Waiting for Commit
summary
Explains why synchronus replication can turn millisecond tail latencies into system-wide pauses via head-of-line blocking, gives measured timings (bare-metal commits ~0.3 ms, cross-AZ a few ms, packet loss ~0.1%), and shows troubleshooting with pg_stat_activity + perf record (and a small perf-analysis tool) to find OS/IO/network slow paths. Recommends quorum commit with ≥2 replicas to hide tails, warns about promotion logic, and suggests Patroni for safe automated failover.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:03Why replicate data
- 3:30Synchronous replication basics
- 4:57How Postgres WAL works
- 8:08Commit process in detail
- 10:53When things go wrong
- 13:25Sync levels and trade-offs
- 15:51Consistency pitfalls
- 18:00Detecting performance issues
- 20:36Solutions and optimizations
- 22:46Summary and recommendations
- 23:35Q&A
related talks
26mMyths and Truths about Synchronous Replication in PostgreSQL
39mBreakout Session
35mUnderstanding and Optimizing commit sychronization for maximum performance
47mA Batch of Commit Batching
01h 00mDistributed transactions implementation trade-offs
45mChasing unicorns : 0 dataloss and 99.999% availability
48mLightning Talks
48mUnderstanding and Reducing Application Contention