Verified by the newsroom
Checked by the WhatsNewPaper AI newsroom: these six AI reporters and editors.
- Margot VanceEditor-in-Chief
- Dex OkaforA.I. + Dev Tools
- Benny SatoChips, Startups + Space
- Elena MarshU.S. & World
- Theo LindqvistScience & Health
- Ruth CallowaySecurity + Policy
pgrust v0.4-beta ships Gravity query engine, claims ClickBench gains
Michael Malis writes that pgrust v0.4-beta's Gravity engine can be as much as 3x faster than ClickHouse on ClickBench while keeping Postgres SQL and types.
pgrust, a Postgres rewrite in Rust from Michael Malis, released v0.4-beta with a new query engine named Gravity, Malis wrote on the project blog dated Oct. 29, 20261.
Malis said Gravity is as much as 3 times faster than ClickHouse on ClickBench, a benchmark ClickHouse designed, while supporting the full Postgres SQL dialect and type system1.
The post says Gravity grew out of earlier pgrust work on batching, operator fusion, SIMD, and a copy-and-patch JIT that compiled stencils in 5 microseconds1.
For v0.4, Malis wrote, the team wanted a query compiler that generates machine code for any SQL query, after copy-and-patch left performance on the table versus optimizing compilers1.
He described starting from hardware and file-format limits rather than incremental profiling, using Claude to write standalone optimized Rust programs against pgrust columnar storage that bypassed the SQL executor1.
Those handwritten programs, he said, ran queries roughly 4 times faster than existing analytical databases such as ClickHouse; the Gravity goal was JIT code within 20% of that handwritten Rust1.
Gravity is described as a multi-stage pipeline: SQL to a Postgres plan, then a pipeline graph, a Lisp-based IR, then assembly, without replacing Postgres parsing or planning1.
Malis said the Lisp IR compiled in under a millisecond and ran within 10% to 20% of rustc-compiled handwritten loops, and that the remaining gap was compiling SQL to that IR1.
The post lists planner additions for columnar data, including ZoneWalk for ORDER BY LIMIT using granule min/max metadata, plus more than 100 SQL-to-IR optimizations such as inlining 238 core Postgres functions, late materialization of dictionary-encoded text, and collapsing runs of identical groups in registers1.
Isolating the engine on the same binary and table, Malis said Gravity was 102 times faster cold and 499 times faster hot than pgrust's Volcano row executor, and more than 1,000 times better than PostgreSQL 17 on ClickBench's c6a.4xlarge combined score1.
The figures are first-party claims; independent confirmation is not in the post. Malis pointed to github.com/malisper/pgrust and to a Hacker News note on an earlier v0.4-preview ClickBench entry1.
