Introduction to NoSQL

Gabriele Pozzani
April 24, 2013

Outline

NoSQL

Definition Categories Related concepts

Modeling NoSQL stores Properties Performance

Relational vs NoSQL DBMSs

Applications

What does NoSQL mean?

NoSQL stands for:

No Relational No RDBMS NonRel Not Only SQL An umbrella term for all databases and data stores that dont follow the RDBMS principles

NoSQL is

A class of products A collection of several (related) concepts about data storage and manipulation

Often related to large data sets

Where does NoSQL come from?

Non-relational DBMSs are not new But NoSQL represents a new incarnation

Due to massively scalable Internet applications Based on distributed and parallel computing Starts with Google

Development

First research paper published in 2003

Continues also thanks to Lucene's developers/Apache (Hadoop) and Amazon (Dynamo) Then a lot of products and interests came from Facebook, Netfix, Yahoo, eBay, Hulu, IBM, and many more

NoSQL and big data

NoSQL comes from Internet, thus it is often related to the big data concept How much big are big data?

Over few terabytes (>1012 240) Enough to start spanning multiple storage units Efciently storing and accessing large amounts of data is difcult, even more considering fault tolerance and backups Manipulating large data sets involves running immensely parallel processes Managing continuously evolving schema and metadata for semi-structured and un-structured data is difcult

Challenges

Why are RDBMS not suitable for big data?

The context is Internet RDBMSs assume that data are

Dense Largely uniform (structured data) Massive and sparse Semi-structured or unstructured

Data coming from Internet are

With massive sparse data sets, the typical storage mechanisms and access methods get stretched

NoSQL products' categories

NoSQL products can be categorized in

Document databases Key/value stores Sorted ordered column-oriented stores Graph databases XML data stores

Document databases

Documents

Loosely structured sets of key/value pairs in documents, e.g., XML, JSON, BSON Encapsulate and encode data in some standard formats or encodings Are addressed in the database via a unique key Documents are treated as a whole, avoiding splitting a document into its constituent name/value pairs

Allow documents retrieving by keys or contents Notable for:

MongoDB (used in FourSquare, Github, and more) CouchDB (used in Apple, BBC, Canonical, Cern, and more)

Document databases, JSON

{ ApacheLogRecord: { ip: 127.0.0.1, ident : -, http_user : frank, time : 10/Oct/2000:13:55:36 -0700, request_line : { http_method : GET, url : /apache_pb.gif, http_vers : HTTP/1.0, }, http_response_code : 200, http_response_size : 2326, referrer : http://www.example.com/start.html, user_agent : Mozilla/4.08 [en] (Win98; I ;Nav), } }

Key/value stores

Store data in a schema-less way Store data as maps

HashMaps or associative arrays Provide a very efcient average running time algorithm for accessing data Couchbase (Zynga, Vimeo, NAVTEQ, ...) Redis (Craiglist, Instagram, StackOverfow, fickr, ...) Amazon Dynamo (Amazon, Elsevier, IMDb, ...) Apache Cassandra (Facebook, Digg, Reddit, Twitter,...) Voldemort (LinkedIn, eBay, ) Riak (Github, Comcast, Mochi, ...)

Notable for:

Sorted ordered column-oriented stores

Data are stored in a column-oriented way

Data efciently stored Avoids consuming space for storing nulls Unit of data is a set of key/value pairs

Identified by row-key Ordered and sorted based on row-key

Columns are grouped in column-families Data isnt stored as a single table but is stored by columnfamilies Google's Bigtable (used in all Google's services) HBase (Facebook, StumbleUpon, Hulu, Yahoo!, ...)

Notable for:

Column-oriented store example

Dealing with big data and scalability

Traditional DBMSs are best designed to run well on a single machine

Larger volumes of data/operations requires to upgrade the server with faster CPUs or more memory

Vertical scaling

NoSQL solutions are designed to run on clusters of low-specification servers

Larger volumes of data/operations requires to add more machines to the cluster

Horizontal scaling

Horizontal scaling

Horizontal scaling needs to handle distributed data

It's tricky It involves tradeofs between speed, scalability, fault tolerance, atomicity, and consistency

One very used approach is the MapReduce algorithm design pattern

Pioneered by Google and adopted by many others

MapReduce (I)

Originating from the functional programming world It consists of three steps

1. A mapper function/script that goes through input data and

outputs an unordered list of keys and values to use for calculating the final results

Keys are used to cluster data that will be needed to calculate a single output result

2. A sort step that ensures that all data having same key are next

to one another in the file/memory Remove unplanned random accesses 3. A reducer stage that goes through the sorted output and calculates result

It could involve aggregating values

MapReduce (II)

Easily allows the process to be split across a large number of machines Each step runs on each node of a distributed cluster

Independently on the data subset local to the node Isolation provides efective parallel processing Values from reduce operations on diferent nodes are combined to get the final result

It boosts performance even on single machines thanks to the increased locality of memory accesses

MapReduce (III)

Modeling NoSQL stores

NoSQL data modeling often starts from the application-specific queries as opposed to relational modeling:

Relational modeling is typically driven by the structure of available data. The main design theme is What answers do I have? NoSQL data modeling is typically driven by application-specific access patterns, i.e. the types of queries to be supported. The main design theme is What questions do I have?

Data duplication and denormalization are first-class citizens

Conceptual techniques (I)

1. Denormalization

the copying of the same data into multiple documents or tables in order to simplify/optimize query processing or to fit data into a particular model

2. Aggregates 3. Application Side Joins

Conceptual techniques (II)

1. Denormalization 2. Aggregates

Soft schema allows one to form classes of entities with complex internal structures (nested entities) and to vary the structure of particular entities.

3. Application Side Joins

Conceptual techniques (III)

1. Denormalization 2. Aggregates 3. Application Side Joins

Joins are rarely supported in NoSQL solutions Joins are often handled at design time instead of at query execution time When joins are inevitable, they should be handled by an application

Data modeling tecniques

Atomic aggregates Dimensional reduction Index tables Composite key index

Querying NoSQL stores

Diferent NoSQL stores provide diferent querying tools and features

From simple filtering of data based on columns names/values (MongoDB, HBase, Redis, ) To SQL-like languages (Google App Engine, HyperTable, Hive, ...)

NoSQL, No ACID

RDBMSs are based on ACID (Atomicity, Consistency, Isolation, and Durability) properties NoSQL

Does not give importance to ACID properties In some cases completely ignores them

In distributed parallel systems it is difcult/impossible to ensure ACID properties

Even with a central coordinator 2PL, 2PC and SS2PL can help

Long-running transactions don't work because keeping resources blocked for a long time is not practical

CAP Theorem

A congruent and logical way for assessing the problems involved in assuring ACID-like guarantees in distributed systems is provided by the CAP theorem

At most two of the following three can be maximized at one time

Consistency - Each client has the same view of the data Availability - Each client can always read and write Partition tolerance - System works well across distributed physical networks

Conjectured by Eric Brewer in 2000 Proved by Seth Gilbert and Nancy Lynch in 2002

CAP theorem, consistency

It alludes to atomicity and isolation

Gilbert & Lynch use the word "atomic" instead of consistent Concurrent operations see the same valid and consistent data state A service that is consistent operates fully or not at all

It means consistent reads and writes

An atomic data object. Atomic, or linearizable, consistency is the condition expected by most web services today (Gilbert & Lynch)

CAP theorem, availability

It means the system is available to serve at the time when its needed If a system is not available to serve a request at the very moment its needed, its not available "Every request received by a non-failing node in the system must result in a response" (Gilbert & Lynch)

CAP theorem, partition tolerance

A partition happens when, say, a network cable gets chopped, and Node A can no longer communicate with Node B Partition tolerance measures the ability of a system to continue to respond in the event a few of its cluster members become unavailable No set of failures less than total network failure is allowed to cause the system to respond incorrectly (Gilbert & Lynch)

Two out of three

CAP theorem

At most two properties on three can be addressed

The choices could be as follows:

1. Availability is compromised but consistency and partition

tolerance are preferred over it

2. The system has little or no partition tolerance.

Consistency and availability are preferred

3. Consistency is compromised but systems are always

available and can work when parts of it are partitioned

Option 1, compromising on availability

Usual choice for transactional RDBMS under horizontal scaling Availability is afected by many factors, including:

Delays due to network lag Communication bottlenecks Resource starvation Hardware failure leading to partitioning

Option 2, compromising on partition tolerance

Modern interpretation of the theorem is: during a network partition, a distributed system must choose either Consistency or Availability

Because a system that is not Partition-tolerant will, by definition, be forced to give up Consistency or Availability during a partition

In the Internet partitions are rare

Option 3, compromising on consistency

Dropping strong consistency we get weak consistency

We accept eventual consistency Given a sufciently long period of time, over which no updates are sent, one can expect that all updates will, eventually, propagate through the system and all the replicas will be consistent In the presence of continuing updates, an accepted update eventually either reaches a replica or the replica retires from service Basically Available Soft-state Eventually consistent

Eventual consistency could be interpreted in two ways

Eventual consistency takes us to the BASE approach

Performance (I)
based on A vendor-independent comparison of NoSQL databases: Cassandra, HBase, MongoDB, Riak, Sergey Bushik, October 2012.

Tested DBMSs

Cassandra (column-oriented store) HBase (column-oriented store) MongoDB (document-oriented database) Riak (key-value store) MySQL Cluster (RDBMS) sharded MySQL (RDBMS)

Performance (II)

Evaluated under diferent workloads:

Workload A: Update heavily Workload B: Read mostly Workload C: Read only Workload D: Read latest Workload E: Scan short ranges Workload F: Read-modify-write Workload G: Write heavily 100,000,000 records manipulated The total size of a record equal to 1Kb 10 fields of 100 bytes each per record Multithreaded communications with the system (100 threads)

In each workload:

Performance, load

Performance, workload A

Performance, workload B

Performance, workload C

Performance, workload E

MongoDB showed a maximum throughput of 20 ops/sec with a latency of 1sec MySQL Cluster was under 10ops/sec with a latency of 400ms Sharded MySQL showed near 40ops/sec with a latency of up to 400ms Riak does not support range scans

Performance, workload G

Performance, conclusions

There is no perfect NoSQL database Every database has its advantages and disadvantages

Depending on the type of tasks (and preferences) to accomplish

Applications

To DWHs

Pentaho Big Data

To GIS

Application to DWH, Pentaho Big Data

Pentaho developed some tools for big data analysis

Pentaho Big Data Analytics

It provides ability to input, output, manipulate and report on data using Cassandra, Hadoop HDFS, Hadoop MapReduce, HBase, Hive, and MongoDB. Pentaho Instaview provides self-service analytics for the leading big data stores including Hadoop, Cassandra, HBase, MongoDB

Automatic data preparation (ETL, modeling)

Extreme-scale in-memory caching using external data grid technologies such as Infinispan and Memcached

Application to GIS

Some NoSQL solutions include support for geospatial data natively or with an extension Some NoSQL stores currently used to manage geospatial data include:

MongoDB (Foursquare, Where, Scrabbly) BigTable (Google Earth) Cassandra (SimpleGeo) CouchDB/GeoCouch (City of Portland, Oregon's open GIS data initiative)

Storing of POIs and GeoJSON documents

Application to GIS, Google Earth

Google BigTable is a column-based store

sparse, distributed, persistent multi-dimensional sorted map

(row:string, column:string, time:int64)

string

One table to preprocess data (70TB)

Each row in the table corresponds to a single geographic segment

Rows are named to ensure that adjacent segments are stored near

A column family to keep track of the sources of data for each segment

It has a large number of columns: one for each raw data image Each segment is only built from a few images, this column family is very sparse

and a diferent set of tables for indexing and serving client data (500GB)

Conclusions

NoSQL is a set of concepts, ideas, technologies, and software dealing with

Big data Sparse un/semi-structured data High horizontal scalability Massive parallel processing

Diferent applications, goals, targets, approaches need diferent NoSQL solutions

References

Tiwari, Shashank. Professional NoSQL. Wrox, 2011. Warden, Pete. Big Data Glossary. O'Reilly Media, 2011. Vogels, Werner (Amazon.com's CTO). All Things Distributed. Werner Vogels' weblog on building scalable and robust distributed systems. http://www.allthingsdistributed.com/ Katsov, Ilya. NoSQL Data Modeling Techniques.
http://highlyscalable.wordpress.com/2012/03/01/nosql-data-modeling-techniques/

Bushik, Sergey. A vendor-independent comparison of NoSQL databases: Cassandra, HBase, MongoDB, Riak. October 2012. Available online. Gilbert, Seth and Lynch, Nancy. Brewer's conjecture and the feasibility of consistent, available, partition-tolerant web services. ACM SIGACT News 33.2 (2002): 51-59. Redmond, Eric, Wilson, Jim R. , and Carter, Jacquelyn. Seven databases in seven weeks: a guide to modern databases and the NoSQL movement. The Pragmatic Programmers, LLC, 2012.

Sharding

Any DB distributed across multiple machines needs to know in what machine a piece of data is stored or must be stored

A sharding system makes this decision for each row, using its key The main issue is to decide a rule for mapping any key to the corresponding shard

The rule must not introduce too much extra complexity and lookup cost The rule must be fexible to adapt to changes in the cluster and data amount growing

MapReduce (IV)