CAP Revisited

Diving into the theory behind the CAP theorem and discussing implications for database systems.

Transcript

1. CAP revisited Philipp Krenn @xeraa
3. Consistency "[...] a total order on all operations such that each operation looks as if it were completed at a single instant."
4. Availability "[...] every request received by a non- failing node in the system must result in a response."
5. Partition Tolerance "In order to model partition tolerance, the network will be allowed to lose arbitrarily many messages sent from one node to another."
6. Proof by Construction If a client writes to one side of a partition, any reads that go to the other side of that partition cannot know about this write
8. Conclusion "most real-world systems today are forced to settle with returning “most of the data, most of the time.”"
10. Fallacy #1: The Network Is Reliable https://aphyr.com/posts/288-the-network-is-reliable Distributed systems can either go for C or A
11. Fallacy #2: Guarantees must hold under all (critical) conditions
12. Fallacy #3: Beating the CAP Theorem Checklist http://ferd.ca/beating-the-cap-theorem-checklist.html
13. 4 you pushed the actual problem to another layer of the system 4 you are not, in fact, designing a distributed system 4 latency is a thing that exists 4 using "infinite timeouts" is not an acceptable solution to lost messages 4 read-only mode is still unavailability for writes 4 you shouldn't be in charge of people's data
15. CAP vs ACID Terms: Distributed systems vs databases
16. Consistency Predicate: Integrity constraint History: Total order / single copy Increment / decrement counter consistency: ACID !, CAP ❓
17. Isolation Often relaxed in RDBMS Serializable, repeatable read, read (un-) committed Consistency model in CAP
18. Atomicity RDBMS "all-or-nothing" not isolated Consistency model in CAP
19. Durability Not covered in the CAP paper
20. Repeating for the 4th time today, fsync is not required for strong consistency, this post saying so is wrong http://antirez.com/news/67 1 https://twitter.com/kellabyte/status/410224523602960385
22. "the impossibility of guaranteeing both safety and liveness in an unreliable distributed system"
23. "some systems may sacrifice both consistency and availability! In doing so they may achieve a trade-off better suited for the application at hand."
25. "CAP prohibits only a tiny part of the design space: perfect availability and consistency in the presence of partitions, which are rare."
26. "retrying communication indefinitely is in essence choosing C over A."
27. "Most systems cannot always merge conflicts [after a partition recovery]."
28. CAP in practice https://aphyr.com/tags/jepsen MongoDB, ElasticSearch, etcd & Consul, RabbitMQ, Redis, Cassandra, Kafka, NuoDB, Zookeeper, Riak, PostgreSQL
29. Pick Two — At Most https://aphyr.com/posts/322-call-me-maybe-mongodb-stale-reads MongoDB replication: CP
30. Play It Safe Write concern majority Read from primary
32. Dirty Read Isolation: Read uncommitted Readable from (minority) primary while async replication is running — network partition with rollback after a timeout
33. Stale Read Reading old data from a (minority) primary until a timeout happens
34. /dev/null breaks CAP: effect of write are always consistent, it's always available, and all replicas are consistent even during partitions. 1 https://twitter.com/ashic/status/591511683987701760
35. More 4 http://webpages.cs.luc.edu/~pld/353/ gilbert_lynch_brewer_proof.pdf 4 http://groups.csail.mit.edu/tds/papers/Gilbert/Brewer2.pdf 4 http://www.infoq.com/articles/cap-twelve-years-later- how-the-rules-have-changed 4 http://blog.thislongrun.com/2015/03/the-cap-theorem- series.html 4 http://henryr.github.io/cap-faq/
36. Questions? @xeraa
37. Images 4 Bad Neighbors 4 CAP: http://berb.github.io/diploma-thesis/community/ 061_challenge.html 4 Robinson Crusoe: https://flic.kr/p/8rbnLQ 4 Perpetuum mobile: https://flic.kr/p/5VvSJ8 4 Lemon: https://flic.kr/p/5wviJe 4 Carly Rae Jepsen: Call me maybe 4 Dirty & stale read: https://aphyr.com/posts/322-call-me-maybe- mongodb-stale-reads