Lecture 6: Query Processing; Hurry

up!

Overview
EXPLAIN
Measuring Performance
Disk Architectures
Indexes

Join Algorithms (ctd.)

Sort-Merge
External Sorting

Costs and Complexities

Mechanics
Parsing

Motivation, Definition,
Demonstration
Classification

Optimization

Primary vs. Secondary

Unique
Clustered vs UnClustered

Join Algorithms
Nested Loop
Simple
Index
CS3/586

10/10/16

Lecture 6

Slide 1

Learning objectives
LO6.1: Use SQL to declare indexes
LO6.2: Determine the I/O cost of finding record(s) using
a B+ tree
LO6.3: Given a join query, calculate the cost using each
join algorithm: Nested loops, Index Nested Loops,
Sort-Merge
LO6.4: Parse a query
LO6.5: Use VP to answer questions about optimization

Slide 2

Today we will start from the bottom

Web Form

SQL interface

Applic. Front end

SQL

Security

Parser
Relational Algebra(RA)

Catalog

Optimizer
Executable Plan (RA+Algorithms)

Concurrency
Crash
Recovery

Plan Executor
Files, Indexes &
Access Methods
Database, Indexes

Operator
algorithms
indexes
how a
disk works

Slide 3

Measuring Query Speed

Our goal this week is to figure out how to execute a
query fast.
But the time a query takes to execute is hard to
measure or predict.
Depends on environment

Simpler, easier to measure and predict: Number of

disk I/Os.
Good: Very roughly proportional to execution time
Bad: Does not take into account CPU time or type of I/O

Therefore: we will use number of disk I/Os to

measure the time it takes a query to execute.
Like looking under the lamppost.
Slide 4

Components of a Disk *
Spindle

platters are always

spinning (say, 7200rpm).

Disk head

Tracks
Sector

one head reads/writes

at any one time.
to read a record:
position arm (seek)
engage head
wait for data to spin by
read (transfer data)

Arm movement

Platters

Arm assembly
Slide 5

More terminology
Spindle
Disk head

Each track is made up of

fixed size sectors.
Page size is a multiple of
sector size.
A platter typically has data on

Tracks

Arm movement

both surfaces.
All the tracks that you
can reach from one
position of the arm isArm assembly
called a cylinder
(imaginary!).

Sector

Platters

Slide 6

Cost of Accessing Data on Disk

Time to access (read/write) a disk block:
seek time (moving arms to position disk head on track)

rotational delay (waiting for block to rotate under head)

Half a rotation, on average
transfer time (actually moving data to/from disk surface)
Key to lower I/O cost: reduce seek/rotation delays!
(you have to wait for the transfer time, no matter what)
The text measures the cost of a query by the NUMBER of
page I/Os, implying that all I/Os have the same cost, and that
CPU time is free. This is a common simplification.
Real DMBSs (in the optimizer) would consider sequential vs.
random disk reads because sequential reads are much faster
and would count CPU time.
Slide 7

Typical Disk Drive Statistics (2009)*

Sector size: 512 bytes
Seek time
Average
4-10 ms
Track to track
.6-1.0 ms
Average Rotational Delay 3 to 5 ms
(rotational speed 10,000 RPM to 5,400RPM)
Transfer Time - Sustained data rate
0.3- 0.1 msec per 8K
page, or 25-75 Meg/second
Density
12-18GB/in2

Rule of Thumb: 100 I-Os/second/page

Slide 8

How far away is the data?

C lock Ticks

10

Andromdeda
Tape /Optical
Robot

10 6 Disk

100
10
2
1

Memory
On Board Cache
On Chip Cache
Registers

2,000 Years

Pluto

Sacramento

2 Years

1.5 hr

This Campus
10 min
This Room
My Head 1 min

From http://research.microsoft.com/~gray/papers/AlphaSortSigmod.doc

Slide 9

Block, page and record sizes

Block According to text, smallest unit of I/O.
Page often used in place of block.
My notation is:
Page is smallest I/O for operating system
Block is smallest I/O for an application
Block is integral number of units

typical record size: commonly hundreds,

sometimes thousands of bytes
Unlike the toy records in textbooks

typical page size 4K, 8K

Slide 10

What Block Size is Faster?*

At times you can choose a block size for an application. How?
In some OS's, e.g., IBM's, you can enforce a block size
Or you can perform several reads at once, imitating a large block
size. This is called asynchronous readahead.
This is like: should I buy one bottle or a case?

What application will run faster with a large block size?

Goal is for the disk to overlap reads with the CPU's processing of
records. Potentially running twice as fast.

What application will run faster with a small block size?

Goal is not to waste memory or read time.

Slide 11

Time for some Magic

You are in charge of a production DBMS for the FEC.
Production: an enterprise depends on the DBMS for its
existence.

Customers will ask queries like find donations from

97223. You must ensure a reasonable response time.
If the queries run forever, customers will be unhappy
and you will be DM.
The DBMS will grind to a halt. Customers will complain to
congress, you will be out of a job.

Wouldn't it be nice to know what plan the optimizer will

choose, and how long that plan will take to execute?
Rub the magic lantern

Slide 12

Postgres EXPLAIN
Output for
EXPLAIN SELECT * FROM indiv WHERE zip = 97223;
Seq Scan on indiv (cost=0.00.. 109495.94 rows=221 width=166)
Filter:(zip = 97223::bpchar)

Sequential
Scan

I/Os to get
first row

I/Os to get
last row*

Rows
retrieved

Average Row
Width

These values are estimates from sampling.

Most DBMS's provide this facility.
Also useful when a query runs longer than expected.
If you are online, try it.

*Actually this includes CPU costs but we will call it I/O costs to simplify
Slide 13

You are now DM

More than 100K I/Os!
Response time is 1,000 seconds, or 17 minutes.

Unacceptable! Customers will complain!

Is there a faster way than Seq Scan?
You must do something or you are out of a
job!!!

Slide 14

To the Rescue: Index

An Index is a data structure that speeds up access to
records based on some search key field(s).
Indexes are not part of the SQL standard
Because of physical data independence

Typical SQL command to create an index:

CREATE INDEX indexname
ON tablename (searchkeyname[s]);
For example
CREATE INDEX indiv_zip_idx ON indiv(zip);
Nota Bene
Search key is not the same as a key for the table.
Attributes in a search key need not be unique.

Slide 15

Index Demonstration: Input, Output

EXPLAIN SELECT * FROM indiv WHERE zip='97223';
Seq Scan on indiv (cost=0.00..109495.94 rows=221 width=166)
Filter: (zip = '97223'::bpchar)

CREATE INDEX indiv_zip_idx ON indiv(zip);

EXPLAIN SELECT * FROM indiv WHERE zip='97223';
Bitmap Heap Scan on indiv (cost=6.06..861.32 rows=221 width=166)
Recheck Cond: (zip = '97223'::bpchar)
-> Bitmap Index Scan on indiv_zip_idx (cost=0.00..6.01 rows=221
width=0)
Index Cond: (zip = '97223'::bpchar)

With an index, the I/Os went from 109,495 to 861!

Thats 17 minutes to 9 seconds!
Slide 16

LO6.1: Practice with indexes*

When you declare a primary key, most modern DBMSs
(including Postgres) create a clustered (sorted) index on
the primary key attribute (s).
Give the SQL for creating all possible single-attribute
indexes on the table Emp(ssn PRIMARY KEY, name)

What are the search keys of each index?

Slide 17

Data Entries*

Before we learn about how indexes are built, we must understand the
concept of data entries.
Given a search key value, the index produces a data entry, which
produces the data record in one I/O.
Other real-life indexes will help motivate this concept.
Each of the following indexes speeds up data retrieval. What is the
search key, data entry, and data record for each one?
Search Key
Data Entry
Data Record
Library Catalog
Google
Mapquest
Slide 18

Essentially all DBMS Indexes are B+ Trees

Oracle, SQLServer and DB2 support only B+Tree indexes.

Postgres supports hash indexes but does not recommend
using them.
B+ tree indexes support range searches (WHERE const <
attribute) and equality searches (WHERE const = attribute).
The next page contains a sample B+ tree index. Think of it as
an index on the first two digits of zip code.
28* is a data entry that points to the donations from zip codes
that start with 28.
Above the data entries are index entries that help find the
correct data entry.

Slide 19

Example B+ Tree
Root

Note how data entries

in leaf level are sorted

17

Entries <= 17
5

2*

3*

27

13

5*

Entries > 17

7* 8*

14* 16*

22* 24*

30

27* 29*

33* 34* 38* 39*

Find 29*? 28*? All > 15* and < 30*

Insert/delete: Find data entry in leaf, then change it. Need to adjust parent
sometimes.
And change sometimes bubbles up the tree
This keeps the tree balanced: each data retrieval takes the same number of I/Os.
Each page is always at least half full.

Slide 20

LO6.2: I/O Cost in a B+ Tree*

Root

17

2*

3*

27

13

5*

7* 8*

14* 16*

22* 24*

30

27* 29*

33* 34* 38* 39*

How many I/Os are required to retrieve data records with search
key values x, 13 < x < 27? Assume x is a unique key.
How many I/Os are required to retrieve data records with search
key values x, 3 < x < 15? Assume x is a unique key.
Slide 21

B+ Tree Indexes
Non-leaf
Pages

Leaf
Pages
(Sorted by search key)

Leaf pages contain data entries, and are chained (prev

Non-leaf pages have index entries; only used to direct s
index entry

P0

K 1

P1

K 2

P 2

K m Pm

Slide 22

Dont get carried away!*

Now I dont want you to run out and index every
attribute and set of attributes in all your tables!
If you define an index, you will incur three costs
Space to store the index
Updates to the search key will be slower why?
The optimizer will take longer to choose the best plan
because it has more plans to choose from.
We will see that sometimes it is better not to use an index

There is one advantage to having an index

Some queries run faster (better be sure about this).

Slide 23

Index Classification
Primary vs. secondary: If the indexs search key
contains the relations primary key, then the index is
called a primary index, otherwise a secondary index.
The index created by the DBMS for the primary key is usually
called the primary index.

Unique index: Search key contains a candidate

key, i.e. no duplicate values of the search key.

Slide 24

Clustered vs. Unclustered indexes

If the order of the data records is the same as, or `close
to, the order of the search key, then the index is called
clustered.

CLUSTERED

Index entries
direct search for
data entries

Data entries

UNCLUSTERED

Data entries
(Index File)
(Data file)

Data Records

Data Records

Slide 25

Comments on Clustered Indexes

If you are retrieving only one record, any index will do.
Retrieve one record in each index and count the I/Os.
Assume the height of the index entry tree is 2.

If you are retrieving many records with the same search

key value, a clustered index is almost always faster.
Retrieve 10 records from each index and count the I/Os.
Clustered:
Unclustered:

Lest you get carried away: a table can have only one
clustered index. Why?
DBMSs make their primary indexes clustered.

PS: DB2, Postgres and MySQL construct clustered indexes as we have

described on the previous slide. Oracle and SQLServer put the data records in
place of the data entries.
Slide 26

Where Are We?

We've now learned two ways to perform a 1-table
SELECT query: Sequential Scan and Index Scan.
EXPLAIN tells you which plan/algorithm the optimizer
will choose; which one it thinks is the fastest.
Now we study possible plans/algorithms for multitable join SELECT queries.

Slide 27

Join Algorithms: Motivation

(apocryphal)

When I was young I was asked to help with a charity

art auction. At the start I got a big stack of bidder
cards with bidder IDs and bidder information.
At the end I got a much bigger stack of bought cards,
each one containing a bidder ID and the cost of a
painting that a bidder bought.
Suddenly there was a long line of bidders who
wanted to go home. For each bidder, I had to give
the cashier the bidders card with the bidders
matching bought cards.
What would you do if you were in this situation?
Slide 28

Computer Science Algorithms

Answers to the previous question will be investigated
on the following pages. They fall into three
categories, the three basic algorithms of computer
science: iteration, sorting and hashing.
Nested Loop Join (iteration) comes in two versions:
Simple Nested Loop
Index Nested Loop

Sort Merge Join

Hash Join (Will not be covered in this course)

Slide 29

Join Algorithms an Introduction

The text discusses algorithms for every relational
operator. We study only join algorithms since join is
so expensive.
L R is very common!
Notation: M pages in L, pL rows per page, N pages in
R, pR rows per page.
In our examples, L is indiv and R is comm.
Our algorithms work for any equijoins.

Slide 30

A simple join
SELECT *
FROM indiv L, comm R
WHERE L.commid=R.commid
Review how to compute this join by hand, with the cl versions
of the tables.
M = 23,224 pages in L, pL = 39 rows per page,
N = 414 pages in R, pR = 24 rows per page.
These (estimated) statistics are stored in the system catalog.
In PostgreSQL, retrieve number of pages with the function
SELECT pg_relation_size('tablename')/8192;
Retrieve rows per page using
SELECT COUNT(*)/(pages in L or R) FROM L or R;
Slide 31

The simplest algorithm: Nested Loops

Join on commid in L and commid in R
foreach row l in L do
foreach row r in R do
if rcommid == lcommid then add <r, s> to result

For each row in the outer table L, we scan the entire inner
table R, row by row.
Cost: M + (pL * M) * N = 23,224 + (39*23,224)*414 I/Os
= 374,997,928 I/Os 3,749,979 seconds 43 days

Assuming approximately 100 I/Os per second

(86,400 secs/day)
Slide 32

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:

Table R
on disk
... 2
13
12
27
1
5

Slide 33

Nested Loops Join

Table L
on disk
2 ...
12
6 ...

Memory Buffers:
2 ...
12
6 ...

... 2
13

Table R
on disk
... 2
13
12
27

1
5
27

1
5

Query Answer
2 2
Slide 34

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

... 2
13

No match:
Discard!

Table R
on disk
... 2
13
12
27
1
5

Query Answer
2 2
Slide 35

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

12
27

No match:
Discard!

Table R
on disk
... 2
13
12
27
1
5

Query Answer
2 2
Slide 36

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

12
27

No match:
Discard!

Table R
on disk
... 2
13
12
27
1
5

Query Answer
2 2
Slide 37

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

1
5

No match:
Discard!

Table R
on disk
... 2
13
12
27
1
5

Query Answer
2 2
Slide 38

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

1
5

No match:
Discard!

Table R
on disk
... 2
13
12
27
1
5

Query Answer
2 2
Slide 39

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

... 2
13

No match:
Discard!

Table R
on disk
... 2
13
12
27
1
5

Query Answer
2 2
Slide 40

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

... 2
13

No match:
Discard!

Table R
on disk
... 2
13
12
27
1
5

Query Answer
2 2
Slide 41

Nested Loops Join

Table L
on disk
2 ...
12
6 ...
1
5
27

Memory Buffers:
2 ...
12
6 ...

12
27

Match!

Query Answer
2 2
12 12

Table R
on disk
... 2
13
12
27
1
5

And so forth

Slide 42

Index Nested Loops Join

IF THERE IS AN INDEX ON r.commid
foreach row l in L do
use the index to find all rows r in R where lcommid =
rcommid
for all such r: add <l, r> to result

Cost: M + ( (M*pL) * cost of finding matching R rows)

= 23224 + ((23224*39)*3) = 2,740,432 I/Os 27,404 secs 8 hours

Cost of finding the rows in R using

the index on commid much
cheaper than scanning all of comm!

Slide 43

External Sorting

Many relational operator algorithms require sorting a table

Often the table wont fit in memory
How do we sort a dataset that wont fit in memory?
Answer: External Sort-Merge algorithm
First pass: Read and write a memoryfull of (sorted) runs at a time.
Second and later passes: Merge runs to make longer runs
Heres a picture of merging two runs:
78 72 68 55 54 54 40

36

92 88 66 51 43

29

Runs on disk

Merging the runs

in memory

23 21 20 18 9 7

The merged output is

a longer run, on disk

Slide 44

External Sorting Cost

Number of passes depends on how many
pages of memory are devoted to sorting
Can sort M pages of data using B pages of
memory in 2 passes if sqrt(M) <= B

Can sort big files M with not much memory B

If page size is 4K:
Can sort 4Gig of data in 4Meg of memory
Can sort 256Gig of data in 32Meg of memory

Each pass is a read and a write, so if sqrt(M) <= B

then sort costs (M+M)+(M+M) so can be done in
4*M I/Os
So its reasonable to assume that sorting M pages
costs 4*M.
Slide 45

Sort-Merge Join

This join algorithm is the one many people think of when asked
how they would join two tables. It is also the simplest to
visualize. It involves three steps.
1. Sort L on lcommid
2. Sort R on rcommid
3. Merge the sorted L and R on lcommid and rcommid.

Weve covered the algorithm and cost of steps 1 and 2 on the

previous pages

Slide 46

The Merge Step

What is the algorithm for step 3, the merge?
Advance scan of L until current L-rows lcommid >= current R rows rcommid, then
advance scan of R until current R-rows rcommid >= current R rows lcommid ; do
this until current R rows lcommid = current R rows rcommid.
At this point, all R rows with same lcommid and all R rows with same rcommid
match; output <l, r> for all pairs of such rows.
Then resume scanning L and R.

What is the cost of the merge step?

Normally, M+N

What if there are many duplicate values of lcommid and rcommid?

What if all values of lcommid are the same and equal to all values of
rcommid?
Then L R = L R and the cost of the merge step is L * R.

BUT, almost every real life join is a foreign key join. One of
the joining attributes is a key, so the duplicate value
problem does not occur.
Slide 47

Cost of Sort-Merge Join

Assuming that sorting can be done in two
passes and that the join is a foreign key join
Cost: (cost to sort L) + (cost to sort R) + (cost
of merge)
= 4M + 4N + (M+N) = 5(M+N)
For our running example the cost is:
5*(M+N) = 5*(23224+414) = 118,190 I/Os
1,181 seconds 20 minutes
In reality the cost is much less because of
optimizations, indexes, and the use of hash
join
Cf. CS587/410
Slide 48

Costs for Join Algorithms

Join Algorithm
Nested Loop
Index Nested Loop
Sort-merge, with 2-pass
sort for both inputs

I/O Cost

O( )

M + PL*M*N M*N

Time for our

example
43 Days

M + PL*M*(cost
of index access*)

8 Hours

5(M+N)

M+
N

20 minutes

*For homework and exercises you may assume this is 3 times

the number of rows retrieved
Slide 49

LO6.3: Costs of Join Algorithms*

Consider this join query:
SELECT *
FROM pas L, comm R
WHERE L.commid = R.commid;
Calculate the cost (in time) of a nested loop, index
nested loop and sort-merge join.

Slide 50

Now we focus on the top of this

diagram
SQL Query

Parser

Search for a cheap plan

Relation Algebra Query

Query Optimizer

Join algorithms,

Relational Operator Algs.

Heap, Index,

Files and Access Methods

Covered in
CS587/410

Buffer Management
Disk Space Management
DB
Slide 51

Detail of the top

SQL Query
(SELECT )

Query Parser
Relational Algebra Expression (Query Tree)
Query Optimizer
Plan
Generator

Plan Cost
Estimator

Catalog
Manager

Query Tree + Algorithms (Plan)

Plan Evaluator
Slide 52

Parsing and Optimization

The Parser

Verifies that the SQL query is syntactically correct, that

the tables and attributes exist, and that the user has the
appropriate permissions.
Translates the SQL query into a simple query tree
(operators: relational algebra plus a few other ones)

The Optimizer:

Generates other, equivalent query trees

(Actually builds these trees bottom up)
For each query tree generated:

Selects algorithms for each operator (producing a query plan)

estimates the cost of the plan

Chooses the plan with lowest cost (of the plans considered,
which is not necessarily all possible plans)
Slide 53

Heres what the parser does

SQL Query:
SELECT
FROM
USING
WHERE

Relational Algebra Tree:

commname
comm JOIN indiv
commid
indiv.zip=97223;

commname
indiv.zip=97223

commid=commid

comm

indiv
Slide 54

LO6.4: Parse a Query*

Describe the parser's output when the input is
SELECT candname
FROM cand JOIN pas
USING candid
WHERE amount > 3000;

Slide 55

What does the optimizer do?

Fortunately, a Master's student at PSU, Tom Raney,
has just added a patch to PostgreSQL (PG) that
allows anyone to look inside the optimizer (PG calls it
the planner).
One of the lead PG developers says its like finding
Sasquatch.
Well use Toms patch to see what the PG planner
does.
The theory behind the PG planner [668] is shared by
all DBMS optimizers*.
*Except SQL Server, though I won't keep saying this.

Slide 56

Overview of DBMS Optimizers

"Optimizing a query" consists of these 4 tasks
1. Generate all trees equivalent to the parser-generated
tree
2. Assign algorithms to each node of each tree

A tree with algorithms is called a plan.

3. Calculate the cost of each generated plan

Using the join cost formulas we learned in previous slides*

4. Choose the cheapest plan

*Statistics for calculating these costs are kept in the system catalog.

Slide 57

Dynamic Programming

A no-brainer approach to these 4 tasks could take

forever. For medium-large queries there are millions
of plans and it can take a millisecond to compute
each plan cost, resulting in hours to optimize a query.
This problem was solved in 1979 [668] by Patsy
Selinger's IBM team using Dynamic Programming.
The trick is to solve the problem bottom-up:
First optimize all one-table subqueries
Then use those optimal plans to optimize all two-table
subqueries
Use those results to optimize all three-table subqueries, etc.
Slide 58

Consider A Query and its Parsed

Form
SELECT commname
FROM indiv JOIN comm USING (commid)
WHERE indiv.zip = '96828';

commname
indiv.zip=96828

I chose 96828 because it

is in Hawaii. Wishful
thinking.

commid=commid

comm

indiv
Slide 59

What Will a Selinger-type Optimizer

Do?
1. Optimize one table subqueries

indiv WHERE zip=96828 , then comm

2. Optimize two-table queries

The entire query

Let's use Raney's patch, the Visual Planner, to see

what PG's Planner does.
We'll watch PG's Planner in two cases
noindex.pln: no index on indiv.zip
index.pln: a nonclustered index on indiv.zip

Slide 60

How to Set Up Your Visual Planner

Download, then unzip, in Windows or *NIX:

Read README.TXT, don't worry about details

Be sure your machine has a Java VM

http://www.java.com/en/download/index.jsp

Click on Visual_Planner.jar

cs.pdx.edu/~len/386/VP1.7.zip

If that does not work, use this at the command line:

java -jar Visual_Planner.jar

In the resulting window

File/Open
Navigate to the directory where you put VP1.7

Navigating to C: may take a while

Choose noindex.pln
Slide 61

Windows in the Visual Planner *

The SQL window holds the (canned) query
The Plan Tree window holds the optimal plan for the query.
The Statistics window holds statistics about the highlighted
node of the Plan Tree's plan
Click a Plan Tree node to see its statistics
Why is the Seq Scan on the right input, indiv, almost the same cost
as the Sort?
Why is there an index scan on the joining attribute of comm?

Why is a merge join the optimal plan?

Almost no cost to sort the right input
No cost to sort the left input because the index is clustered
Slide 62

Visualize Dynamic Programming*

Recall the first steps of Dynamic Programming:
Optimize indiv, then comm.
Postgres calls these the ROI steps and they are
displayed in the ROI window of VP.
In the ROI window, click on indiv to see how the PG
Planner optimized indiv. What happened?
In the ROI window, click on comm. What happened?
The Planner saved the index scan even though it was slower
than the Seq Scan, because it had an interesting order.
The index scan is ordered on commid, which is a joining
attribute, so it is an interesting order.
Slide 63

The Last Act

The last step of Dynamic Programming is to optimize
the entire query, the two-table join.
Click on indiv/comm in the ROI Window.
Blue plans are those that have the fastest total cost or the
fastest startup cost, either overall or for some interesting
order.
Red plans are dominated by another plan.
Dominated means there is a faster plan with the same order.

To see a plan in a separate window, Shift-click it.

Plans are listed in alphabetical order, then in order of total
cost, then in order of startup cost.

Slide 64

What Happened in the Last Act?*

The first blue plan is the optimal plan we've been
looking at.
Why is the second blue plan there?
Look at the other Merge Join plans. Why are they
red?
Find and describe the most expensive plan. What
makes it so expensive?

Slide 65

Index to the Rescue*

File/Open, navigate to index.pln

Without the index the optimal plan cost 35,471
What is the cost of the optimal plan now?
Why?

Slide 66

LO6.2 EXERCISE*
Consider the B+-tree index on slide 21. Assume none
of the tree is in memory and the index is unique.
Assume that in the data file, every data record is on a
different page. How many disk I/Os are needed to
retrieve all records with search key values x, 7 < x <
16?

Slide 67

LO6.3: EXERCISE
Consider the join query:
SELECT *
FROM comm L, cand R JOIN ON (assoccand = candid )
Calculate the cost of a nested loop, index nested loop
and sort-merge join.

Slide 68

LO6.4: EXERCISE
Follow the instructions on slide 61 to set up the Visual
Planner. Open the file noindex.pln
What is the startup cost and the total cost of the left input?

Open the file index.pln

Click on the "Bitmap Index Scan". What index is being used?
What is the order of the left input?

Slide 69