Speaker(s): Matt Kennedy, Architect at DataStax Apache Cassandra is singular in its ability to provide a single logical database dispersed over multi-region datacenters. Cassandra databases can lose entire datacenters due to network outages or natural disasters and the database will remain available for application users. This talk will discuss how Cassandra is able to achieve this feat, and dive deep into the programming realities of distributed systems and consistency models. It turns out that the consistency model is the critical feature of successful distributed systems. When an application and a database can have an honest conversation about the consistency needs of each, a powerful system is born. This honesty is at the core of Cassandra's dominant position in the world of distributed databases.

1. Cassandra: Consistency & Tolerance
(a morality guide for databases)
Matt Kennedy, Architect
@thetweetofmatt

2. ©2014 DataStax Confidential. Do not distribute without consent. 2

3. ©2014 DataStax Confidential. Do not distribute without consent. 3

4. ©2014 DataStax Confidential. Do not distribute without consent. 4

5. ©2014 DataStax Confidential. Do not distribute without consent. 5

6. ©2014 DataStax Confidential. Do not distribute without consent. 6

7. ©2014 DataStax Confidential. Do not distribute without consent. 7

8. ©2014 DataStax Confidential. Do not distribute without consent. 8

9. ©2014 DataStax Confidential. Do not distribute without consent. 9

10. ©2014 DataStax Confidential. Do not distribute without consent. 10

11. ©2014 DataStax Confidential. Do not distribute without consent. 11

12. ©2014 DataStax Confidential. Do not distribute without consent. 12
?

13. ©2014 DataStax Confidential. Do not distribute without consent. 13
Georgetown
American
20 minutes
Worth the wait?

14. ©2014 DataStax Confidential. Do not distribute without consent. 14
Georgetown
American

15. ©2014 DataStax Confidential. Do not distribute without consent. 15

16. Lessons?
• High Availability means:
• Have to maintain some redundancy
• Need to distribute your stuff
• Distributed Stuff means:
• Need to accommodate unreliable networks
• Need to have strategies for different path costs
• Plan for outright failures
• Cooperation!
Also, don’t underestimate a squirrel with a Jesuit education.
©2014 DataStax Confidential. Do not distribute without consent. 16

17. ©2014 DataStax Confidential. Do not distribute without consent. 17

18. ©2014 DataStax Confidential. Do not distribute without consent. 18
Eventual Consistency

19. Meh.
©2014 DataStax Confidential. Do not distribute without consent. 19
http://nekrocodile-izumi.deviantart.com/art/Lazy-college-student-322186707

20. ©2014 DataStax Confidential. Do not distribute without consent. 20

21. ©2014 DataStax Confidential. Do not distribute without consent. 21

22. ©2014 DataStax Confidential. Do not distribute without consent. 22

23. Consistency vs. Consistency
©2014 DataStax Confidential. Do not distribute without consent. 23

24. ©2014 DataStax Confidential. Do not distribute without consent. 24

25. Tunable Consistency & High Availability
©2014 DataStax Confidential. Do not distribute without consent. 25
Example assumes replication factor = 3
write
read

26. Quorum-write, RF=5
©2014 DataStax Confidential. Do not distribute without consent. 26
Client
6-node cluster, rf=5
OK!

27. Quorum read RF=5
©2014 DataStax Confidential. Do not distribute without consent. 27
Client
6-node cluster, rf=5
00101010
00101010
0010101000101010
? ?
Read value: 00101010

28. ©2014 DataStax Confidential. Do not distribute without consent. 28

29. Linearizable Consistency
• AKA:
• Compare & Set
• Check & Set
• Lightweight Transactions
• CQL “IF” Statements
• Uses Paxos
• Can specify local datacenter or all datacenters
• (CASSANDRA-5797)
©2014 DataStax Confidential. Do not distribute without consent. 29

30. Linearizable Consistency
• Put X in the database if X is not already in the
database
• Set foo=bar if foo=baz
• Set foo=bar if boo=faz
©2014 DataStax Confidential. Do not distribute without consent. 30

31. ©2014 DataStax Confidential. Do not distribute without consent. 31
MONSTER MASH!!!
Alice Bob
Create “Stinky”
C*
Create “Stinky”
Insert “Stinky”
Insert “Stinky”Quorum insert
Quorum insert
Now Bob owns “Stinky” and Alice doesn’t know
why she can’t see the monster she just made

32. ©2014 DataStax Confidential. Do not distribute without consent. 32
MONSTER MASH!!!
Alice Bob
Create “Stinky”
C*
Create “Stinky”
Insert “Stinky”
Insert “Stinky”
LWT Insert
LWT insert
NO!

33. Consistency & Cassandra
• Tunable Consistency
• Linearizable Consistency
• Your App can tell C* what it needs
©2014 DataStax Confidential. Do not distribute without consent. 33

34. THIS IS IMPORTANT
• The Consistency Mechanisms are King
• Cassandra’s core strengths…
• High Availability
• Tolerance of node failures
• Tolerance of datacenter failures
• Tolerance of network failures between datacenters
• All due to the Consistency Mechanisms
• Cassandra isn’t Cassandra without Eventual
Consistency
©2014 DataStax Confidential. Do not distribute without consent. 34

35. Consistency is a Key Value!
• Get it? Key-Value?
• It’s a NoSQL joke.
• Always-on Big Data demands multi-region DCs
• Multi-region DCs must face the reality of network partitions
• Eventual Consistency enables applications built on
Cassandra to do what they need to do correctly and flexibly
while remaining constantly available.
©2014 DataStax Confidential. Do not distribute without consent. 35

36. Thank You
©2014 DataStax Confidential. Do not distribute without consent.

37. Tunable Consistency & High Availability
©2014 DataStax Confidential. Do not distribute without consent. 37
write
read
Example assumes replication factor = 3
1.
2.
write
read
3.
Options 1 & 2 achieve strong
consistency at the expense of
availability. Option 3 requires
only 2 replicas on write and 2
on read, so it achieves high
availability & strong
consistency.

38. Eventual Consistency The Netflix Experiment
• “Eventual Consistency != Hopeful Consistency”
• Netflix, big C* user, performed an experiment
©2014 DataStax Confidential. Do not distribute without consent. 38
http://planetcassandra.org/blog/post/a-netflix-experiment-eventual-consistency-hopeful-consistency-by-christos-kalantzis/

39. ©2014 DataStax Confidential. Do not distribute without consent. 39

40. C* Datacenters A Single Cluster
©2014 DataStax Confidential. Do not distribute without consent. 40
One
Datacenter
Peer-to-peer ring of nodes

41. Datacenters can go offline…
©2014 DataStax Confidential. Do not distribute without consent. 41

42. For a variety of reasons
©2014 DataStax Confidential. Do not distribute without consent. 42

43. C* Datacenters A Single Cluster
©2014 DataStax Confidential. Do not distribute without consent. 43
One
Datacenter
Peer-to-peer ring of nodes

44. C* Datacenters What happens when I lose one?
Los
Angeles
©2014 DataStax Confidential. Do not distribute without consent. 44
Chicago
New
York

45. C* Datacenters What is possible?
Los
Angeles
©2014 DataStax Confidential. Do not distribute without consent. 45
Chicago
New
York
Cloud

46. C* Datacenters Cloud Deployments
©2014 DataStax Confidential. Do not distribute without consent. 46
Availability
Zone 1
AWS

47. C* Datacenters Cloud Deployments
©2014 DataStax Confidential. Do not distribute without consent. 47
Availability
Zone 1
Availability
Zone 2
AWS

48. C* Datacenters Cloud Deployments
©2014 DataStax Confidential. Do not distribute without consent. 48
Availability
Zone 1
Availability
Zone 2
Google
Compute
Engine
AWS

49. C* Datacenters Cloud Deployments
©2014 DataStax Confidential. Do not distribute without consent. 49
Off-Cloud
Migration
On-Prem
Start in
Cloud