ChatGPT Ain’t Got $%@& On Me! Next Generation Automated Database Tuning
summary
A new ProtoX holistic tuner for PostgreSQL 15 runs on Ubuntu 22 with 192 GB RAM and SSD and uses hypo_pg and pg_hint_plan plus small planner hooks to read per-node timings and estimate plan cost without fully building indexes. ProtoX searches knobs, table and index options (arbitrary-width, type, include columns) and query hints together, converges in about 14–16 hours, and finds configs that cut runtime ~78% vs PGTune+Dexter and ~28% vs PGTune+DTA+AutoSteer; LLM baselines (GPT‑4 and fine‑tuned Llama) are 3.6x and 1.7x worse.
AI-written summary · about
chapters (24)
- 0:00Introduction and background
- 1:58What is database tuning
- 4:48Role of DBAs and decline
- 5:54Current automated tuning tools
- 8:21How tuning tools work
- 10:05Limitations of individual tuners
- 11:09Sequential tuning approach
- 13:02Prisoner's dilemma in tuning
- 15:33Sequential tuning results
- 16:54State of the art holistic tuning
- 17:59Challenges of giant configuration space
- 18:40Key insight action similarity
- 20:28Defining similarity in tuning
- 24:00ProtoX evaluation setup
- 26:51Comparison with other tuners
- 30:03ProtoX results and performance
- 31:32Variety supported by ProtoX
- 34:01Integration challenges with PostgreSQL
- 36:25Future of tuning with language models
- 38:29Evaluating language models for tuning
- 41:41Accelerating tuning with prior knowledge
- 43:46Accelerating tuner setup
- 44:50Key takeaways and future directions
- 46:00Q&A session
related talks
27mBuilding safety tooling for risk-free AI tuning of Postgres
18mWhy Postgres Is Better Than MySQL for Automated Knob Tuning
28mAutotuning PostgreSQL on Azure Flexible Server
48mDBtune: AI-driven performance tuning for all PostgreSQL flavors
50mPerformance Archaeology
04mPostgres Tuning is saved by Sovenath Shaw
53mMake Your Database Dream of Electric Sheep: Designing for Autonomous Operation
20mDebate: Autotuning