SQL Execution Internals – From Database to OS and Hardware Levels
summary
Workshop shows a reproducible methodology using YB stats snapshots and OS metrics to locate latency: a SELECT over 1M rows measures 12.3 s end-to-end, with tablet-server reads ~11.2 s and physical IO accounting for ~58% of that; RocksDB reports RocksDB SSD read micros with ~0.3 ms per SSD read, then OS page cache drops reads to microseconds on re-run. First backend bootstrap hits master and costs ~200 ms vs ~20 ms typical logon. Uses PG gate and tablet-service traces to separate parse/plan/execute, local vs remote dispatch, and aggregate pushdown (SELECT COUNT() ~1.5 s).
AI-written summary · about
chapters (14)
- 0:00SQL execution internals
- 3:20SQL tuning importance
- 5:25YB stats introduction
- 9:39Using YB stats
- 13:10YugabyteDB cluster architecture
- 22:53SQL connection backend
- 39:59Postgres query phases
- 47:54YSQL and PG gate layers
- 55:02Query execution with YB stats
- 1:09:28RocksDB and storage
- 1:22:24OS caching and I/O
- 1:30:31Parallel query execution
- 1:35:38Key takeaways
- 1:46:23Q&A on indexes
related talks
26mUnderstanding the Query Execution in YugabyteDB
56mLightning Talks
30mOptimising your Database for Analytics
48mModern SQL features in PostgreSQL
31mProcessing data: Ruby or SQL?
37mDatabase Wizardry for Legacy Applications
41mUsing Databases to Pull Your Application Weight
45mPerformance tricks you have never seen before 2.0