be inferred from
context. In the textual form, these objects had to be qualified by integer suffixes,
378 BON TEXTUAL EXAMPLES

dynamic_diagram Move_group
component
scenario "Scenario 2: Move example group"
action
"1" "Next outer group member requested"
"2" "Square asked to move"
"3" "Next square point requested"
"4" "Point asked to move"
"5" "Outer circle asked to move"
"6" "Next outer circle point requested"
"7" "Point asked to move"
"8" "Inner group asked to move"
"9" "Next inner group member requested"
"10" "Inner circle asked to move"
"11" "Next inner circle point requested"
"12" "Point asked to move"
"13" "Text asked to move"
end
object SQUARE
object CIRCLE.1
object CIRCLE.2
object TEXT
object GROUP.outer
object GROUP.inner
object SET.1
object SET.2
object SET.3
object SET.4
object SET.5
object_stack POINT.1
object_stack POINT.2
object_stack POINT.3
GROUP.outer calls SET.1 "1"
GROUP.outer calls SQUARE "2"
SQUARE calls SET.2 "3"
SQUARE calls POINT.1 "4"
GROUP.outer calls CIRCLE.1 "5"
CIRCLE.1 calls SET.3 "6"
CIRCLE.1 calls POINT.2 "7"
GROUP.outer calls GROUP.inner "8"
GROUP.inner calls SET.4 "9"
GROUP.inner calls CIRCLE.2 "10 "
CIRCLE.2 calls SET.5 "11 "
CIRCLE.2 calls POINT.3 "12 "
GROUP.inner calls TEXT "13"
end

Figure B.17 Move example group (cf. figure 5.21)

probably appended by a generating case tool.

Similarly, we need an identification for unnamed object groups in order to
refer to it by message relations. This is illustrated in figure B.19, where the
name Group was assigned to the nameless group.
 DYNAMIC DIAGRAMS 379

dynamic_diagram Claim_settlement
− − This dynamic diagram groups typical sequential subtasks.
component
scenario "Scenario 5: Settlement of claims for damages resulting from car accident"
action
"1-3 " "Owner obtains necessary statements and certificates\
\ from involved parties, fills in damage report, and\
\ sends it to insurance company."
"4-7 " "Insurance adjuster evaluates damage claims and sends\
\ settlement statement back to owner."
"8-9 " "Owner agrees on car rental and repair details based\
\ on settlement. "
end
object_group Accident_report
component
object WITNESS
object POLICE
object INSURANCE_INSPECTOR
end
object_group Evaluation
component
object INSURANCE
object STANDARD_PRICE_LIST
object APPROVED_GARAGES
end
object_group Repair
component
object RENTAL_COMPANY
object GARAGE
end
object OWNER
object INSURANCE_ADJUSTER
object DAMAGE_REPORT
object SETTLEMENT
OWNER calls Accident_report "1"
OWNER calls DAMAGE_REPORT "2"
OWNER calls INSURANCE_ADJUSTER "3"
OWNER calls SETTLEMENT "8"
OWNER calls Repair "9"
INSURANCE_ADJUSTER calls DAMAGE_REPORT "4"
INSURANCE_ADJUSTER calls Evaluation "5"
INSURANCE_ADJUSTER calls SETTLEMENT "6"
INSURANCE_ADJUSTER calls OWNER "7"
end

Figure B.18 Grouping into sequential subtasks (cf. figure 5.23)

380 BON TEXTUAL EXAMPLES

dynamic_diagram Evaluate_paper
component
scenario "Scenario 2: Accept or reject a paper and notify authors"
action
"1-2 " "A paper is selected"
"3-5 " "Acceptance or rejection date is entered"
"6-7 " "The first author of the paper is selected"
"8" "A notification letter is created"
"9-11 " "The letter is sent to first author"
end
nameless object_group Group
component
object ACCEPTANCE_LETTER
object REJECTION_LETTER
end
object LETTER_FORM
object_stack PERSON
object PROGRAM_COMMITTEE
object CONFERENCE
object_stack PAPER
object STATUS
CONFERENCE calls PROGRAM_COMMITTEE "1"
PROGRAM_COMMITTEE calls PAPER "2, 3, 6"
PROGRAM_COMMITTEE calls PERSON "7"
PROGRAM_COMMITTEE calls Group "8, 9"
PAPER calls STATUS "4"
Group calls LETTER_FORM "10"
Group calls Outside_world "11: Letter printed"
Outside_world calls STATUS "5: User input"
end

Figure B.19 Paper evaluation (cf. figure 9.24)

Appendix C:
BON quick reference

BON deliverables: dependencies

LEGEND:
intellectual help to create/update
Class dictionary
possible (partial ) automatic SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS
SSSSSSSSSS

SSSSSSSSSS

generation SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

SYSTEM CHART CLUSTER CHART CLASS CHART Class interface

SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSS SSSSSSSSSSSS Class text
SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSS SSSSSSSSSSSS
(source code)
SSSSSSSSSSSSSS SSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS SSSSSSSSSSSS SSSSSSSSSSSS

Static architecture

STATIC MODEL
DYNAMIC MODEL
EVENT CHART SCENARIO CHART Object scenario CREATION CHART
SSSSSSSSSSSSS SSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS
1 SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

SSSSSSSSSSSSS SSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

2
SSSSSSSSSSSSS SSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

381
382 BON QUICK REFERENCE

BON deliverables: description (◊ indicates the most important ones)

System chart SYSTEM CHART

Definition of system and list of associated clusters. Only one SSSSSSSSSSSSSS SSSSSSSSSSSSSS

system chart per project; subsystems are described through

SSSSSSSSSSSSSS SSSSSSSSSSSSSS

SSSSSSSSSSSSSS SSSSSSSSSSSSSS

corresponding cluster charts.

Cluster charts CLUSTER CHART
Definition of clusters and lists of associated classes and SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

subclusters, if any. A cluster may represent a full subsystem

SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

SSSSSSSSSSSSSSS SSSSSSSSSSSSSSS

or just a group of classes.

Class charts CLASS CHART
Definition of analysis classes in terms of commands, queries, SSSSSSSSSSSS SSSSSSSSSSSS

and constraints, understandable by domain experts and non-

SSSSSSSSSSSS SSSSSSSSSSSS

SSSSSSSSSSSS SSSSSSSSSSSS

technical people.
Class dictionary Class dictionary
Alphabetically sorted list of all classes in the system, showing SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

the cluster of each class and a short description. Should be

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

SSSSSSSSSS SSSSSSSSSS SSSSSSSSSS

generated automatically from the class charts/interfaces.

◊ Static architecture Static architecture
Set of diagrams representing possibly nested clusters, class
headers, and their relationships. Bird’s eye view of the
system (zoomable).

◊ Class interfaces Class interface

Typed definitions of classes with feature signatures and
formal contracts. Detailed view of the system.

Creation charts CREATION CHART

List of classes in charge of creating instances of other classes. SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

Usually only one per system, but may be repeated for

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

subsystems if desirable.
Event charts EVENT CHART
Set of incoming external events (stimuli) triggering SSSSSSSSSSSSS SSSSSSSSSSSSS

interesting system behavior and set of outgoing external

SSSSSSSSSSSSS SSSSSSSSSSSSS

SSSSSSSSSSSSS SSSSSSSSSSSSS

events forming interesting system responses. May be

repeated for subsystems.
◊ Scenario charts SCENARIO CHART
List of object scenarios used to illustrate interesting and SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

representative system behavior. Subsystems may contain

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

SSSSSSSSSSSSSSSS SSSSSSSSSSSSSSSS

local scenario charts.

◊ Object scenarios Object scenario
Dynamic diagrams showing relevant object communication 1

for some or all of the scenarios in the scenario chart. 2

 BON QUICK REFERENCE 383

BON process: tasks and activities

TASK DESCRIPTION BON DELIVERABLES

Delineate system borderline. Find
G 1 SYSTEM CHART, SCENARIO CHARTS
A
subsystems, user metaphors, use cases.
T
H List candidate classes. Create glossary of
2 CLUSTER CHARTS
E technical terms.
R
I SYSTEM CHART, CLUSTER CHARTS,
N Select classes and group into clusters.
G 3 STATIC ARCHITECTURE,
Classify; sketch principal collaborations.
CLASS DICTIONARY

D Define classes. Determine commands,

4 CLASS CHARTS
E queries, and constraints.
S
C Sketch system behaviors. Identify events, EVENT CHARTS, SCENARIO CHARTS,
R 5 object creation, and relevant scenarios CREATION CHARTS,
I
B drawn from system usage. OBJECT SCENARIOS
I
N Define public features. Specify typed CLASS INTERFACES,
G 6
signatures and formal contracts. STATIC ARCHITECTURE

CLASS INTERFACES,
Refine system. Find new design classes, STATIC ARCHITECTURE,
7
D add new features. CLASS DICTIONARY, EVENT CHARTS,
E OBJECT SCENARIOS
S
I CLASS INTERFACES,
G
N 8 Generalize. Factor out common behavior. STATIC ARCHITECTURE,
I CLASS DICTIONARY
N
G Complete and review system. Produce
Final static and dynamic models;
9 final static architecture with dynamic
all BON deliverables completed.
system behavior.

STANDARD ACTIVITIES
1 Finding classes
2 Classifying
3 Clustering
4 Defining class features
5 Selecting and describing object scenarios
6 Working out contracting conditions
7 Assessing reuse
8 Indexing and documenting
9 Evolving the system architecture
384 BON QUICK REFERENCE

BON notation: charts and interfaces

INFORMAL CHARTS

SYSTEM SYSTEM_NAME Part: EVENTS SYSTEM_NAME Part:

PURPOSE INDEXING COMMENT INDEXING

Cluster Description External (in/out) Involved object types

CLUSTER CLUSTER_NAME Part: SCENARIOS SYSTEM_NAME Part:

PURPOSE INDEXING COMMENT INDEXING

Class / (Cluster) Description Scenario 1:

Description 1

CREATION SYSTEM_NAME Part:

CLASS CLASS_NAME Part:
COMMENT INDEXING
TYPE OF OBJECT INDEXING
Class Creates instances of
Inherits from

Queries

Commands

Constraints

CLASS INTERFACE FEATURE SIGNATURES

name * , name+ , name++ deferred / effective / redefined
CLASS_NAME name: TYPE result type
name:{ TYPE aggregation result type
Indexing information name:(n) TYPE shared result type
{ ^CLASS_NAME.name } rename clause
Inherits: PARENT – name: TYPE input argument
CLASSES ?, ! pre- and postcondition

Public features
ASSERTION SYMBOLS
A, B, C
∆ name attribute name may change
Features only visible
old expr previous value of expr
to classes A, B, C
@, ∅ current object, void reference
Invariant +, −, ∗, /, ^, //, \\ arithmetic op
Class invariant =. ≠, <, ≤, >, ≥ relational op
→ , ↔ , ¬, and, or, xor boolean op
∃, ∀, | , • predicate logic op
∈, ∉, { }, .. set op
: TYPE type op
 BON QUICK REFERENCE 385

BON notation: static and dynamic diagrams

CLASS HEADERS SCENARIO BOX

NAME reused Scenario:

1 description
● 2 …
NAME persistent

NAME parameterized OBJECT HEADERS

[G, H]
NAME one object
* deferred
NAME
NAME one object (qualified)
( id )
+
NAME effective
NAME one or more objects
▲
NAME interfaced
OBJECT GROUPING
NAME root

object group
(unnamed)
CLUSTERING

NAME NAME

cluster object group

NAME NAME

cluster (iconized) object group (iconized)

STATIC LINKS DYNAMIC LINKS

inheritance message passing

client association DYNAMIC LABELS

1, 2 sequence number
{ client aggregation
STATIC LABELS
MULTIPLICITY name
name1, name2
3 24 of relation TYPE […]
name: TYPE […]
1 of shared instances (name1, name2): TYPE […]
→ TYPE […]
Appendix D:
Other approaches

Below is a list of references to other approaches to object-oriented analysis and

design. Only work that presents a distinct method has been included.

ASTS Development Method 3 (ADM3)

[Firesmith 1993]
the Booch Method
[Booch 1994]
Class-Relation
[Desfray 1992]
Class, Responsibility, Collaboration (CRC)
[Beck 1989]
the Fusion Method
[Coleman 1994]
General Object-Oriented Design (GOOD)
[Seidewitz 1987]
Hierarchical Object-Oriented Design (HOOD)
[Robinson 1992]
Methodology for Object-Oriented Development (MooD)
[Siemens 1993]
Methodology for Object-Oriented Software Engineering of Systems (MOSES)
[Henderson-Sellers 1994]
Object Behavior Analysis (OBA)
[Rubin 1992]

386
 OTHER APPROACHES 387

Object Modeling Technique (OMT)

[Rumbaugh 1991]
Object-Oriented Analysis/Object-Oriented Design (OOA/OOD)
[Coad 1991a, Coad 1991b]
Object-Oriented Analysis/Object-Oriented Design Language (OOA/OODLE)
[Shlaer 1992]
Object-Oriented Role Analysis, Synthesis, and Structuring (OORASS)
[Reenskaug 1992]
Object-Oriented Semantic Networks
[Berard 1991]
the Object-Oriented Software Development Method
[Colbert 1989]
Object-Oriented Specification (OOS)
[Bailin 1989]
Object-Oriented Structured Design (OOSD)
[Wasserman 1990]
Object-Oriented Systems Analysis (OSA)
[Embley 1992]
Objectory / OOSE
[Jacobson 1992]
Ptech
[Martin 1992]
Responsibility-Driven Design (RDD)
[Wirfs-Brock 1990]
Semantic Object Model (SOM)
[Velho 1992, Velho 1994]
Semantic Object Modelling Approach (SOMA)
[Graham 1994]
Synthropy
[Cook 1994]
Uniform Object Notation/Synthesis
[Page-Jones 1990]
Appendix E:
Glossary of terms

ABSTRACT CLASS. See DEFERRED CLASS.

ABSTRACT DATA TYPE. A type of data structure (often abbreviated ADT)
defined exclusively through its external behavior. An ADT is specified by a
number of applicable operations, how each operation may be invoked (the
signature), and its effect (the semantics).
ABSTRACT FEATURE. See DEFERRED FEATURE.
ABSTRACTION. A conceptual model of something, which suppresses details
that are less interesting in a certain context and instead emphasizes what is
essential. Abstraction is the key tool for humans to understand the world—
without it reasoning would not be possible. However, since essentiality is
not an objective quality, the value of an abstraction depends entirely on
what it is used for.
ACCESS CONTROL. The selective mechanism by which clients are granted
access to the features of a class. Public features are accessible to all clients,
while restricted features may only be used by a specified group of clients.
A private feature can only be called on the current object, and only by other
features in the same class (or recursively by the feature itself). See also
INFORMATION HIDING.
ACTIVE OBJECT. An object playing an active role, that is driving a thread of
control by calling other objects. The called objects then play the part of
passive objects. Since being active is a role, the same object can be active
in one scenario and passive in another. The role of acting in both capacities
(being called and also calling other objects) is sometimes referred to as
agent, and the corresponding objects are then called agents.
ACTOR. Anything that can call a system object from the outside, such as a
human operator, a sensory device, or another system.

388
 GLOSSARY OF TERMS 389

AGENT. See ACTIVE OBJECT.

AGGREGATION. A composition of a group of objects into an integral unit.
AGGREGATION RELATION. A relation between an object and one of its integral
parts. The exact semantics of this relation may vary from one case to
another within the same system.
ANALYSIS. In normal usage, the process of breaking something down into its
constituent parts to find out how it works. However, object-oriented
analysis is often used in the alternative meaning of building a model of a
problem domain.
ANCESTOR CLASS. A class from which another class, the DESCENDANT, inherits
directly or indirectly.
APPLICATION. An executable system modeling some problem domain.
ARCHITECTURE. The major components of a system and the structural
relations between them.
ASSERTION. A software contracting element expressed as a logical statement
(predicate). The assertions discussed in this book are routine pre- and
postconditions and class invariants, which are used to define the semantics
of a class without regard to implementation details. Complementary
assertions, such as loop invariants, loop variants, and free assertions, may
be used for reasoning about the correctness of a given implementation.
ASSOCIATION. A relationship denoting a logical dependency between two
classes.
ATTRIBUTE. A property of an object manifested as a function returning a
value.
BEHAVIOR. The way an object reacts to calls from other objects. The top-level
object of a system, the root object (or root objects, in the case of concurrent
object execution), is considered called by whatever starts the corresponding
process.
BINARY ASSOCIATION. A relationship between two entities in an
entity−relationship model.
BINDING. The attachment of an object to an entity. An entity, in BON
terminology, is any symbol used to refer to an object. See also STATIC
BINDING and DYNAMIC BINDING.

BOOLEAN ALGEBRA. See PROPOSITIONAL LOGIC.

390 GLOSSARY OF TERMS

BOTTOM-UP DESIGN. The construction of something new from already existing

parts. Opposite of TOP-DOWN DESIGN. Extensive reuse implies (by
definition) a great deal of bottom-up design.
BROKER. A special object, acting as mediator between client and server
objects in distributed object-oriented applications. The term was introduced
in the “Object Management Architecture Guide” published by the Object
Management Group and X/Open [OMG 1990], and is an abbreviation for
object request broker (ORB).
BUSINESS RULES. Norms of a business that should not be violated by any
system action. Such rules often manifest themselves as constraints on the
persistent objects of the system.
CANDIDATE CLASS. A concept that is a candidate to become a class in an
object-oriented model. The first set of candidate classes usually results
directly from a study of the problem domain.
CASE DISCRIMINATION. The explicit enumeration of a set of possible system
states and specification of a corresponding action for each case. Much of
the potential of the object-oriented approach lies in reducing the number of
case discriminations in a system by an order of magnitude through the
systematic use of polymorphism and dynamic binding. (See also SINGLE
CHOICE PRINCIPLE.)

CHILD CLASS. A class that inherits directly from another class (the PARENT).
CLASS. A description of the behavior of certain kinds of objects, called
instances of the class. In this book, a class is viewed as an implementation
of an abstract data type, and the semantics of the corresponding operations
is specified through software contracts.
CLASS ATTRIBUTE. A state variable (sometimes called instance variable) of a
class. In this book, we do not consider the concept of metaclass as used in
Smalltalk.
CLASS FEATURE. See FEATURE.
CLASS INSTANCE. An object built according to the description in a class; any
number of instances may be built from the same class.
CLASS INTERFACE. The collective interface of all the features of a class. The
public interface is defined by the set of public features.
CLASS INVARIANT. An assertion about every object of a certain class. The
class invariant must be satisfied before and after execution of any public
 GLOSSARY OF TERMS 391

feature of the class. This is more committing than the contracts of

individual features, since not only all existing features but also all features
added in the future must obey the class invariant.
CLASS METHOD. In Smalltalk, a class method refers to an operation applicable
to the class itself rather than to an instance of the class. In this book,
however, a class is not viewed as an object but merely as a description of
object behavior. Therefore, a class method (or class feature) just means a
feature of a class.
CLASSIFICATION. Systematic placement of objects in categories with respect to
similarities and differences, often resulting in hierarchical orderings such as
the Linnaean system for classification of plants and animals. The science or
practice of classification is often called taxonomy after the branch of
biology concerned with classification of organisms.
CLIENT. Either a client object calling another object (the supplier object), or a
client class encapsulating a call to an object of another class (the supplier
class). See also SUPPLIER.
CLIENT−SERVER. A general model in which a consumer (the client) can
request service from a producer (the server) without knowing where the
server is located (apart from its name) or how the services are implemented.
CLUSTER. A group of classes and/or other clusters selected according to some
criteria to form a conceptual unit. A system may be partitioned into a
hierarchical structure of non-overlapping clusters, called a system view.
The same set of classes can be partitioned into different system views, but
for practical purposes only one is usually maintained for a particular
system. Clusters may be related to other clusters and also to individual
classes (in the same, or in other clusters).
CLUSTERING. The act of designing a cluster structure. The selection criteria
are usually highly variable from one system to another, and also within the
same system. Some clusters may represent fully fledged subsystems, while
others merely collect a small group of classes with a common ancestor.
COHESION. A qualitative measure of how strongly the components of a
module are related to each other. A highly cohesive module concentrates
on a single task, and is therefore much easier to understand and maintain,
compared to one with low cohesion, which coincidentally groups many
unrelated tasks into one. See also COUPLING.
COLLABORATION. A term used to signify the cooperation of a set of objects
calling each other to perform some overall task. See also RESPONSIBILITY.
392 GLOSSARY OF TERMS

COMMAND. An operation on an object which does not return any value, but
which may change the system state. See also QUERY.
COMPONENT LIBRARY. A repository containing reusable components.
COMPRESSION. In BON notation, most graphical elements can be compressed
into less space consuming forms, or hidden completely. A case tool would
keep the full information in its internal model, so compressed elements may
again be expanded. Compression and expansion can be applied recursively,
and the level of detail chosen for each part of a system diagram.
CONCRETE CLASS. See EFFECTIVE CLASS.
CONSTRAINT. Signifies some rule or restriction that applies to an object type or
some of its operations. Complements the queries and commands of BON
class charts to capture the semantics of a class, or general business rules.
Some identified constraints usually translate into formal assertions during
later phases, while others may instead influence the system design.
CONSTRUCTOR. An operation which is given control directly after the creation
of an object (before returning to whatever object issued the creation call) in
order to complete the initialization. Constructors are needed when the
default initialization values for the attributes of an object are not enough. In
particular, a constructor must ensure that the class invariant is satisfied
directly after an object is created. Some implementation languages permit
the specification of alternative constructors for a given class.
CONTAINER CLASS. A class representing a data structure that can hold objects
as elements.
CONTRACT MODEL. See DESIGN BY CONTRACT.
CONTRAVARIANT RULE. A rule stating that argument types of a feature
signature may be redefined to (and only to) ancestor types. See also
COVARIANT RULE.
CONTROL FLOW. The logical paths taken during execution of a system
depending on various conditions of the system state.
COUPLING. A qualitative measure of module interconnection in a software
structure. Low coupling means that each module is relatively independent
of other modules. This makes the system easier to understand and maintain,
since it is possible to investigate one part in isolation without having to deal
with the rest of the system concurrently. See also COHESION.
COVARIANT RULE. A rule stating that argument types of a feature signature
 GLOSSARY OF TERMS 393

may be redefined to (and only to) descendant types. See also

CONTRAVARIANT RULE.
CURRENT OBJECT. Except for the root object (which is called from the
surrounding operating system environment), the only way to invoke a
feature of a class is to apply it to an object. While the feature is executing,
this object then becomes the current object and is implicitly referred to
whenever an unqualified call or an assignment to a state variable occurs.
DATA ABSTRACTION. An encapsulation of data that hides internal structural
details and instead presents an external interface to clients.
DATA DICTIONARY. A description of all data types used in a software system.
The term is typically used in connection with data modeling and relational
database management systems.
DATA FLOW DIAGRAM. A graph showing the major data flows of a system at
execution time and the successive transformation of data. Data flow
diagrams (DFDs) are part of structured analysis methods.
DATA HIDING. See INFORMATION HIDING.
DATA MODELING. A modeling technique concentrating on the central data
structures of a system, particularly the persistent parts. The most common
approaches used in this context are variants of the entity−relationship (ER)
modeling [Chen 1976].
DATABASE. A repository for storage of persistent data. A database could be
anything from a flat file to a multi-disk, multi-platform distributed data set.
DBMS. Acronym for database management system. The mainstream DBMS
products have developed from hierarchical systems, over network systems,
into relational systems (RDBMS) and object-oriented database management
systems (ODBMS).
DEFENSIVE PROGRAMMING. A programming style, also referred to as blind
checking, where redundant error control is inserted in the program code.
This leads to increased complexity and programs which are more difficult to
understand and maintain. The reason for defensive programming is
generally that no clear division of responsibility between client and supplier
exists in the system. By contrast, the software contracting model specifies
exactly who is responsible for what, and therefore represents the opposite of
defensive programming.
DEFERRED CLASS. A class containing at least one feature which has no
implementation (and never will have). Deferred classes cannot be
394 GLOSSARY OF TERMS

instantiated, but serve as partial specification for all corresponding

descendant classes, which will fill in the missing details. Deferred classes
are sometimes called abstract, and if no features are implemented, a
deferred class can come very close to the mathematical specification of an
abstract data type. See also EFFECTIVE CLASS.
DEFERRED FEATURE. A feature that has no implementation (and never will
have). It serves instead as a (full or partial) specification of all implemented
versions of the feature that may occur in descendant classes. The opposite
is EFFECTIVE FEATURE.
DELEGATION. Refers to shared behavior in object-oriented systems using
prototypes instead of classes [Lieberman 1986]. Delegation enables objects
to reuse part of the knowledge stored in a prototype object representing the
default behavior of a concept. This approach is adopted by the language
Self [Ungar 1987].
DERIVED ATTRIBUTE. An attribute in ER modeling that can be inferred from
other attributes.
DESCENDANT CLASS. A class which inherits directly or indirectly from another
class, the ANCESTOR.
DESIGN. In normal usage, the process of arranging or building something from
a set of components. However, object-oriented design is often used in the
alternative meaning of building a computer representation of an analysis
model.
DESIGN BY CONTRACT. A view of software development as a series of
documented contracting decisions [Meyer 1992b]. In this book we use the
phrase “software contracting”.
DESTRUCTOR. A class feature which is executed when an object of the
corresponding class is finally destroyed. May be invoked by an automatic
garbage collector or through some other mechanism.
DOMAIN ANALYSIS. The analysis of a certain problem domain, particularly
with the intent of finding common concepts that could be captured as
reusable abstractions.
DOMAIN EXPERT. A person with great skills and experience in a certain
problem domain area.
DYNAMIC BINDING. A mechanism permitting different behavior to result from
the same feature call. Depending on the exact type of object attached to the
entity on which a feature is called, different implementations may be
 GLOSSARY OF TERMS 395

invoked at different points in time. The mechanism is crucial to eliminate

the abundance of case discriminations that occur in most systems developed
with traditional techniques. Dynamic binding is usually combined with
inheritance. See also POLYMORPHISM.
DYNAMIC CHART. An informal BON chart recording dynamic behavior. There
are currently three types of dynamic chart: event chart, scenario chart, and
creation chart. See also STATIC CHART.
DYNAMIC MODEL. A model capturing a high-level description of the typical
execution of a system in terms of time-ordered events. It concentrates on
the how part to serve as complement to the what part represented by the
STATIC MODEL.
DYNAMIC TYPE. The actual type of object attached to an object reference at
any given moment. See also STATIC TYPE.
DYNAMIC TYPING. A typing scheme where each object has a well-defined
type, but where the type of each symbol used in a description is not
explicitly specified. Type correctness can only be checked at run-time.
Smalltalk is an example of an object-oriented language with dynamic
typing. See also STATIC TYPING.
EFFECTIVE CLASS. A class all of whose features are effective. Objects taking
part in a system execution must be instances of effective classes. See also
DEFERRED CLASS.
EFFECTIVE FEATURE. A feature which has an implementation attached to its
specification. Opposite of DEFERRED FEATURE.
ENCAPSULATION. See INFORMATION HIDING.
ENTITY. The term entity in BON stands for any symbol used to refer to an
object, such as a state variable, a feature argument, or a local variable.
ENTITY−RELATIONSHIP DIAGRAM. A diagram showing the data entities in a
system and the relationships between them, usually adorned with semantic
labels. Often abbreviated ER diagram or ERD.
ER MODELING. Abbreviation for entity−relationship modeling, which uses ER
diagrams to describe the most important data elements in a system
(particularly the persistent parts).
EVENT. See SYSTEM EVENT.
EVENT TRACE. A graph or table showing a series of events that may occur in a
system. Event trace diagrams usually represent time on one axis and
communicating objects on the other.
396 GLOSSARY OF TERMS

EXCEPTION HANDLING. The handling of contract violations at execution time.

A contract violation occurs when something happens which is outside the
scope of the system specification.
EXPANSION. See COMPRESSION.
EXTENSIBILITY. A system quality permitting many requirements changes to be
incorporated in the system without extensive redesign and corresponding
high cost.
FEATURE. An operation applicable to an object. The term feature also covers
the concept of class attribute at implementation time, that is state variables.
FEATURE CALL. The invocation of an operation on an object.
FINITE STATE MACHINE (FSM). A machine which accepts a sequence of input
events and produces a sequence of output events. Each input and output
event belongs to a finite set of events, and each output event depends both
on the corresponding input event and on the past history of input events.
Typical examples of FSMs are telephone switching circuits and vending
machines.
FIRST-ORDER LOGIC. A logic where the formal elements may range over
objects in sets, but not over all subsets of sets.
FRAMEWORK. A set of classes representing an object-oriented abstract design.
Typically, a user may tailor the behavior of a framework by client calls to
public framework routines, or by definition or redefinition of features
inherited from framework classes.
FSM MODELING. Using finite state machines (FSMs) to describe the dynamic
behavior of a system or some of its parts.
GARBAGE COLLECTION. The reclamation of allocated object space that is no
longer needed in a system. Memory management is a central issue in
object-oriented systems, since very often large numbers of objects are
created which only live for a relatively short time. The primary memory
occupied by these transient objects needs to be reused lest the system run
out of available object space. Modern automatic garbage collectors can
relieve the user of the tedious and extremely error-prone task of manual
memory administration with little execution overhead.
GENERALIZATION. The act of detecting a general pattern in a concrete problem
solution and capturing the higher-level essence, so it can be reused to solve
related problems. This is often done by introducing new deferred classes in
the inheritance hierarchies of a system. See also SPECIALIZATION.
 GLOSSARY OF TERMS 397

GENERIC CLASS. See PARAMETERIZED CLASS.

GROUP. See OBJECT GROUP.
HIERARCHY. A structure with a top-level component, where each component
may have any number of subcomponents, and where each component
(except the top component) is a subcomponent of exactly one other
component.
HYBRID TECHNIQUE. A technique mixing different paradigms in the same
approach.
IDENTITY. See OBJECT IDENTITY.
IMPLEMENTATION. The refinement of a software design to the point where the
resulting description is directly executable on some supporting platform.
INCREMENTAL DEVELOPMENT. A software development technique which starts
by producing a system with minimal functionality, barely enough to be
useful, and then proceeds to gradually add functionality in small
increments. Since each successive system version is tested and validated by
actual usage, the risk of going in the wrong direction is greatly reduced and
the cost estimation for each increment becomes much more feasible
compared to large monolithic developments. Incremental development
matches the basic object-oriented ideas very well.
INDEXING. The addition of keywords and index terms to classes in order to
improve documentation and facilitate the search for reusable components in
class libraries.
INFORMATION HIDING. A technique whereby implementation details likely to
change in the future are hidden from the authors of client modules, who are
instead presented with a public interface. In spite of the name, the intent is
usually not so much to hide the information as to prevent clients from
becoming dependent on it.
INHERITANCE. A mechanism allowing a module to share the behavior defined
in another module, but still be able to modify it and to add new behavior.
Some researchers advocate separation between inheritance of specification
and inheritance of implementation, but this book takes the opposite view.
INSTANCE. See CLASS INSTANCE.
INSTANCE MULTIPLICITY. The number of instances of a given class that may
be attached to another object through a named reference. Often expressed
as a range and referred to as a “cardinality constraint” in data modeling.
See also ROLE MULTIPLICITY.
398 GLOSSARY OF TERMS

INSTANCE SHARING. The same object being attached to several other objects,
which then share its resources.
INSTANCE VARIABLE. The Smalltalk name for state variables.
INSTANTIATION. The process of creating new objects according to the
description in a class.
INTERFACE CLASS. A class serving as (part of) the interface to a cluster of
classes. Classes outside the cluster only use the interface classes, while
these in turn use the rest of the classes in the cluster to implement the
external behavior.
INVARIANT. See CLASS INVARIANT.
LATE BINDING. See DYNAMIC BINDING.
LAYER. A set of classes and corresponding operations at the same level of
abstraction. See also CLUSTER.
LEVEL OF ABSTRACTION. A set of system components representing distinct
concepts in a given context. The levels of abstraction in a system are
partially ordered so that a given component depends only on components at
the same or lower levels.
LINK. The graphical representation of a relation between elements in BON.
Client links and inheritance links relate classes and clusters in static
diagrams, while message links relate objects and object groups in dynamic
diagrams.
MEMBER FUNCTION. A class feature in C++ terminology. Corresponds to a
query or a command in BON terminology, depending on whether or not it
returns a value.
MESSAGE. A feature call in Smalltalk terminology.
METACLASS. A Smalltalk concept whose instances are classes. There is no
correspondence to metaclasses in BON, since classes are strictly regarded as
descriptions of object behavior, and not as objects.
METALANGUAGE. A language used to reason about another language. Mixing
language and metalanguage without a clear syntactic separation between the
two levels can lead to severe misunderstandings.
METHOD. A technique or arrangement of work for a particular field or subject.
METHOD (OF A CLASS). A class feature in Smalltalk terminology.
 GLOSSARY OF TERMS 399

METHODOLOGY. Often used as a synonym for method, but signifies more

properly the system of methods used in a discipline, or the general science
of method and procedure.
MIXIN. In CLOS terminology, a class which is inherited to get access to
certain utility features.
MULTIPLE INHERITANCE. A mechanism allowing one class to inherit
simultaneously from more than one parent class.
MULTIPLICITY. See INSTANCE MULTIPLICITY and ROLE MULTIPLICITY.
NAMING. The procedure of giving names to concepts in a system. Proper
naming is extremely important for the understandability of a system.
OBJECT. The basic components of an object-oriented system, exhibiting well-
defined behavior in response to a number of applicable operations. Also
called CLASS INSTANCE.
OBJECT CLASS. Sometimes used in data modeling to signify a type of object,
as opposed to OBJECT INSTANCE. In object-oriented contexts, the term is
confusing and should not be used.
OBJECT DIAGRAM. A graphical description of a scenario showing message
passing between objects.
OBJECT GROUP. A group of objects that are treated collectively in some
context. In a BON dynamic diagram, messages may be passed to an object
group signifying that one or more of the objects are called.
OBJECT GROUPING. The process of finding adequate object groups.
OBJECT IDENTITY. The concept of a unique identity which distinguishes an
object from all other objects, regardless of whether their internal structures
and values coincide or not. All objects are created with a unique identity in
an object-oriented system, so the user need not worry about introducing
access keys.
OBJECT INSTANCE. Sometimes used in data modeling to signify an instance of
a type, as opposed to OBJECT CLASS. In object-oriented contexts, the term is
confusing and should not be used.
OBJECT MANAGEMENT. The storage and retrieval of persistent objects.
OBJECT-ORIENTED. A term signifying that a certain system or method includes
support for abstract data types combined with inheritance, polymorphism,
and dynamic binding.
400 GLOSSARY OF TERMS

OBJECT-ORIENTED ABSTRACTION. An abstraction represented as a class, which

describes the behavior of a certain type of object through applicable
operations.
OBJECT-ORIENTED ANALYSIS. An object-oriented model of a problem
domain—or the process of creating such a model.
OBJECT-ORIENTED DATABASE. A database and corresponding environment
with direct support for storing not only the data parts of an object, but also
its operations and references to other objects.
OBJECT-ORIENTED DESIGN. An object-oriented model of a computer
representation of a problem—or the process of creating such a model.
OBJECT-ORIENTED ENCAPSULATION. See OBJECT-ORIENTED ABSTRACTION.
OBJECT-ORIENTED EXECUTION. System execution viewed exclusively as a set
of objects passing messages to each other.
OBJECT-ORIENTED IMPLEMENTATION. Implementation of software using an
object-oriented language.
OBJECT-ORIENTED MODEL. A model expressed as a structure of classes with
specified operations, related through inheritance and client dependencies.
OBJECT-ORIENTED MODELING. The process of creating an object-oriented
model.
OBJECT-ORIENTED PROGRAMMING LANGUAGE. An executable language which
directly supports abstract data types, inheritance, and dynamic binding. The
three object-oriented languages attracting most attention in industry today
are C++, Smalltalk, and Eiffel.
OBJECT-ORIENTED SOFTWARE DEVELOPMENT. The process of developing
software using object-oriented analysis, design, and implementation.
OBJECT-ORIENTED SPECIFICATION. The process of specifying a system as a set
of cooperating objects with well-defined behavior. Software contracting
combined with strong typing can be used to capture the semantics of such a
system.
OBJECT-ORIENTED SYSTEM. A system built using object-oriented techniques.
Notice that this is not the same as a system exhibiting object-oriented user
metaphors, such as graphical items that may be manipulated by the user,
resizing of windows, drag and drop, and so forth. The latter kinds of system
need not be object oriented at all (although this is certainly an advantage for
their future maintenance).
 GLOSSARY OF TERMS 401

OBJECT STATE. That part of the system state that has an effect on the future
behavior of an object. Object state often corresponds to one or more
ATTRIBUTES of the object, but this does not necessarily mean that data is
stored in the object itself.
ODBMS. Acronym for object database management systems. (This term is
gradually replacing the longer name OODBMS as the “-oriented” part is
being dropped.)
OOA. Acronym for object-oriented analysis.
OOD. Acronym for object-oriented design.
OOPL. Acronym for object-oriented programming language.
OPERATION. See FEATURE.
OVERLOADING. A mechanism permitting the same name to be used in
different contexts with different meanings.
PARADIGM. A fundamental way of viewing a problem area which is generally
accepted by a large number of people. A paradigm is generally difficult to
question, because its perceived truth is often an integral part of a culture and
has a profound impact on how we think. Its followers may not be aware of
the paradigm at all; they just unconsciously take its implications for
granted. Switching to object-oriented modeling means a paradigm shift for
some people, notably those heavily trained in the so-called structured
techniques.
PARAMETERIZED CLASS. A class from which a number of related types may be
derived, depending on a number of furnished type parameters. STACK
[BOOK] and STACK [INTEGER] are typical examples of two types
derived from the same parameterized class STACK [T]. Also called generic
class.
PARENT CLASS. A class from which another class, the CHILD, inherits.
PARTIAL FUNCTION. A function which is not defined for all possible argument
values (as opposed to a TOTAL FUNCTION).
PARTIAL ROUTINE. A query or command which is not defined for all possible
argument values and system states (thus having a precondition which is
different from true). Partial routines play a central role in software
development, since in many cases they permit a much simpler and concise
solution compared to a corresponding total routine, which has to handle lots
of irrelevant cases for which there is no reasonable response. Combining
402 GLOSSARY OF TERMS

partial routines with software contracts allows precise specification of who

is responsible for checking: the client or the supplier.
PASSIVE OBJECT. See ACTIVE OBJECT.
PERSISTENT CLASS. A class whose instances are persistent.
PERSISTENT OBJECT. An object whose life span is independent of any system
session. It may reside in either primary or secondary storage.
PHYSICAL OBJECT. An object perceived as physical by people in a problem
domain.
POLYMORPHISM. A mechanism by which a named reference, also called an
entity, can be attached to objects of different type at different points in time.
Unconstrained polymorphism (as in Smalltalk) permits any entity to refer to
any type of object, while in a typed language (such as Eiffel) an entity may
only be attached to objects of the declared type or a descendant type.
Polymorphism is sometimes used as a synonym for DYNAMIC BINDING.
POSTCONDITION. A predicate which must be true just after the execution of a
public feature, provided the feature’s precondition was fulfilled when the
execution started. See also PRECONDITION.
PRECONDITION. A predicate that must be true when a public feature is called
by a client. See also POSTCONDITION.
PREDICATE LOGIC. A logic that apart from constant objects also uses formal
variables ranging over objects in its expressions. Predicate logic is a
generalization of propositional (or sentential) logic, and is also known as
predicate calculus.
PRIVATE FEATURE. A feature which is not publicly available to clients. It may
only be invoked on the current object by features in the same class.
PROBLEM DOMAIN. A certain application area or other context, in which a
problem is to be solved with the help of computer software.
PROGRAMMING-IN-THE-LARGE. The act of combining high-level components
into tailored configurations to fit the needs of a particular application. The
term was coined in [DeRemer 1976], which discusses module
interconnection languages.
PROPERTY. An (abstract) attribute of an object, either accessible directly
through some query or else indirectly inferable through system behavior.
PROPOSITIONAL LOGIC. A logic whose expressions are built from primitive
 GLOSSARY OF TERMS 403

sentences (considered true or false) and logical connectives, such as not,

and, or, implies, equals. Also known as propositional calculus, sentential
logic, or boolean algebra.
PROTOCOL. A Smalltalk term denoting the collective signatures of all features
(methods) of a class.
PROXY. An object acting on behalf of another object, but appearing to the
client as being that other object. Typically used in distributed object-
oriented systems to implement transparent calls across networks in
client−server architectures.
PUBLIC FEATURE. A feature which is publicly available to all clients.
PURE TECHNIQUE. A technique which does not mix elements of other
paradigms with its own view.
QUERY. An operation on an object which returns a value, but does not alter the
system state. See also COMMAND. Strict separation of object behavior into
state changing procedures (commands) and side-effect-free value returning
functions (queries) is of fundamental importance for the ability to reason
about the correctness of a class.
RECURSIVE SPECIFICATION. A specification technique in which the conceptual
base for expressing specification and implementation is the same.
Removing the artificial barriers between specification and implementation
(which is possible with object-oriented abstraction) gives developers a
chance to keep code and specification consistent during the lifetime of a
system, since the two will move in concert.
REDEFINITION. A modification in a descendant class of the implementation of
an inherited feature.
RELATION. A dependency between two entities; used in this book as an
abbreviation for the term relationship. BON includes three types of
relation: inheritance, client, and message.
RELATIONSHIP. See RELATION.
RENAMING. A modification of the name of an inherited feature in a descendant
class. This is a purely syntactic device, which allows more proper names to
be used when the role of an inherited concept is altered.
REPEATED INHERITANCE. The same feature being inherited more than once
from a class.
RESPONSIBILITY. A term used to emphasize the behavior of an object rather
404 GLOSSARY OF TERMS

than its structure. An object may be responsible for certain system actions
and for maintaining certain information, but instead of dealing directly with
the processing or data storage, the object may call on other objects behind
the scenes to do the job. See also COLLABORATION.
RESTRICTED FEATURE. A feature which is only publicly available to specific
clients.
REUSABLE LIBRARY. See COMPONENT LIBRARY.
REUSE MANAGER. Someone responsible for the component libraries and for
maintaining the reuse policy of an organization.
REUSED CLASS. A class that is reused and possibly adapted from previous
developments. Marked with underscore in BON static diagrams.
REVERSIBILITY. The possibility of seamlessly translating changes made during
a certain development phase back into earlier phases, so as to maintain
consistency.
ROBUSTNESS. The ability of a system to function reasonably also in situations
that should never occur according to the system specifications. Robustness
cannot be defined precisely, but includes things like restoring class
invariants, guarding persistent data from corruption, stating the problem as
clearly as possible to a system operator, saving traces for manual recovery,
etc.
ROLE. The purpose for which a certain class is used. Roles in combination
with their corresponding types are crucial for understanding a system, since
they reflect not only what abstractions are employed, but also how they are
used in various contexts.
ROLE MULTIPLICITY. The number of client relations from a client class to a
supplier class. Each relation corresponds to the supplier being used by the
client to fulfill a certain role. The number of instances involved in each
relation may vary. See also INSTANCE MULTIPLICITY.
ROLE NAME. The name of a role, usually captured as a symbol (entity) which
is attached to an object. Sometimes the role name is the same as the class
name, which means the class is just used in its general sense. In other cases
the role name may reflect a typical specialization or a type of usage that was
never anticipated.
ROOT CLASS. A class of which one instance will be created when an object-
oriented process is started, and whose initialization routine drives the
execution.
 GLOSSARY OF TERMS 405

ROOT OBJECT. The root class instance created when an object-oriented

execution starts.
SCALABILITY. The ability of a method and notation to scale up from small
textbook examples to sizable real applications without losing usability. For
a notation, a “zooming” capability is essential so views can be presented at
different levels without losing the overall system structure. BON uses
compression and expansion to achieve this.
SCENARIO. A script of a possible system execution showing the objects
involved, which other objects they call, and the temporal order of these
calls.
SEAMLESSNESS. The quality of natural translation from problem domain
specification, over system design, into executable code. To be really
effective with respect to reuse and ease of maintenance, seamlessness needs
to be combined with REVERSIBILITY.
SELF. The Smalltalk term for the CURRENT OBJECT. Corresponds to this in
C++ and Current in Eiffel.
SEMANTIC DATA MODEL. A data model capturing the usage and meaning of
information in a system that can be used as a basis for structuring the
system data.
SEMANTIC LABEL. In ER modeling, a label on a relationship between two
entities stating its meaning. Sometimes used in BON as an attachment to an
inheritance link to convey the designer’s intent, or as a complement to the
role name of a client link.
SERVER. An object which provides a number of services to clients, often in
parallel.
SERVICE. One or a set of related operations that may be requested from a
server.
SIDE EFFECT. A system state change taking place as a result of calling a value
returning function. Side effects are dangerous, because when they occur,
there is no guarantee that a query will return the same result when called
repeatedly. This makes logical reasoning about program correctness much
more difficult. Maintaining a clear separation between state changing
commands and side-effect-free queries may sometimes require a few extra
instructions in an implementation, but the cost is negligible in view of the
added clarity of the system.
SIGNATURE. The type of the return value of a feature, if any, and the number
406 GLOSSARY OF TERMS

and types of all arguments, if any. The signature specifies the syntax of a
feature call, while contracting elements in the form of assertions capture its
semantics.
SINGLE CHOICE PRINCIPLE. A principle stating that whenever there is a choice
between a number of alternatives which the system needs to keep track of,
the exact list should be known in only one module. The advantage is that
updates to the list can be done in one place and will immediately take effect
in all other parts of the system. With object-oriented techniques it is
possible to come close to a strict application of this principle by relying
heavily on dynamic binding.
SINGLE INHERITANCE. The process of inheriting from at most one parent class.
SOFTWARE CONTRACTING. See DESIGN BY CONTRACT.
SOFTWARE ENGINEERING. A term which was invented by Douglas McIlroy in
1968 and presented at a NATO conference [McIlroy 1976]. It expressed a
wish for the future rather than a description of the way software was
produced at the time. McIlroy envisioned a software components industry,
and the greatest potential of the object-oriented approach today lies in the
eventual realization of this dream, which has proved much more difficult
than expected.
SOFTWARE SYSTEM. A set of executable components cooperating to fulfill a
certain purpose.
SPECIALIZATION. The act of adding more detail to a general pattern in order to
solve a problem which fits the principles of the pattern. This is often done
by introducing new descendant classes in the inheritance hierarchies of a
system. See also GENERALIZATION.
STATE. A summary of the current situation in some context (state of affairs).
See also SYSTEM STATE and OBJECT STATE.
STATE TRANSITION DIAGRAM. A graph depicting a number of states and
possible transitions between them.
STATE VARIABLE. A physical class attribute holding a value.
STATIC BINDING. The act of deciding which version of a routine will be
invoked as a result of a feature call based only on the static type of the
reference used for calling the feature. Static binding contradicts one of the
most important ideas in the object-oriented paradigm: to favor
polymorphism and dynamic binding. If it is to be applied for efficiency,
this should be done automatically behind the scenes by an optimizing
 GLOSSARY OF TERMS 407

compiler, since only then can it be guaranteed that the statically bound
version is also the correct one.
STATIC CHART. An informal BON chart recording static structure. There are
currently three types of static chart: system chart, cluster chart, and class
chart. See also DYNAMIC CHART.
STATIC MODEL. A model which shows the structure of a system: how modules
and objects are related and depend on each other. A static model
concentrates on the description of a system’s behavior embodied in the
specification given by the system classes, representing the what part. See
also DYNAMIC MODEL.
STATIC TYPE. The declared type of an object reference. During system
execution, only objects of this type or descendant types may be attached to
the reference. See also DYNAMIC TYPE.
STATIC TYPING. A typing scheme (also called strong typing) where each object
has a well-defined type, and where the type of each symbol used in a
description is explicitly specified. In a statically typed language, a compiler
can check before execution that there will always be an appropriate
implementation to take care of every feature call. Eiffel and (to a lesser
extent) C++ are examples of object-oriented languages with static typing.
STEPWISE REFINEMENT. A method of successively refining a high-level design
until an executable level is reached.
STIMULUS. See SYSTEM EVENT.
STRONG TYPING. See STATIC TYPING.
SUBCLASS. A class which inherits from another class, directly or indirectly.
Same as DESCENDANT CLASS.
SUBSYSTEM. A sizable part of a system, which is relatively independent of
other parts and reasonably complete with regards to its function.
SUPERCLASS. The Smalltalk term for ANCESTOR CLASS.
SUPPLIER. Either a supplier object being called by another object (the client
object), or a supplier class encapsulating a feature which is called by a
feature of another class. See also CLIENT.
SYSTEM. See SOFTWARE SYSTEM.
SYSTEM BEHAVIOR. The collective behavior of the objects in a system.
SYSTEM EVENT. Something to which a system will respond with a certain
408 GLOSSARY OF TERMS

behavior. Typical events are keyboard input, data from sensory devices,
and calls from other systems.
SYSTEM STATE. The sum of all information stored in a system. The system
state is a representation of its history of events, although only those events
that were actually recorded (explicitly or implicitly) may influence the
future behavior of the system. Depending on system state and incoming
events, the state may be changed by the commands in the system. Such a
change is also called a transition from one state to another.
SYSTEM VIEW. In BON terminology, a system view is a hierarchical
partitioning of the classes of a system into possibly nested clusters. The
clustering structure is orthogonal to the classes and does not change the
semantics or interfaces of these. The structure is merely a help for a human
reader to better understand the system, albeit a very important one. Several
system views are possible for the same set of classes, but in practice only
one is elaborated.
TANGIBLE OBJECT. See PHYSICAL OBJECT.
THIS. The C++ term for CURRENT OBJECT. Corresponds to self in Smalltalk
and Current in Eiffel.
TOP-DOWN DESIGN. A method of starting with the functions required at the
highest level of a system and recursively decomposing these functions into
smaller parts until an executable level is reached.
TOTAL FUNCTION. A function which is defined for all possible argument
values, as opposed to a PARTIAL FUNCTION.
TRANSIENT OBJECT. The opposite of PERSISTENT OBJECT. Transient objects are
destroyed when the execution of a system is over.
TRANSITION. A change from one state to another. See SYSTEM STATE.
TRADER. Roughly equivalent to object request broker. The term was
introduced by the Advanced Network Systems Architecture (ANSA) project
[Marshak 1991].
TUPLE. An ordered set of values, each from a specific domain. The term is
used in the relational data model, and usually called row in commercial
DBMS products and industry. A relation in this model (usually called table
in industry) is a set of tuples drawn from the same cartesian product of
domains.
TYPE. The pattern of behavior of a certain kind of object, specified as a
 GLOSSARY OF TERMS 409

number of applicable operations and their semantics. A type is roughly

equivalent to a class in this book, except that parameterized classes give rise
to different types, depending on the type of the furnished parameters.
TYPE DERIVATION. The act of deriving a specific type from a generic
(parameterized) class.
TYPE INSTANTIATION. See TYPE DERIVATION.
TYPING. The policy with regard to types employed by a language. See also
STATIC TYPING and DYNAMIC TYPING.
USE. In this book, using a class means being a client of the class.
USE CASE. A type of usage of a system. Roughly equivalent to a scenario in
BON.
USER METAPHOR. The mental model a system user has of the internal
workings of a system.
VIEW. See SYSTEM VIEW.
VIRTUAL FUNCTION. In C++, a feature whose calls will be effected using
dynamic binding.
VISIBILITY. Specifies whether an operation is available to a client or not.
Public features are always available, while restricted features may only be
used by a certain group of classes and private features may only be called
on the current object by features of the same class.
VOID REFERENCE. A reference which is not attached to any object.
410 GLOSSARY OF TERMS
References

Adams 1994.
Greg Adams and Jean-Pierre Corriveau. “Capturing Object Interactions.”
Proc. Technology of Object-Oriented Languages and Systems (TOOLS 13,
Versailles, France, Mar. 1994), Prentice Hall, Hemel Hempstead, England,
1994.
Alexander 1977.
Christopher Alexander, Sara Ishikawa, and Murray Silverstein. A Pattern
Language. Oxford University Press, New York, 1977.
Alexander 1987.
James H. Alexander. “Painless Panes for Smalltalk Windows.” Proc.
Object-Oriented Programming Systems, Languages, and Applications
(OOPSLA ’87) in ACM SIGPLAN Notices 22(12), Orlando, FL, Oct. 1987,
pp. 287−294.
Allworth 1990.
R. T. Allworth and R. N. Zobel. Introduction to Real Time Software
Design. 2nd edn, Macmillan, London, 1990.
Arango 1989.
Guillermo Arango. “Domain Analysis: From Art to Engineering
Discipline.” Proc. Fifth Int. Workshop on Software Specification and
Design, Pittsburgh, PA, May 1989, pp. 152−159.
Bailin 1989.
Sidney C. Bailin. “An Object-Oriented Requirements Specification
Method.” Commun. ACM 32(5), May 1989, pp. 608−623.
Beck 1989.
Kent Beck and Ward Cunningham. “A Laboratory for Teaching Object-
Oriented Thinking.” Proc. Object-Oriented Programming Systems,
Languages, and Applications (OOPSLA ’89) in ACM SIGPLAN Notices
24(10), New Orleans, LA, Oct. 1989, pp. 1−6.
Berard 1991.
Ed Berard. “Object-Oriented Semantic Networks.” Berard Software
Engineering Inc., Gaithersburg, MA, 1991.

411
412 REFERENCES

Blaha 1994.
Michael Blaha, William Premerlani, and Hwa Shen. “Converting OO
Models into RDBMS Schema.” IEEE Software 11(3), May 1994, pp.
28−39.
Boehm 1981.
Barry W. Boehm. Software Engineering Economics. Prentice Hall,
Englewood Cliffs, NJ, 1981.
Booch 1994.
Grady Booch. Object-Oriented Analysis and Design with Applications. 2nd
edn, Benjamin/Cummings, Redwood City, CA, 1994.
Booth 1989.
P. Booth. An Introduction to Human−Computer Interaction, chapter 4.
Lawrence Erlbaum Assoc., Hillsdale, NJ, 1989.
Bracha 1993.
Gilad Bracha and David Griswold. “Strongtalk: Typechecking Smalltalk in
a Production Environment.” Proc. Object-Oriented Programming Systems,
Languages, and Applications (OOPSLA ’93) in ACM SIGPLAN Notices
28(10), Washington, DC, Oct. 1993, pp. 215−230.
Brooks 1975.
Frederick P. Brooks, Jr. The Mythical Man-Month. Addison-Wesley,
Reading, MA, 1975.
Brooks 1987.
Frederick P. Brooks, Jr. “No Silver Bullet: Essence and Accidents of
Software Engineering.” IEEE Computer 20(4), Apr. 1987, pp. 10−19.
Carroll 1982.
J. M. Carroll and J. Thomas. “Metaphor and the Cognitive Representation
of Computing Systems.” IEEE Trans. Systems, Man, and Cybernetics
SMC−12(2), Mar./Apr. 1982.
Cattel 1991.
R. G. G. Cattel. Object Data Management: Object-Oriented and Extended
Relational Database Systems. Addison-Wesley, Reading, MA, 1991.
Charette 1989.
Robert N. Charette. Software Engineering Risk Analysis and Management.
McGraw-Hill, New York, 1989.
Chen 1976.
Peter Pin-Shan Chen. “The Entity−Relationship Model—Toward a Unified
View of Data.” ACM Trans. Database Systems 1(1), Mar. 1976, pp. 9−36.
Coad 1991a.
Peter Coad and Edward Yourdon. Object-Oriented Analysis. 2nd edn,
Prentice Hall, Englewood Cliffs, NJ, 1991.
 REFERENCES 413

Coad 1991b.
Peter Coad and Edward Yourdon. Object-Oriented Design. Prentice Hall,
Englewood Cliffs, NJ, 1991.
Codd 1970.
E. F. Codd. “A Relational Model of Data for Large Shared Data Banks.”
Commun. ACM 13(6), June 1970, pp. 377−387.
Codd 1985a.
E. F. Codd. “Is your DBMS really relational?” Computer World, Oct. 14,
1985, pp. ID/1−9.
Codd 1985b.
E. F. Codd. “Does your DBMS run by the rules?” Computer World, Oct.
21, 1985, pp. 49−60.
Codd 1990.
E. F. Codd. The Relational Model for Database Management: Version 2.
Addison-Wesley, Reading, MA, 1990.
Colbert 1989.
E. Colbert. “The Object-Oriented Software Development Method: A
practical approach to object-oriented development.” Proc. Tri-Ada, New
York, 1989.
Coleman 1994.
Derek Coleman, Patrick Arnold, Stephanie Bodoff, Chris Dollin, Helena
Gilchrist, Fiona Hayes, and Paul Jeremes. Object-Oriented Development:
The Fusion Method. Prentice Hall, Englewood Cliffs, NJ, 1994.
Constantine 1968.
Larry L. Constantine. “The Programming Profession, Programming
Theory, and Programming Education.” Computers and Automation 17(2),
1968, pp. 14−19.
Cook 1994.
Steve Cook and John Daniels. Designing Object Systems. Prentice Hall,
Hemel Hempstead, England, 1994.
Cook 1992.
William R. Cook. “Interfaces and Specifications for the Smalltalk-80
Collection Classes.” Proc. Object-Oriented Programming Systems,
Languages, and Applications (OOPSLA ’92) in ACM SIGPLAN Notices
27(10), Vancouver, Canada, Oct. 1992, pp. 1−15.
Cypher 1986.
A. Cypher. “The Structure of User Activities.” In User Centered System
Design, D. A. Norman and S. W. Draper (eds). Lawrence Erlbaum Assoc.,
Hillsdale, NJ, 1986.
414 REFERENCES

Date 1983.
C. J. Date. An Introduction to Database Systems, Vol. II. Addison-Wesley,
Reading, MA, 1983.
Date 1990.
C. J. Date. An Introduction to Database Systems, Vol. I. 5th edn, Addison-
Wesley, Reading, MA, 1990 (6th edn to appear 1994).
Date 1993.
C. J. Date with Hugh Darwen. A Guide to the SQL Standard. Addison-
Wesley, Reading, MA, 1993.
DeRemer 1976.
Frank DeRemer and Hans H. Kron. “Programming-in-the-Large versus
Programming-in-the-Small.” IEEE Trans. Software Engineering SE-2(2),
June 1976, pp. 80−86.
Desfray 1992.
P. Desfray. Ingénerie des Objets: Approche class-relation application à
C++. Editions Masson, Paris, 1992.
Dijkstra 1972.
Edsger W. Dijkstra. “Notes on Structured Programming (EWD 249,
Technical Univ. Eindhoven, 1969).” In Structured Programming, C.
Anthony R. Hoare (ed.). Academic Press, London, 1972, pp. 1−82.
Dijkstra 1976.
Edsger W. Dijkstra. A Discipline of Programming. Prentice Hall,
Englewood Cliffs, NJ, 1976.
Dijkstra 1982.
Edsger W. Dijkstra. “How to Tell Truths that Might Hurt?” In Selective
Writings on Computing: A Personal Perspective. Springer-Verlag, New
York, 1982, pp. 129−131.
Douglas 1983.
S. Douglas and Thomas P. Moran. “Learning text-editing semantics by
analogy.” Proc. ACM Human Factors in Computing Systems (CHI ’83),
Boston, MA, Dec. 1983.
Embley 1992.
David W. Embley, Barry D. Kurtz, and Scott N. Woodfield. Object-
Oriented Systems Analysis: A Model-Driven Approach. Prentice Hall,
Englewood Cliffs, NJ, 1992.
Farmer 1990.
William M. Farmer. “A Partial Functions Version of Church’s Simple
Theory of Types.” The Journal of Symbolic Logic 55(3), Sept. 1990, pp.
1269−1291.
 REFERENCES 415

Firesmith 1993.
Donald G. Firesmith. Object-Oriented Requirements Analysis and Logical
Design: A Software Engineering Approach. John Wiley, New York, 1993.
Floyd 1967.
Robert W. Floyd. “Assigning Meanings to Programs.” Proc. Symposium in
Applied Mathematics 19, J. T. Schwartz (ed), American Mathematical
Society, 1967, pp. 19−31.
Gabriel 1991.
Richard P. Gabriel, Jon L. White, and Daniel G. Bobrow. “CLOS:
Integrating Object-Oriented and Functional Programming.” Commun. ACM
34(9), Sept. 1991, pp. 28−38.
Gilb 1988.
Tom Gilb. Principles of Software Engineering Management. Addison-
Wesley, Reading, MA, 1988.
Goldberg 1991.
Adele Goldberg. “Object-Oriented Project Management.” Technology of
Object-Oriented Languages and Systems (TOOLS 4, Paris, Mar. 1991),
tutorial documentation, SOL, Paris, 1991.
Goldberg 1983.
Adele Goldberg and David Robson. Smalltalk-80: The Language and its
implementation. Addison-Wesley, Reading, MA, 1983.
Gougen 1978.
Joseph A. Gougen, Jim W. Thatcher, and E. G. Wagner. “An Initial
Algebra Approach to the Specification, Correctness, and Implementation of
Abstract Data Types.” In Current Trends in Programming Methodology,
Raymond T. Yeh (ed.). Prentice Hall, Englewood Cliffs, NJ, 1978, pp.
80−149.
Gould 1988.
J. D. Gould. “How to Design Usable Systems.” In Handbook of Human
Computer Interaction, chapter 35, M. Helander (ed.). Elsevier Science
Publisher, North-Holland, New York, 1988.
Graham 1994.
Ian Graham. Object-Oriented Methods. Addison-Wesley, Wokingham,
England, 1994.
Grape 1992.
Per Grape and Kim Waldén. “Automating the Development of Syntax Tree
Generators for an Evolving Language.” Proc. Technology of Object-
Oriented Languages and Systems (TOOLS 8, Santa Barbara, Aug. 1992),
Prentice Hall, Englewood Cliffs, NJ, 1992, pp. 185−195.
416 REFERENCES

Guttag 1977.
John V. Guttag. “Abstract Data Types and the Development of Data
Structures.” Commun. ACM 20(6), June 1977, pp. 396−404.
Halbert 1987.
Daniel C. Halbert and Patrick D. O’Brien. “Using Types and Inheritance in
Object-Oriented Programming.” IEEE Software 4(5), Sept. 1987, pp.
71−79.
Harel 1988.
David Harel. “On Visual Formalisms.” Commun. ACM 31(5), May 1988,
pp. 514−530.
Hehner 1984.
Eric. C. R. Hehner. The Logic of Programming. Prentice Hall, Englewood
Cliffs, NJ, 1984.
Henderson-Sellers 1994.
Brian Henderson-Sellers. Book Two of Object-Oriented Knowledge: The
Working Object. Prentice Hall, Englewood Cliffs, NJ, 1994.
Hoare 1969.
C. Anthony R. Hoare. “An Axiomatic Basis for Computer Programming.”
Commun. ACM 12(10), Oct. 1969, pp. 576−583.
Hoare 1978.
C. Anthony R. Hoare. “Communicating Sequential Processes.” Commun.
ACM 21(8), Aug. 1978, pp. 666−677.
Hofstadter 1980.
Douglas R. Hofstadter. Gödel, Escher, Bach: An Eternal Golden Braid.
Random House, Vintage Books, New York, 1980.
Hucklesby 1989.
Philip Hucklesby and Bertrand Meyer. “The Eiffel Object-Oriented Parsing
Library.” Proc. Technology of Object-Oriented Languages and Systems
(TOOLS 1, Paris, Nov. 1989), SOL, Paris, 1989, pp. 501−507.
ISO 1990.
ISO WD 9241-11. Ergonomic Requirements for Office Work with Visual
Display Terminals (VDTs), Part 11: Usability Statements, Version 2.5
Committee draft. International Organization for Standardization, 1990.
Jacobson 1992.
Ivar Jacobson, Magnus Christerson, Patrik Jonsson, and Gunnar Övergaard.
Object-Oriented Software Engineering: A Use Case Driven Approach.
Addison-Wesley, Wokingham, England, 1992.
Johnson 1992.
Ralph E. Johnson. “Documenting Frameworks using Patterns.” Proc.
Object-Oriented Programming Systems, Languages, and Applications
(OOPSLA ’92) in ACM SIGPLAN Notices 27(10), Vancouver, Canada, Oct.
 REFERENCES 417

1992, pp. 63−76.

Johnson 1988.
Ralph E. Johnson and Brian Foote. “Designing Reusable Classes.” Journal
of Object-Oriented Programming 1(2), June/July 1988, pp. 22−30, 35.
Kay 1977.
Alan C. Kay. “Microelectronics and the Personal Computer.” Scientific
American, Sept. 1977, pp. 231−244.
Kim 1990.
Won Kim. Introduction to Object-Oriented Databases. The MIT Press,
Cambridge, MA, 1990.
Korson 1990.
Tim Korson and John D. McGregor. “Understanding Object-Oriented: A
Unifying Paradigm.” Commun. ACM 33(9), Sept. 1990, pp. 40−60.
Krasner 1988.
Glenn E. Krasner and Stephen T. Pope. “A Cookbook for Using the Model-
View-Controller User Interface Paradigm in Smalltalk-80.” Journal of
Object-Oriented Programming 1(3), Aug./Sept. 1988, pp. 26−41, 48−49.
Kraut 1983.
R. Kraut, S. Hanson, and J. Farber. “Command Use and Interface Design.”
Proc. ACM Human Factors in Computing Systems (CHI ’83), Boston, MA,
Dec. 1983.
Krol 1992.
Ed Krol. The Whole Internet: User’s Guide and Catalog. O’Reilly and
Associates, Inc., Sebastopol, CA, 1992.
Krueger 1992.
Charles W. Krueger. “Software Reuse.” Computing Surveys 24(2), June
1992, pp. 131−183.
Lakoff 1980.
George Lakoff and Mark Johnson. Metaphors We Live By. University of
Chicago Press, Chicago, 1980.
Lano 1994.
Kevin Lano and Howard Haughton (eds). Object-Oriented Specification
Case Studies. Prentice Hall, Hemel Hempstead, England, 1994.
LeJacq 1991.
Jean Pierre LeJacq. “Semantic-Based Design Guidelines for Object-
Oriented Programs.” In Journal of Object-Oriented Programming: Focus
on Analysis and Design. SIGS Publications, New York, 1991, pp. 86−97.
Lieberherr 1989.
Karl L. Lieberherr and Ian M. Holland. “Assuring Good Style for Object-
Oriented Programs.” IEEE Software 6(5), Sept. 1989, pp. 38−48.
418 REFERENCES

Lieberman 1986.
Henry Lieberman. “Using Prototypical Objects to Implement Shared
Behavior in Object-Oriented Systems.” Proc. Object-Oriented
Programming Systems, Languages, and Applications (OOPSLA ’86) in
ACM SIGPLAN Notices 21(11), Portland, OR, Nov. 1986, pp. 214−223.
Linton 1989.
Mark A. Linton, John M. Vlissides, and Paul R. Calder. “Composing User
Interfaces with InterViews.” IEEE Computer 22(2), Feb. 1989, pp. 8−22.
Loomis 1991.
Mary E. S. Loomis. “Objects and SQL.” Object Magazine 1(3), 1991, pp.
68−78.
Lovgren 1994.
John Lovgren. “How to Choose Good Metaphors.” IEEE Software 11(3),
May 1994, pp. 86−88.
Marmolin 1993.
Hans Marmolin. User Centered System Design. UI Design AB (in
Swedish), Linköping, Sweden, June 1993.
Marshak 1991.
David S. Marshak. “ANSA: A Model for Distributed Computing.”
Network Monitor 6(11), Patricia B. Seybold, Boston, MA, Nov. 1991.
Martin 1992.
James Martin and James J. Odell. Object-Oriented Analysis and Design.
Prentice Hall, Englewood Cliffs, NJ, 1992.
McIlroy 1976.
M. Douglas McIlroy. “Mass-produced Software Components.” In Software
Engineering Concepts and Techniques (1968 NATO Conf. on Software
Engineering), J. M. Buxton, Peter Naur, and Brian Randell (eds). Van
Nostrand Reinhold, New York, 1976, pp. 88−98.
Mellor 1993.
Stephen J. Mellor, Sally Shlaer, Grady Booch, James Rumbaugh, Jim
Salmons, Timlynn Babitsky, Sam Adams, and Rebecca J. Wirfs-Brock.
“Open letter to the industry: Premature Methods Standardization Considered
Harmful.” Journal of Object-Oriented Programming 6(4), July/Aug. 1993.
Meyer 1985.
Bertrand Meyer. “On Formalism in Specifications.” IEEE Software 2(1),
Jan. 1985, pp. 6−26.
Meyer 1988a.
Bertrand Meyer. Object-Oriented Software Construction. Prentice Hall,
Hemel Hempstead, England, 1988.
 REFERENCES 419

Meyer 1988b.
Bertrand Meyer. “Cépage: Toward Computer-Aided Design of Software.”
The Journal of Systems and Software 8(5), Dec. 1988, pp. 419−429.
Meyer 1990a.
Bertrand Meyer. Introduction to the Theory of Programming Languages.
Prentice Hall, Hemel Hempstead, England, 1990.
Meyer 1990b.
Bertrand Meyer. “Tools for the New Culture: Lessons from the Design of
the Eiffel Libraries.” Commun. ACM 33(9), Sept. 1990, pp. 69−88.
Meyer 1992a.
Bertrand Meyer. Eiffel: The Language. Prentice Hall, Hemel Hempstead,
England, 1992.
Meyer 1992b.
Bertrand Meyer. “Design by Contract.” In Advances in Object-Oriented
Software Engineering, Dino Mandrioli and Bertrand Meyer (eds). Prentice
Hall, Hemel Hempstead, England, 1992, pp. 1−50.
Meyer 1992c.
Bertrand Meyer. “Applying ‘Design by Contract’.” IEEE Computer
25(10), Oct. 1992, pp. 40−51.
Meyer 1993a.
Bertrand Meyer. “What is an Object-Oriented Environment?” Journal of
Object-Oriented Programming 6(4), July/Aug. 1993, pp. 75−81.
Meyer 1993b.
Bertrand Meyer. “Systematic Concurrent Object-Oriented Programming.”
Commun. ACM 36(9), Sept. 1993, pp. 56−80.
Meyer 1994a.
Bertrand Meyer. Reusable Software: The Base Object-oriented Component
Libraries. Prentice Hall, Hemel Hempstead, England, 1994.
Meyer 1994b.
Bertrand Meyer. An Object-Oriented Environment: Principles and
Application. Prentice Hall, Hemel Hempstead, England, 1994.
Miller 1956.
G. A. Miller. “The Magical Number Seven, Plus or Minus Two: Some
Limits on Our Capacity for Processing Information.” The Psychological
Review 63(2), Mar. 1956, pp. 81−97.
Miller 1975.
G. A. Miller. “The Magical Number Seven after Fifteen Years.” In Studies
in Long Term Memory, A. Kennedy (ed.). John Wiley, New York, 1975.
Mills 1975.
Harlan D. Mills. “How to Write Correct Programs and Know It.” Proc. Int.
Conf. Reliable Software, Los Angeles, CA, Apr. 1975, pp. 363−370.
420 REFERENCES

Moran 1983.
T. Moran. “Getting into a System: External, Internal Task Mapping
Analyses.” Proc. ACM Human Factors in Computing Systems (CHI ’83),
Boston, MA, Dec. 1983.
Nerson 1991.
Jean-Marc Nerson. “Extending Eiffel Toward O-O Analysis and Design.”
Proc. Technology of Object-Oriented Languages and Systems (TOOLS 5,
Santa Barbara, July/Aug. 1991), assoc. paper, Prentice Hall, Englewood
Cliffs, NJ, 1991, pp. 377−392.
Nerson 1992a.
Jean-Marc Nerson. “O-O development of a date and time management
cluster.” Journal of Object-Oriented Programming 5(1), Mar./Apr. 1992,
pp. 39−46.
Nerson 1992b.
Jean-Marc Nerson. “Applying Object-Oriented Analysis and Design.”
Commun. ACM 35(9), Sept. 1992, pp. 63−74.
Nygaard 1970.
Kristen Nygaard, Ole-Johan Dahl, and Bjørn Myhrhaug. The Simula 67
Common Base Language. Norwegian Computing Center, Publ. No. S−22,
Oslo, Oct. 1970.
Ogden 1923.
C. K. Ogden and I. A. Richards. The meaning of meaning. London, 1923.
OMG 1990.
OMG, Object Management Group and X/Open. “Object Management
Architecture Guide 1.0.” OMG TC Document 90.9.1, Nov. 1990.
OMG 1991.
OMG, Object Management Group and X/Open. “The Common Object
Request Broker: Architecture and Specification.” OMG Document Number
91.12.1, Dec. 1991.
Omohundro 1991.
Stephen M. Omohundro. “The Sather Language.” Report: International
Computer Science Institute, Berkeley, CA, 1991.
Page-Jones 1990.
Meilir Page-Jones, Larry L. Constantine, and Steven Weiss. “Modeling
Object-Oriented Systems: The Uniform Object Notation.” Computer
Language 7(10), Oct. 1990, pp. 69−87.
Parnas 1972.
David Lorge Parnas. “On the Criteria To Be Used in Decomposing Systems
into Modules.” Commun. ACM 15(12), Dec. 1972, pp. 1053−1058.
 REFERENCES 421

Parnas 1979.
David Lorge Parnas. “Designing Software for Ease of Extension and
Contraction.” IEEE Trans. Software Engineering SE-5(2), Mar. 1979, pp.
128−137.
Parnas 1993.
David Lorge Parnas. “Predicate Logic for Software Engineering.” IEEE
Trans. Software Engineering 19(9), Sept. 1993, pp. 856−862.
Parnas 1986.
David Lorge Parnas and Paul C. Clements. “A Rational Design Process:
How and Why to Fake It.” IEEE Trans. Software Engineering SE-12(2),
Feb. 1986, pp. 251−257.
Penrose 1989.
Roger Penrose. The Emperor’s New Mind. Oxford University Press, New
York, 1989.
Premerlani 1994.
William J. Premerlani and Michael R. Blaha. “An Approach for Reverse
Engineering of Relational Databases.” Commun. ACM 37(5), May 1994,
pp. 42−49.
Prieto-Diaz 1987.
Ruben Prieto-Diaz. “Domain Analysis for Reusability.” Proc. IEEE
Computer Software and Applications Conference (COMPSAC ’87), Tokyo,
Oct. 1987, pp. 23−29.
Quarterman 1990.
John S. Quarterman. The Matrix: Computer Networks and Conferencing
Systems Worldwide. Digital Press, Digital Equipment Corporation,
Maynard, MA, 1990.
Rahayo 1993.
Wenny J. Rahayo and Elizabeth Chang. “A Methodology for Transforming
an Object-Oriented Data Model to a Relational Database.” Proc.
Technology of Object-Oriented Languages and Systems (TOOLS 12,
Melbourne, Australia, Nov./Dec. 1993), Prentice Hall, Englewood Cliffs,
NJ, 1993.
Reenskaug 1992.
Tryggve Reenskaug, Egil P. Andersen, Arne Jorgen Berre, Anne Hurlen,
Anton Landmark, Odd Arild Lehne, Else Nordhagen, Eirik Nêss-Ulseth,
Gro Oftedal, Anne Lise Skaar, and Pål Stenslet. “OORASS: Seamless
support for the creation and maintenance of object oriented systems.”
Journal of Object-Oriented Programming 5(6), Oct. 1992, pp. 27−41.
Robinson 1992.
Peter Robinson. Hierarchical Object-Oriented Design. Prentice Hall,
Englewood Cliffs, NJ, 1992.
422 REFERENCES

Rubin 1992.
Kenneth S. Rubin and Adele Goldberg. “Object Behavior Analysis.”
Commun. ACM 35(9), Sept. 1992, pp. 48−62.
Rumbaugh 1991.
James Rumbaugh, Michael Blaha, William Premerlani, Frederick Eddy, and
William Lorensen. Object-Oriented Modeling and Design. Prentice Hall,
Englewood Cliffs, NJ, 1991.
Scott 1991.
A. C. Scott, J. E. Clayton, and E. L. Gibson. A Practical Approach to
Knowledge Acquisition. Addison-Wesley, Reading, MA, 1991.
Seidewitz 1987.
E. Seidewitz and M. Stark. “Towards a General Object-Oriented Design
Methodology.” Ada Letters 7(4), July/Aug. 1987.
Shlaer 1992.
Sally Shlaer and Stephen J. Mellor. Object Lifecycles: Modeling the World
in States. Prentice Hall, Englewood Cliffs, NJ, 1992.
Shneiderman 1983.
Ben Shneiderman. “Direct Manipulation: A Step Beyond Programming
Languages.” IEEE Computer 16(8), Aug. 1983, pp. 57−69.
Siemens 1993.
Siemens Nixdorf Informationssysteme AG. “MooD V1.0—Methodology
for Object-Oriented Development: Introduction.” Frankfurt / Main,
Germany, 1993.
Stroustrup 1992.
Bjarne Stroustrup. The C++ Programming Language. 2nd edn, Addison-
Wesley, Reading, MA, 1992.
Strunk 1979.
William Strunk, Jr. and E. B. White. The Elements of Style. Macmillan,
New York, 1979.
Ungar 1987.
David Ungar and Randall B. Smith. “Self: The Power of Simplicity.” Proc.
Object-Oriented Programming Systems, Languages, and Applications
(OOPSLA ’87) in ACM SIGPLAN Notices 22(12), Orlando, FL, Oct. 1987,
pp. 227−242.
Velho 1992.
Amândio Vaz Velho and Rogerio Carapuça. “SOM—A Semantic Object
Model: Towards an Abstract, Complete and Unifying Way to Model the
Real World.” Proc. Third Int. Conference on Dynamic Modeling of
Information Systems, Noordwijkerhout, The Netherlands, 1992, pp. 65−93.
 REFERENCES 423

Velho 1994.
Amândio Vaz Velho and Rogerio Carapuça. “Attribute: A Semantic and
Seamless Construct.” Proc. Technology of Object-Oriented Languages and
Systems (TOOLS 13, Versailles, France, Mar. 1994), Prentice Hall, Hemel
Hempstead, England, 1994.
Wasserman 1990.
Anthony Wasserman, Peter A. Pircher, and Robert J. Muller. “The Object-
Oriented Structured Design Notation for Software Design Representation.”
IEEE Computer 23(3), Mar. 1990, pp. 50−63.
Wilkie 1993.
George Wilkie. Object-Oriented Software Engineering: The Professional
Developer’s Guide. Addison-Wesley, Wokingham, England, 1993.
Wirfs-Brock 1991.
Allen Wirfs-Brock. “Does Smalltalk need types?” Journal of Object-
Oriented Programming 4(6), Oct. 1991, pp. 62−67.
Wirfs-Brock 1989.
Rebecca J. Wirfs-Brock and Brian Wilkerson. “Object-Oriented Design: A
Responsibility-Driven Approach.” Proc. Object-Oriented Programming
Systems, Languages, and Applications (OOPSLA ’89) in ACM SIGPLAN
Notices 24(10), New Orleans, LA, Oct. 1989, pp. 71−75.
Wirfs-Brock 1990.
Rebecca J. Wirfs-Brock, Brian Wilkerson, and Lauren Wiener. Designing
Object-Oriented Software. Prentice Hall, Englewood Cliffs, NJ, 1990.
Wirth 1971.
Niklaus Wirth. “Program Development by Stepwise Refinement.”
Commun. ACM 14(4), Apr. 1971, pp. 221−227.
Zloof 1977.
M. M. Zloof. “Query-by-Example: A Data Base Language.” IBM Systems
Journal 16(4), 1977, pp. 324−343.
424 REFERENCES
Index
BON keywords are in boldface. Page numbers in boldface indicate places of
definition or main discussions of the corresponding topic. Page numbers in
italics refer to a corresponding entry in appendix E: “Glossary of terms.”

A ANSA (Advanced Network Systems

Architecture), 408
Abstract class, see Class, deferred Anthropomorphism, 185
Abstract data type, 179−180, 183, 186, 206, Application, 389
242, 388 Architecture, 389
Abstract feature, see Feature, deferred Assertion, see under Software contracting
Abstraction, 388 Assertion language, see BON assertion
for reuse, 170−171 language
layer of, 183, 217, 221 Association, see under Client
premature, 220−221 Attribute, 389
strong, 221, 222 abstract, 50
weak, 221, 222 change, see Delta
Access control, 388 exported, 35
Accidents, see under Software engineering of class, see under Class
Acid test for object-orientedness, 225 Attribute (relational), 295
Active object, 388
Actor, 388
Ada, 167 B
ADT, see Abstract data type
Agent, 389 Barks, Carl, 94
Aggregation, see under Client Behavior, 389
Aggregation relation, 389 Bidirectional link, see under Client and
Alexander, Christopher, 217 Message
Algol, 5 Binary association, 389
Analysis, 389, 400 Binding, 389
meaning of, 122−123, 151, 153, Blind checking, see Defensive programming
172−173 BNF (Backus Naur Form), 350
recursive, 123 BON acronym, 11
tasks, see under BON process BON activities, 120
Analysis class, 9, 98, 122, 136, 143, 151, BON approach, 11−25
172, 182, 184, 197, 252, 260, 275, 277 general, 11−12
Ancestor, see under Class position, 16−17

425
426 INDEX

what is not included, 12−15 364−378

Bonasonic, 270, 278, 282 typing, 19−20
BON assertion language, 49−59 BON process, 120, 143−177
basic boolean expressions, 51−52 acceptance criteria, 150, 151, 152
basic object expressions, 50−51 analysis tasks, 151, 152−172
logical quantifiers, 53−54 candidate classes, 160−161, 235−236,
numeric expressions, 51 274, 390
predicate logic, 52−57, 402 class definition, 150, 166−167, 242−251,
propositions, 55−56 279−283
quantified variables, 54−55 class selection, 162−166, 236−241,
range restriction, 56−57 275−279
relational expressions, 51 clustering, 162−166, 236−241, 275−279
semi-strict boolean operators, 47, 51 completing the design, 175−177
set operations, 53 design tasks, 151, 172−177
state changes, 50 feature definition, 171−172, 260−269,
BON deliverables, 143−148, 152 286
roles of, 146−148 finding classes vs. objects, 160−161
BON dynamic model, 90−116, 168−171 formal contracts, 171−172
dynamic diagram, 99−116 generalization, 174−175
external event, 93−94 input sources, 152
internal event, 93−94 refining system, 172−174
object creation, 168−169 system behavior, 168−171, 252−259,
object-oriented execution, 91, 164, 400 283−285
purpose, 90 system borderline, 152−160, 232−235,
relations, see under Message 270−274
system event, 93−96, 156, 407 BON standard activities, 178−228
telephone call analogy, 185−186 assessing reuse, 212−215
what should be captured, 92−93 classifying, 189−194
BON modeling charts, 30−35, 93, 143−145 clustering, 194−198
class chart, 32−35, 242−251, 263, defining class features, 198−202
279−283 evolving the system architecture,
cluster chart, 32, 280 220−228
creation chart, 97−99, 157, 168, 176 finding classes, 178−189
event chart, 94−96, 169, 176, 252, 253, indexing and documenting, 215−220
254, 283 selecting and describing object scenarios,
scenario chart, 97, 157, 169, 252−253, 203−205
255, 283−284 working out contracting conditions,
system chart, 31−32, 157, 238 205−212
BON notation BON static model, 29−64, 65−89
and case tools, 24 relations, see under Client and
characteristics of, 17−24 Inheritance
economy of space, 23 BON textual examples, 364−378
generality, 17 class headers, 367−368
graphical form, 29, 45 client relations, 369−373
reversibility, 17−18 cluster structure, 367−368
scalability, 18−19 dynamic diagrams, 374−378
seamlessness, 17 informal charts, 365−367
simplicity, 22 inheritance relations, 367−369
textual form, 29−30, 43, 44, 349−362, static diagrams, 367−374
 INDEX 427

BON textual grammar, 349−362 attribute, 35

class interface description, 355−357 basic layer, 199
comments, 360 behavior, 32−35, 160−161
construct, 350−352 child, 65, 390
dynamic diagrams, 358−359 class or operation, see under Modeling
formal assertions, 357−358 completeness, 188
free operators, 360 compression of, see under Compressed
identifier, 359 form
informal charts, 352−354 constraint, see Constraint
lexical components, 359−362 convenience feature, 199
notational tuning, 359 de facto, 135, 160, 236
production, 350 deferred, 38, 39, 98−99, 162, 169, 183,
reserved words, 361, 361−362 187, 198, 223, 239, 244, 248, 274, 278,
special symbols, 362, 363 393
static diagrams, 354−355 definition, see under BON process
strings, 360−361 dependencies, 69−70, 86−87
syntax notation, 350−352 descendant, 65, 394
Booch, Grady, 110, 200 distinction from instance, 100, 200
BOOLEAN, 260, 261 effective, 38, 162, 187, 198, 244, 395
Boolean algebra, 389 factored dependencies, 82−83, 86−87
Boolean expression, see under BON feature, see Feature
assertion language flattener, 217
Bottom-up design, 223, 390 formal interface, see typed interface
Broker, 197, 390 generic, see parameterized
Brooks, Fred, 3 header annotations, 37, 38, 59, 60
Business rules, 390 how to find, see under BON standard
activities
implementation oriented, 192
C instantiation, 74, 398
interfaced, 38, 162
C, 5 invariant, 10, 21, 35, 47−48, 50, 55, 56,
C++, 19, 47, 126, 190, 217, 398, 400, 405, 90, 171, 187, 194, 208, 209, 210, 212,
407, 408, 409 217, 243, 260, 263, 266, 269, 390
Candidate class, see under BON process layered design, 188
Case discrimination, 208, 225, 240, 390 letter case conventions, 60, 64, 100, 200
Case studies mixin, see Mixin
1, see Conference management system naming, 181−182
2, see Video recorder number of features, 187
3, see Relational and O-O coexistence object, 405
Celsius, A., 305 operation, see Feature
Changed attribute, see Delta parameterized, 38, 162, 401
Chart, see BON modeling charts parent, 65, 401
Child, see under Class pattern of behavior, 161
Class, 390 persistent, 38, 248
abstract, see deferred relations, see under Inheritance and
abstractor, 217 Client
ancestor, 65, 389 representation, see under Graphical and
as blueprint for object creation, 25 Textual representation
as specification element, 189, 209−210 responsibility, see under Software
428 INDEX

contracting sharing marker, 81

reused, 38 unidirectional link, 72, 86
root, 38, 176, 239, 251, 404 usage vs. inheritance, 190−191
split and merge, 176 Client−server, 196−197, 215, 391
typed interface, 19−20, 36−49, 171−172, Client vs. supplier role, 183−185
216, 260−269, 286 CLOS, 192, 399
untyped interface, see under BON Cluster, 391
modeling charts client relations, see under Client
Class attribute, 390 compression of, see under Compressed
Class chart, see under BON modeling charts form
Classification, 20, 61, 109, 124, 125, 136, inheritance, see under Inheritance
163, 165, 176, 189, 192, 195, 198, 216, letter case conventions, 62
240, 241, 244, 246, 249, 264, 278, 391 name, 62
Class instance, 390 representation, see under Graphical and
Class interface, 390 Textual representation
Class method, 391 semantics of compressed relations,
Class-oriented technology, 161 66−69, 84−86
Client, 10, 391 tag, 62
aggregation, 70−71, 262, 276, 389 Cluster chart, see under BON modeling
aggregation and value types, 80 charts
aggregation link, 71 Clustering, 18−19, 60−64, 157, 162−166,
association, 70, 389 236−241, 275−279, 391
association link, 71 by gradual abstraction, 62−63
bidirectional link, 72−73, 86 by informal charts, 31−32
compacted generic link, 76−78 client−server interface, 196−197
compression of relation, see under cohesion and coupling, 196, 391, 392
Compressed form different views of same classes, 163
entity, 80, 84 gradual compression, 67−68
generic label marker, 75 hierarchical structure, 62, 195
hiding local dependencies, 108, 253−255 interface classes, 196
indirect dependencies, 88 layered approach, 239
instance multiplicity, 80, 397 local naming, 63−64
instance sharing, see shared association nested cluster, 32, 61
label, 71, 84 points of view, 60−61, 236
label placement, 72, 73−74 reused cluster, 62
link, 71 same level of abstraction, 197
links involving generic parents, 78 system view, 61, 163−164, 195, 408
links involving generic suppliers, 74−78 Codd, E. F., 297
multidirectional link, 73−74, 86 Cohesion, see under Clustering
multiplicity marker, 78 Collaboration, 163, 189, 391
object sharing, see shared association Column (relational), 295
proxy, 197 Command, see under Feature
relation, 69−86, 253−255 Comments, see under BON textual grammar
relations involving clusters, 83−86 Component library, 392
responsibility, 41, 209 Compressed form, 19, 392, 396
role, 70, 71, 404 of class, 60
role multiplicity, 78−80, 404 of client labels, 86
role naming, 79 of client relations, 85−86
shared association, 70, 81−83, 398 of cluster, 64
 INDEX 429

of inheritance relations, 66−69 Data-driven design, 180

of interface sections, 36 Data flow, 156, 257
of message relations, 105 Data flow analysis, 92
of object group, 104−105 Data flow diagram, 393
rules for, 36, 60, 66−69, 85−86, 104−105 Data hiding, see Information hiding
Compression, see Compressed form Data integrity, 294
Concatenator, 359, 365 Data modeling, 393
Concrete object, see Modeling, physical DBMS, 393
object Deep persistency principle, 292
Concurrent modeling, 165 Defensive programming, 206, 209, 393
Conference management system, 231−269 Deferred class, see under Class
candidate classes, 235−236 Deferred feature, see under Feature
class definition, 242−251 Delegation, 394
class selection and clustering, 236−241 Delta, 46, 50
final static architecture, 268 Demeter, Law of, 222
formal class description, 260−269 Department of Defense, 149
system behavior, 252−259 Derived attribute, see under ER modeling
system borderline, 232−235 Descendant, see under Class
Constantine, Larry, 196 Design, 394, 400
Constrained generic derivation, 192−193 detailed, 16, 36, 39, 197, 260, 270, 279,
Constraint, 33, 166, 171, 238, 243, 244, 247, 286
263, 392 meaning of, 122−123, 151, 153,
Construct, see under BON textual grammar 172−173
Constructor, 392 of libraries, see Library design
Container class, 53, 55, 162, 188, 189, 202, recursive, 123
256, 261, 370, 392 tasks, see under BON process
Contract model, see Software contracting Design by contract, 206−212, 394
Contract violation, see under Software Design class, 9, 122, 151, 172, 173, 182
contracting Destructor, 394
Contravariant rule, 42, 392 Dijkstra, Edsger, 123, 185, 214
Control flow, 392 Distributed applications, 196
Correctness, see under Software contracting Domain analysis, 169, 394
Coupling, see under Clustering Domain expert, 29, 30, 143, 161, 162, 167,
Covariant rule, 42, 392 169, 182, 394
CPU, 226 Domain integrity, see under Integrity
CRC method, 29, 30 constraints
Creation chart, see under BON modeling Domain (relational), 295
charts Dual role of designer, 183−185
CSP language, 49 Dummy object, see under Object
Current, 43, 45, 46, 47, 55, 211 Dynamic binding, 14, 56, 82−83, 86−87,
Current object, 50, 393 108−109, 112, 167, 208, 225−226, 240,
Current result, see Result 394
Dynamic chart, 395
Dynamic diagram, see Object diagram
D Dynamic model, 24−25, 145, 146, 147, 395
graphical editor example, 108−112, 374
Data abstraction, 393 separation from static model, 25, 100
Database, 393 Dynamic type, 395
Data dictionary, 20, 393 Dynamic typing, see under Typing
430 INDEX

E Extensibility, 396
External communication, 156−157, 257
Editor example, see under Dynamic model External event, see under BON dynamic
Effective class, see under Class model
Effective feature, see under Feature
Egocentric naming, 180
Eiffel, 18, 19, 35, 47, 126, 199, 202, 217, F
228, 400, 402, 405, 407, 408
EiffelBuild, 126 Factoring, see Modeling, generalization
Eiffel parsing library, 228 Fahrenheit, G. D., 305
EiffelVision, 126, 228 Feature, 35, 39−46, 396
Encapsulation, see Information hiding abstract, see deferred
End-user, 29, 30, 167, 173, 203, 242 arguments, see signature
Enterprise modeling, 121−122, 153 command, 33, 166, 171, 260, 282, 392
Entity, 395 deferred, 39−40, 198, 388, 394
Entity−relationship, see ER modeling effective, 40, 395
Entity−relationship diagram, 395 header comment, 219
ER modeling, 16, 395 implemented, see effective
case against, 12−15 letter case conventions, 64, 200
derived attribute, 14, 394 local variable, 19, 88, 108, 200, 210, 253
transitive law, 13 name, 39−40
Essence, see under Software engineering naming, 199−205
Event, 395 number of arguments, 201
Event chart, see under BON modeling charts pre- and postcondition, 10, 21, 35,
Event handler, 279, 284, 285 41−46, 48, 50, 90, 171, 183, 194, 206,
Event trace diagram, 92, 93, 110, 395 207, 208, 209, 212, 217, 219, 225, 226,
Exception handling, see under Software 243, 260, 402
contracting private, 39, 253, 402
Execution model, see under BON dynamic public, 36, 39, 43, 143, 151, 171, 176,
model 403
Exercises, 332−346 query, 32, 166, 171, 260, 282, 403
assertions and classification, 335 redefinition, 39−41, 217, 264−266, 403
car rental company, 340−341 renaming, 40−41, 217, 266, 403
class relationships, 334−335 restricted, 39, 173, 404
dice game (Greed), 339−340 signature, 19, 40−41, 171, 260, 405
dynamic behavior (claim processing), type, see signature
336−337 Feature call, 41, 44, 47, 50, 396
prescription and description (cabbage, Finite state machine, see FSM
lamb, and wolf), 337−338 First-order logic, 22, 30, 52, 396
problem domain clustering (viewpoints), For_all, 43, 45, 46, 53, 54, 55, 56, 57, 58,
332−333 211, 267
real-time process control (perfume Formal interface, see Class, typed interface
production), 345−346 Formal specification, see BON assertion
traffic control system, 338−339 language
truck freight, 341−345 Formal syntax, see BON textual grammar
Existential quantifier, see Exists Formatted memo, 242
Exists, 43, 45, 46, 53, 54, 55, 58, 211 Fortune magazine, 179
Expanded form, see Compressed form Framework, 6, 197−198, 202, 216, 396
Expansion, see Compressed form black-box, 196, 198, 224
 INDEX 431

documentation, 217−218 Hybrid technique, 397

toolkits, see under Library design Hybrid vs. pure techniques, 12−15, 167
white-box, 198, 224
Free Software Foundation, 158
FSM, 92, 228, 279, 396 I
FSM modeling, 12, 16, 17, 92−93, 97, 396
Fuzzy requirements, 157, 252 IBM, 200
Icon, 19
Identifier, see under BON textual grammar
G Identity, see Object identity
Implementation, 397
Garbage collection, 174, 290, 396 as road to abstraction, 224
Generalization, see under Modeling Implementation class, 182
Generic class, see Class, parameterized Implemented feature, see Feature, effective
Generic derivation, 74−78 Incremental development, 157−158, 397
Glazier, 227 Indexing, 397
Glossary of terms, see under Problem Indexing clause
domain of BON modeling chart, 32
Graphical BON, see under BON notation of typed interface, 37
Graphical editor example, see under Indexing guidelines, 218−219
Dynamic model Informal chart, see BON modeling charts
Graphical representation Informal interface, see BON modeling charts
of call, see Message Information hiding, 397
of class header, 59, 60 Inheritance, 397
of class interface, 36−48 as classification, 163, 240−241
of client relations, 71−86 compression of, see under Compressed
of cluster, 62 form
of inheritance relations, 65−68 for grouping, 192
of message, 101−103 involving clusters, 66−69
of object, 100 link, 65
of object group, 104 multiple, 40, 65, 66, 174, 175, 176, 191,
of object set, 100 193, 209, 217, 241, 399
vs. textual representation, 48−49 of facilities, 191
Green Delivery, 207 of non-orthogonal concepts, 194
Group, see Object group of secondary properties, 192−193
GUI (Graphical User Interface), 94, 126, of specification vs. representation, 224
139, 153, 173, 227, 228 protecting descendants, the case against,
190
relation, 65−69
H repeated, 65, 66, 175, 403
single, 65, 66, 175, 406
Hardware encapsulation, 270, 273, 274, 277 vs. client usage, 190−191
Harel, David, 17, 93 Inheritance clause
Hierarchy, 397 of BON class chart, 34
High-level language, 5 of typed interface, 39
Hofstadter, Douglas, 130 Instance, see Class instance
Hugo, Victor, 49 Instance multiplicity, see under Client
Human−computer interaction, see User Instance sharing, see under Client
centered design Instance variable, 398
432 INDEX

Integrity constraints, 294−295 M

domain integrity, 294
referential integrity, 294 Macintosh, 134, 138
user-defined integrity, 294 Marmolin, Hans, 130
Interaction diagram, see Event trace diagram McIlroy, Douglas, 205, 406
Interface MCU (Micro Computer Unit), 270, 272, 277
formal, see Class, typed interface Member function, 398
informal, see BON modeling charts Member_of, 43, 45, 46, 53, 54, 55, 56, 57,
Interface class, 398 265, 267, 286, 287
Internal event, see under BON dynamic Memo, 242, 263
model Message, 398
Internet, 158 bidirectional link, 102
InterViews, 126 caller, 102
Invariant, see under Class compression of, see under Compressed
ISO (International Organization for form
Standardization), 129, 296 concurrent, 103
It_holds, 43, 45, 46, 55, 56, 57, 58, 211, 267 equivalence with call, 101
label, 101
label placement, 102, 103
J link, 101
multidirectional link, 102−103
Join (relational tables), 297 receiver, 102
recursion, 105
relation, 93, 101−106
K time order, 101
Message passing, see Feature call
Keyword_prefix, 359, 365 Metaclass, 398
Metalanguage, 398
Method, 398
L faking an ideal process, 120−121
generic approach, 119
Language design, 180 no substitute for skill, 119, 120
Late binding, 398 vs. methodology, 120
Law of Demeter, 222 Method (of a class), 398
Layer, 398 Methodology, 119, 120, 399
Level of abstraction, 398 Mills, Harlan, 142
Lexical components, see under BON textual MIT, 158
grammar Mixin, 192, 399
Library design, 226−228 Modeling
consistency, 226−227 class or operation, 188
inversion of control, 227 conceptual summaries, 224−225
precondition checking, 226 confusing inheritance with aggregation,
toolkits, 227−228 193−194
Link, 398 factoring, see generalization
Linnaeus, Carolus, 391 generalization, 150, 174−175, 222−223,
Local variable, see under Feature 396
Logical quantifier, see under BON assertion human problem solving, 121, 132
language importance of free classification, 136
Love, Tom, 339 interface classes, 139−140
 INDEX 433

no a priori structure, 124−125, 142, 236, NUMERIC, 262

288 Numeric expressions, see under BON
physical object, 125, 135, 182−183, 236, assertion language
270, 274, 288, 402
problem vs. solution, 124−125 O
puzzle analogy, 165−166
roles of analysis objects, 135−136 OBA (Object Behavior Analysis), 169
several interface levels, 138 Object, 399
specialization, 222−223, 406 active, 106−107
strong design means strong caller, 253
dependencies, 221−222 cloning, 316
system front and backend, 137−138 communication, see Message
system interface, 136−140 creation, 97−99, 168−169, 191
Modeling chart, see BON modeling charts de facto, see under Class
Modula 2, 167 dependencies, 69−70, 86−87, 253−255
Moscow, 138 dummy, 191
Motif, 228 indirect call chain, 107
Multidirectional link, see under Client and letter case conventions, 100
Message life cycle, 174
Multiple inheritance, see under Inheritance multiplicity, see Client, instance
Multiple language metaphor, 123−124, 180, multiplicity
184 passive, 106−107
Multiplicity, 399 persistent, 162, 169, 174, 248, 263,
Multiplicity marker, see under Client 289−292, 293, 295, 301, 302, 309, 311,
MVC (Model−View−Controller), 126, 227 312, 315, 316, 402
physical, see under Modeling
qualifier, 100
N receiver, 253
relations, see under Client and Message
Name space, 191 representation, see under Graphical and
Naming, 399 Textual representation
abbreviation, 200−201 root, 137, 168, 169
consistency, 202−205 scenario, see Scenario
grammatical categories, 201 sharing, see under Client
importance of, 180−181 state, 35
of class, see under Class transient, 290, 408
of feature, see under Feature Object class, 399
overqualification, 200−201 Object diagram, 99−116, 399
style guidelines, 199−201, 202−205 claim settlement example, 114−115, 379
word separation, 200 moving nested group of figures, 109−112
NASA (Naval Aeronautic and Space need for grouping, 111−112
Administration), 149 need for two dimensions, 110−111
NATO (North Atlantic Treaty Organization), Object group, 103−105, 399
205, 406 compression of, see under Compressed
Nested cluster, see under Clustering form
Newton, Isaac, 207 label, 104
Nextstep, 126 letter case conventions, 104
Nip trouble spots in the bud, 149 Object grouping, 93, 103−106, 399
No Hidden Clauses rule, 208 Object identity, 399
434 INDEX

Object instance, 399 P

Objective-C, 126
Object management, 399 Panning, 15
Object-oriented, 399 Paradigm, 401
Object-oriented abstraction, 400 Paradigm shift, 142
as application language, see Multiple Parameterized class, see under Class
language metaphor Parent, see under Class
as investment, 6, 184 Parnas, David, 214
benefits of, 3−6, 16 Partial function, 51, 57, 58, 401
encapsulated behavior, 186−187 Partial routine, 401
generality of, 7, 9, 16 Pascal, 5, 187
guidelines and principles, 179 Passive object, 402
mappings to other domains, 16 Pattern language, 217
organizing functionality, 194−195 Pavlov, Ivan Petrovich, 126
Object-oriented analysis, see Analysis Penrose, Roger, 130
Object-oriented database, 169, 400 Persistent class, 292, 293, 300, 306, 310,
Object-oriented design, see Design 316, 402
Object-oriented encapsulation, 400 Persistent object, see under Object
Object-oriented execution, see under BON Persistent root, 292
dynamic model Petri net, 17
Object-oriented implementation, 400 PG toolkit, 228
Object-oriented method, see Method Physical object, see under Modeling
Object-oriented model, 400 Polymorphism, 14, 56, 108−109, 112, 167,
Object-oriented modeling, 400 176, 208, 225−226, 402
Object-oriented programming language, 400 Pre- and postcondition, see under Feature
Object-oriented software development, Predicate logic, see under BON assertion
3−10, 400 language
Object-oriented specification, 400 Premature abstraction, 220−221
Object-oriented system, 400 Premature standardization, 12
Object set, see under Graphical Primary key, 295, 302, 320
representation Private feature, see under Feature
Object SQL, 312 Problem domain, 48, 52, 58, 90, 120, 123,
Object state, 401 125, 127, 129, 132, 135, 141, 142, 143,
ODBMS, 401 150, 151, 152, 155, 156, 160, 161, 162,
Old, 43, 44, 44, 45, 46, 50 167, 168, 171, 172, 182, 189, 205, 221,
OMG (Object Management Group), 390 239, 274, 402
OMT, 110 glossary of terms, 161−162, 167
OOA, 401 Process structure, 164−165
OOD, 401 Production, see under BON textual grammar
OOPL, 401 Programming-in-the-large, 195, 213, 402
OOPSLA, 339 Program templates, 223
OOSE, 110, 115, 136, 140 Program verification, 206
Open−closed principle, 128 Property, 402
OpenLook, 228 Propositional logic, 402
Operation, 401 Protocol, 403
ORACLE, 192 Proxy, 403
ORB (Object Request Broker), 390 Pure technique, 403
Other methods, 15−16 Pure vs. hybrid techniques, 12−15, 167
Overloading, 401 Puzzle analogy, see under Modeling
 INDEX 435

Q Relational model, 295−297

Relation label, see Client, label
Quality assurance, 151 Relationship, see Relation
Quantified variables, see under BON Renaming, see under Feature
assertion language Repeated inheritance, see under Inheritance
Query, see under Feature Reserved words, see under BON textual
Query-by-Example, 312 grammar
Query frame, 312−313 Responsibility, see under Software
contracting
Responsibility-driven design, 180
R Result, 43, 45, 46, 50
Reusable library, 129, 404
README file, 202 Reuse, 4−6, 170−171, 212−215, 270, 277
Recursive analysis, see under Analysis assess impact of future changes, 215
Recursive design, see under Design benefits of, 4
Recursive language design, see Multiple changed attitude, 213
language metaphor component base, 4, 177, 215
Recursive specification, 52, 403 cost/benefit, 127
Redefinition, see under Feature data structures, 6
Referential integrity, see under Integrity encapsulation of existing software,
constraints 214−215
Relation, 403 graphical interfaces, 6
client, see under Client importance of quality, 9
compression of, 66−69, 83−86, 105 initial difficulties, 4−5
direction of arrows, 71−72 investment in human capital, 127
inheritance, see under Inheritance levels, 125−126
join, 66, 73−74, 103 look for components, 158
link, see under Client, Inheritance, and “not invented here” syndrome, 128
Message not only of classes, 214
message, see under Message of analysis and design elements, 17
Relational and O-O coexistence, 289−331 of class, 38
automatic data transfer, 299−301 of cluster, 62
database wrapper, 295−303 of knowledge, 5−6
dynamic routing, 317 of the complex parts, 213
factors determining the schema, 304−305 planned vs. accidental, 127−128, 190
full retrieval scenario, 321−324 policy, 126−127, 158−159
full static architecture, 331 reasons for failure, 5
generation of SQL statements, 324−331 reinventing the wheel, 125
interfacing an existing schema, 303−311 through discovery, 223
mapping non-basic attributes, 301−303 through generalization, 175
object models and relational models, Reused class, 404
293−295 Reuse manager, 128−129, 404
object persistence, 290−292 Reuse repository, 129
persistent object, 315−324 Reversibility, 6, 7−9, 11, 12, 15, 17−18, 69,
querying a persistent model, 312−315 92, 145−148, 176, 404
retrieval by example, 312−314 Risk-driven development, 148−149, 154
schema dependency, 305 Robustness, see under Software contracting
storage manager, 315−324 Role, see under Client
virtual database, 305−306 Role multiplicity, see under Client
436 INDEX

Role name, 404 Side effect, 212, 405

Role naming, see under Client Signal-to-noise ratio, 219
Root class, see under Class Signature, see under Feature
Root object, see under Object Similarity, 163, 189, 195, 225, 226
Row (relational), 295 Simplicity principle, 22
Simula, 206
Single choice principle, 226, 406
S Single inheritance, see under Inheritance
Smalltalk, 19, 20, 42, 126, 202, 227, 390,
Saint-Exupéry, Antoine de, 22 391, 395, 398, 400, 402, 403, 405, 407,
Sather, 217 408
Scalability, 18−19, 159, 405 Software components, see Reuse, component
level of description, 59, 67−69, 83−86, base
109 Software contracting, 6, 9−10, 11, 20−22,
Scenario, 96−97, 140, 143, 157, 169−171, 171, 205−212, 406
203−205, 405 and inheritance, 42, 208−209
analysis, 203−205 and partial functions, 57−59
as guide to static modeling, 170−171 as mutual benefit, 207−208
box, 101 assertion, 10, 21, 35, 41−48, 49−59, 243,
design oriented, 173 389
how many, 170 assertion monitoring, 210−212
internal, 145 CITIZEN example, 34, 42−48, 365
name, 101 class invariant, see under Class
of conference system, 252−259 consistency, 176
of VCR system, 283−285 contract violation, 10, 212
roles, 205 correctness, 10, 20
travel guide example, 204−205 exception handling, 212, 396
Scenario chart, see under BON modeling importance of precondition, 206
charts laws of subcontracting, 208−209, 217,
Seamlessness, 6−7, 11, 12, 14, 15, 20, 70, 225
91, 92, 135, 145, 175, 210, 216, 290, 405 pre- and postcondition, see under Feature
in both directions, see Reversibility responsibility, 209, 237, 243, 403
Self, 405 robustness, 20, 404
Self (language), 394 Software correctness, see under Software
Semantic data model, 405 contracting
Semantic label, 76, 78, 189, 405 Software engineering, 3, 7, 10, 14, 125, 127,
Semantic link, 65, 89, 189 406
Semantic specification, see BON assertion accidents of, 3, 4
language essence of, 3, 4
Semi-strict boolean operators, see under life cycle, 150, 158
BON assertion language separate generalization phase, 175
Server, 405 too much too soon, 213, 220, 222
Server proxy, 197 Software specification
Service, 405 and program code, see Reversibility
SET, 42, 48, 53, 56, 72, 99, 107, 110, 189, classes for, 189, 209−210, 260−262
244, 256, 257, 260, 261, 263 common abstraction base, 20, 21, 210
Set operations, see under BON assertion problems with traditional approaches,
language 7−9
Shared association, see under Client recursive approach, 21−22, 210
 INDEX 437

semantics, see BON assertion language grammar

usefulness, 8−9 System
Software system, 406 as a “black box”, 93, 136, 156
Specialization, see under Modeling delineate borderline, 150, 152−160,
Special symbols, see under BON textual 232−235, 270−274
grammar event, see under BON dynamic model
Specification, see Software specification exception, see under Software
SQL gateway, 293 contracting
SQL (Structured Query Language), 293, functionality, 157
295, 296, 297, 298, 305, 306, 309, 311, scenario, see Scenario
312, 319, 320, 321, 323, 324−331 state, see under State
Standardization stimulus, 93, 156, 252
premature, 12 view, see under Clustering
Stanford, 158 System behavior, 90−116, 150, 176, 407
State, 406 System chart, see under BON modeling
change, see Delta charts
of system, 35 System documentation, 177, 180−181
previously saved, see Old abstract vs. incomplete, 216
Statecharts, 17 importance of quality, 177
State transition diagram, 406 newspaper headline style, 199
State variable, 21, 35, 406 of class interfaces, 217
Static binding, 406 style guidelines, see under Naming
Static chart, 407 style of comments, 219−220
Static model, 24−25, 143, 146, 407 System event, see under BON dynamic
separation from dynamic model, 25, 100 model
Static type, 407 System execution, see BON dynamic model
Static typing, see under Typing System usability, 129
Stepwise refinement, 223, 407 System usage, 140−142
Stimulus, see BON dynamic model, system patterns as abstraction, 115
event volatility, 115−116
Stockholm, 207
Stored procedure, 294
String_marks, 359, 365 T
Strong abstraction, 221, 222
Strong typing, 407 TABLE, 53, 76, 99
Strunk, William Jr., 219 Table (relational), 295
Subclass, see Class, descendant Tangible object, see Modeling, physical
Subsystem, 59, 60, 61, 145, 153−154, 157, object
159, 163, 165, 169, 173, 176, 196, 197, TBD (To Be Decided), 261
205, 407 Tektronix, 29
Subtype vs. subclass, 42 Telephone call analogy, see under BON
Such_that, 46, 56, 57, 58, 267 dynamic model
Superclass, see Class, ancestor Textual BON, see under BON notation
Supplier, 10, 407 Textual representation
generically derived, 75 vs. graphical representation, 48−49
relation, see Client relation This, 408
responsibility, 42, 209 Tool class, 187
Supplier vs. client role, 183−185 Toolkit, see under Library design
Syntax notation, see under BON textual Too much too soon, see under Software
438 INDEX

engineering VHS (Video Home System), 278

Top-down design, 408 Video recorder, 270−288
Total function, 408 candidate classes, 274
Trader, 197, 408 class definition, 279−283
Transition, 408 class selection and clustering, 275−279
Transitive closure, 253 external operation, 272−273
Tuple, 295, 408 final static architecture, 288
Type, 20, 408 first static architecture, 277
Type derivation, 74, 409 formal class description, 286
Typed interface, see under Class general menu classes, 278, 282
Type instantiation, see Type derivation internal operation, 270−272
Typing, 409 system behavior, 283−285
as classification, 20 system borderline, 270−274
dynamic, 19, 395 View, 409
importance of, 19−20 Views
information conveyed by, 20 at different levels, see Scalability
static, 19, 171−172, 407 Virtual function, 409
Visibility, 409
Void, 43, 44, 45, 46, 58, 262, 266, 267
U Void reference, 43, 50, 260, 409
Voltaire, 185
Universal quantifier, see For_all
UNIX, 131, 133, 202
Untyped interface, see under BON modeling
W
charts
Use, 409
Weak abstraction, 221, 222
Use case, 115, 140−142, 409
Whiteboard, 17, 29, 48, 62, 71, 127, 145,
User centered design, 129−135
362
a reference model, 130−133
White, E. B., 219
conceptual aspect, 132−133
Whole vs. parts, see Client, aggregation
interface aspect, 133
Windows, 126, 138
pragmatic aspect, 132
Workspace, 126
principles, 154−156
roles of users and computers, 130
User-defined integrity, see under Integrity
X
constraints
User metaphor, 134−135, 138, 155, 157,
Xerox PARC, 134
159, 160, 274, 278, 279, 409
X/Motif, 126, 138

V
Z
VALUE, 260, 261, 264
VCR (Video Cassette Recorder), 270, 272, Zooming capability, 15, 19
273, 278, 279, 282, 284, 286, 288 Zorn, Anders, 208