2. Disclaimer
Any views or opinions presented in this presentation are
solely those of the author and do not necessarily represent
those of Verisign.
2
3. Outline
• Introduction
• Scalability and Availability
• Difficulties to scale
• CAP Theorem
• NoSQL Goals
• NoSQL Taxonomy
• Concepts and Patterns
• Existing implementations
• NoSQL in the Real World
• Conclusion
3
5. NoSQL Term
• Mandatory name disambiguation :
NoSQL stands for Not Only SQL.
• The term NoSQL is more or less attributed to Eric Evans,
a Rackspace employee, who used it in early 2009 when
Johan Oskarsson, a Last.fm employee, wanted to
organize an event to discuss open-source distributed
databases.
5
6. Wikipedia Definition
• [Wikipedia] NoSQL is a term used to designate database
management systems that differ from classic relational
database management systems (RDBMS) in some way.
These data stores may not require fixed table schemas,
usually avoid join operations, do not attempt to provide
ACID (atomicity, consistency, isolation, durability)
properties and typically scale horizontally.
6
8. Why NoSQL ?
• Is there really a problem with SQL and RDBMS ?
No there isn't !
SQL is powerful, ACID (atomicity, consistency, isolation,
durability) properties are well-established, developers and DBAs
have dominated it.
• So why the hell need we NoSQL ?
• What was the motivations of Google and Amazon to
invest huge amount in research around NoSQL ?
• Why “social sites” are so hard to scale ?
8
10. Scalability definition
• [Wikipedia] Scalability is a desirable property of a
system, a network, or a process, which indicates its
ability to either handle growing amounts of work in a
graceful manner or to be readily enlarged.
In summary : handle load and peaks.
• Scalability in two dimensions :
• Scale up → scale vertically (increase RAM in an existing node)
• Scale out → scale horizontally (add a node to the cluster)
10
11. Availability definition
• [Wikipedia] Availability refers to the ability of the users to access
and use the system. If a user cannot access the system, it is said
to be unavailable. Generally, the term downtime is used to refer to
periods when a system is unavailable.
In summary : minimize downtime.
Availability % Downtime per year Downtime per month Downtime per week
90% ("one nine") 36.5 days 72 hours 16.8 hours
95% 18.25 days 36 hours 8.4 hours
99% ("two nines") 3.65 days 7.20 hours 1.68 hours
99.9% ("three nines") 8.76 hours 43.2 minutes 10.1 minutes
99.99% ("four nines") 52.56 minutes 4.32 minutes 1.01 minutes
99.999% ("five nines") 5.26 minutes 25.9 seconds 6.05 seconds
99.9999% ("six nines") 31.5 seconds 2.59 seconds 0.605 seconds
11
13. RDBMS Scalability
• RDBMS are hard to scale. Hard should be understood
as costly
RDBMS licenses, hardware, DBAs' and operational costs grow
non linearly with the load.
RDBMS have either :
• single point of failure (SpoF)
→ the master DB
• or replication latency
→ distributed transactions : two-phase commit, paxos algorithm.
13
14. Hardware Scalability
• Commodity hardware and appliances are I/O bound
• But network throughput is cheaper than hard disk throughput.
• Hard disks is (was?) the main bottleneck in today's
computing.
• Random access have a high latency (disk seek)
• Throughput have increased, but not proportionally with the
storage
→ Distributing data across a network of small computers
(and applying the data locality concept) scale better
(cheaper) than a huge appliance.
14
15. Pioneers in Scaling The Big
• Google : huge amount of data (hundreds of petabytes, do not
fit on a single appliance)
• need to be partitioned MTBF is proportional to number of machines.
→ BigTable.
• Amazon : high availability (99.999%)
• need to have redundancy and write scalability
→ Dynamo
• Facebook, Twitter, "social sites”
• no cluster of data to be partitioned,
• long tail (old data still time to time accessed),
• lots of users connected in the same time,
• user specific content (predictions hard to achieve, few static pages)
15
17. Remind of CAP theorem
• Consistency : all nodes see the same data at the same
time
• Availability : node failures do not prevent survivors from
continuing to operate
• Partition Tolerance : the system continues to operate
despite arbitrary message loss
According to the theorem, a distributed system can
satisfy any two of these guarantees at the same time,
but not all three.
17
18. NoSQL trade-offs
• NoSQL datastores have typically done other trade-offs
than RDBMS to the CAP theorem
• Most of them gave up the “C” of the theorem, giving up the ACID
properties in the same way.
• NoSQL datastores also have a more simpler data
access pattern
• value = get(key)
• put(key, value)
• remove(key)
18
20. NoSQL promises
• NoSQL finality is to achieve (horizontal) scalability and
high availability.
• Business goal : Keep cost growing proportionally with the load
(tight provisioning).
• Operational goal :
• Scale the system by simply adding node (or removing).
• The system runs on commodity hardware.
20
22. Taxonomy
NoSQL most common types are :
• Document store
• store document which structure can be explored.
• Key/value store
• simple hash map data access pattern.
• Column oriented store
• Something between simple key/value and complex document store
• Graph database
• store node and edges, walk-through data access
• Object database
22
23. Classification
23 Picture from http://blog.nahurst.com/visual-guide-to-nosql-systems
25. Implementation side : Consistent Hashing
• [Wikipedia] Consistent hashing is a scheme that
provides hash table functionality in a way that the
addition or removal of one slot does not significantly
change the mapping of keys to slots.
25 Picture from http://www.lexemetech.com/2007/11/consistent-hashing.html
26. Implementation side : Bloom Filter
• [Wikipedia] Bloom filter is a space-efficient probabilistic
data structure that is used to test whether an element is
a member of a set.
26 Picture from http://en.wikipedia.org/wiki/File:Bloom_filter.svg
27. Implementation side : Quorum
• [Wikipedia] A quorum is the minimum number of votes
that a distributed transaction has to obtain in order to be
allowed to perform an operation in a distributed system. A
quorum-based technique is implemented to enforce
consistent operation in a distributed system.
• Quorum : R + W > N
N : number of replica, R : number of node read, W : number of
node written.
• R = 1, W = N
• R = N, W = 1
• R = N/2, W = N/2 (+1 if N is even)
27
28. Implementation side : Vector Clocks
• [Wikipedia] Vector Clocks is an algorithm for generating
a partial ordering of events in a distributed system and
detecting causality violations.
28 Picture from http://en.wikipedia.org/wiki/File:Vector_Clock.svg
29. Implementation side : other Common Concepts
• Replication
• Multi-master (Gossip, agents, P2P)
• Master-slave (+failover)
• Merkle Tree
• The main use of hash trees is to make sure that data blocks received from
other peers in a peer-to-peer network are received undamaged and unaltered
• Multiversion concurrency control
• MVCC is a concurrency control method commonly used by database
management systems to provide concurrent access to the database and in
programming languages to implement transactional memory.
• SEDA
• Staged Event-Driven Architecture
• LMT (Log Merge Tree)
• Efficient replacement of B-Tree that require less disk seeks.
29
30. User side : Map Reduce
• [Wikipedia] MapReduce is a framework for processing
huge datasets on certain kinds of distributable problems
using a large number of computers (nodes). It is inspired
by the map and reduce functions commonly used in
functional programming.
30 Picture from http://code.google.com/edu/parallel/mapreduce-tutorial.html
31. User side : Inverted Indexes
• [Wikipedia] An inverted index is an index data structure
storing a mapping from content, such as words or
numbers, to its locations in a database file, or in a
document or a set of documents. The purpose of an
inverted index is to allow fast full text searches, at a cost
of increased processing when a document is added to
the database (data denormalization).
31 Picture from http://developer.apple.com/library/mac/
32. User side : other Patterns
• Idempotent updates
• Repeating the update twice do not lead to inconsistent data.
• Offline (asynchronous) processing
• Push on change (vs. pull on demand), batches
• SOA and services isolation
• Service failure resilience. Product is a mash-up of back-end
services.
• Schemaless (ColumnFamily-based data model)
• Schema update is no more required
• Data locality
• Processing is sent to data instead of data sent do workers.
32
36. Real World Example
• Facebook new messaging system
• Based on Hbase (consistency model simplier than Cassandra)
• Amazon shopping cart
• Based on Dynamo, never delete a row. Only add delta (+1, +2,
-1) in the cart.
• Twitter
• Use Cassandra for real time analytics
• Google Megastore
• ACID within partitions, lower consistency across partitions
• Synchronous replication with Paxos algorithm
36
37. Real World Example (2)
• Digg, Last.fm, NYT, Guardian, Yahoo, Flickr, NetFlix,
Adobe, Mozilla, Github, Linkedin, StumbleUpon, ...
• MapReduce usage is increasing daily
• Google and Amazon have sat their domination in being the first
to masterize big data.
• Can your business jump into the NoSQL wave ?
37
39. Conclusion
• NoSQL is not a general purpose datastore !
• Eventually consistent model can be tricky, and can reserve
nasty surprises if not used carefully.
• MapReduce search job have high latency
• SQL and NoSQL are complementary
• Knowing both allow to put the right technology to the right place.
• NoSQL has challenged RDBMS supremacy
• New ideas for RDBMS are emerging
• See HandlerSocket, a mysql NoSQL plugin.
39
40. Thanks for listening
Questions ?
Benoit Perroud
bperroud@verisign.com
@killerwhile on Twitter
40