Jepsen II: Linearizable Boogaloo
summary
Jepsen runs generative tests and a nasus verifier checks histories for linearizability; an automated 48-core moon laser lab consumes 97% CPU to stress clusters. Findings are concrete: RabbitMQ’s minority-recovery mode destroys roughly 35% of acknowledged inserts; etcd exposes stale reads because some reads bypass the Raft log; Elasticsearch loses hundreds of writes (examples: 688 inserts lost, 645 CAS failures). Fixes include timeouts, docs changes and optional linearizable reads.
AI-written summary · about
chapters (9)
- 0:00Introduction to Jepsen
- 1:06Distributed system realities
- 2:18Consistency models and CAP
- 7:05Jepsen testing methodology
- 10:59Tooling and implementation
- 18:04Case study: RabbitMQ
- 24:47Case study: etcd and Consul
- 30:00Case study: Elasticsearch
- 40:23Conclusions and takeaways
