A Batch of Commit Batching
summary
PostgreSQL 9.2 introduces a lock and fast‑path change that makes PostgreSQL 9.2 group commit batch hundreds of logical commits into one fsync, yielding order‑of‑magnitude gains on tiny transactions (a laptop limited to ~120 raw commits/sec reaches ~12,000 with batching). Wide‑area synchronous replication suffers ~100 ms RTT so single‑client throughput is ~10 tps but scales to ~2,000 tps with ~300 clients. 9.3 adds a leader‑side commit_delay (≈half raw sync time ~4 ms) and per‑transaction synchronous_commit knobs so different durability levels coexist and work.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:44Storage Latency
- 3:49Synchronous Performance
- 5:25Network Latency
- 15:21Counter Tuned INSERT scaling
- 22:599.2 INSERT scaling
- 26:40User Selectable Durability
- 43:082nd Quadrant Training
related talks
35mUnderstanding and Optimizing commit sychronization for maximum performance
25mWaiting for Commit
42mlesser known Superpower of Postgres Logical Replication
48mReplication Refresh
48mLightning Talks
54mBreakout Session
51mHigh-concurrency distributed snapshots
26mMyths and Truths about Synchronous Replication in PostgreSQL