Partitions for Everyone!
summary
A Jepsen harness runs five-node clusters under injected partitions and logs operations to check linearizability and loss. Zookeeper behaves as CP: 401/401 acknowledged writes all appear in the final set. NeoDB blocks for minutes and can stall long after partition heals. Kafka’s “in-sync replica set” design admits data loss when replicas shrink to one. Cassandra shows probabilistic isolation with about 1 in 250 corrupted rows and lightweight transactions that deadlock; use majority quorums and crash-and-partition tests.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:40What are our petitions
- 8:11Zookeeper
- 11:10Technical Terms
- 15:46Kafka Replication
- 21:49Cassandra
- 28:06Isolation
- 37:14Apache Cassandra
- 39:03Conclusion
- 40:40Clocks
- 41:19Liveness guarantees
- 41:59Write a test program
- 42:15Kill nodes
related talks
55mJepsen IV: Hope Springs Eternal
53mKeynote - Jepsen 7: Anna Concurrenina
45mJepsen II: Linearizable Boogaloo
42mJepsen 9: A Fsyncing Feeling
29mCassandra - an introduction
30mApache Cassandra: What’s New and What’s Next
41mCall me maybe Jepsen and flaky networks
25mImproving Bad Partition Handling in Apache Cassandra