PerfEvolve: AI Agent Tunes PostgreSQL Outperforming Documentation by 35%
A research article advocates for a transition from static documentation to active strategies in database tuning, presenting PerfEvolve. This innovative system converts expert tuning techniques into actionable skills for LLM-based agents, facilitating version-consistency checks, profiling tailored to specific workloads, and joint optimization across multiple parameters. When evaluated on PostgreSQL using TPC-C and TPC-H benchmarks, PerfEvolve surpassed leading documentation-based methods by as much as 35.2%. The study contends that documentation merely reflects conclusions without the underlying reasoning, resulting in three main issues: obsolescence due to software updates, ineffectiveness with diverse workloads, and a lack of awareness regarding inter-parameter relationships. PerfEvolve is offered as open-source software.
Key facts
- PerfEvolve translates expert tuning methodologies into executable skills for LLM-based agents.
- Outperforms documentation-driven baselines by up to 35.2% on PostgreSQL.
- Tested under TPC-C and TPC-H benchmarks.
- Addresses three deficiencies of documentation: staleness, heterogeneous workloads, inter-parameter dependencies.
- Paper title: 'A Case for Agentic Tuning: From Documentation to Action in PostgreSQL'.
- Published on arXiv under Computer Science > Software Engineering.
- Tool is available at a provided URL.
Entities
Institutions
- arXiv