Training a 4B model to produce 81% faster query plans than Postgres
...or how to make Qwen learn query optimization via agentic reinforcement learning
Database Engineering
Summary
pgrust 0.2 achieves 300x speedup over Postgres on analytical workloads by replacing the row-at-a-time Volcano model with a batched query engine, operator fusion, and SIMD vectorization, targeting modern CPU and memory bottlenecks.
Highlights
- Replaces Postgres' row-at-a-time Volcano model with a batched executor, reducing per-row function call overhead
- Operator fusion and SIMD vectorization further reduce CPU and memory bandwidth per query
- Postgres was designed for disk I/O bottleneck era; modern workloads are CPU and memory bound
- A 500M-row SUM takes ~20s in Postgres vs 358ms in plain Rust, illustrating the gap
- pgrust 0.2 is 10x faster than its predecessor and even outperforms ClickHouse on Clickbench
auto-generated
Context
Audience
Database engineers, systems programmers, and developers interested in query engine internals and performance optimization
DomainDatabase Query Optimization
Formatlong-form technical blog post
Accessfree online
Source Overview
pgrust 0.2 is 10x faster than its predecessor, 30% faster than Postgres on OLTP workloads, and 300x faster on the Clickbench analytical benchmark.
The performance gain primarily comes from replacing Postgres' row-at-a-time Volcano model query engine with a batched executor that reduces per-row function call overhead. Additional optimizations include operator fusion and SIMD vectorization to lower CPU and memory bandwidth usage per query.
Topics
Related
Volcano modelSIMD vectorizationClickHouseClickbenchRust database enginesVectorized query execution
