pgrust, the agent-built Postgres rewrite in Rust, ships v0.4-beta in the open after a ClickBench removal

pgrust is a rewrite of PostgreSQL in Rust, built largely with coding agents. On 22 October it published v0.4-beta, and the code that the project had kept private behind monthly drops is now developed in the open. The release lands a month after ClickBench threw pgrust off its leaderboard, and two days before Dan Luu held it up as his main example of what agent-built software can now do. His essay passed 500 points on Hacker News over the weekend.

What v0.4-beta adds

The release post lists four changes:

• Gravity, a new query engine. A new Lisp-based intermediate representation feeds a rewritten JIT that now targets x86, ARM and WASM, where the old JIT only targeted ARM64. The project says Gravity runs “as much as 3× faster than ClickHouse” and that row-store queries over more than a few hundred rows can be up to 5x faster.
• pgrcolumnar gains full MVCC, fast inserts, updates and deletes, and a columnar index that sits on top of an ordinary row-store table. Standard index types such as B-tree, GIN and GiST work on columnar tables.
• Automatic memory: work_mem and the buffer cache size themselves from free memory, which pgrust says its single-process, thread-per-connection design makes possible.
• Native statistics for jsonb columns, so the planner can estimate selectivity inside JSON.

The post is plain about readiness: pgrust “is still experimental and not ready to be the primary database for data you cannot afford to lose.” It suggests analytical replicas or test workloads, and lists memory as a known limit, with some large analytical queries refused rather than risk an out-of-memory crash. The repository is AGPL-3.0.

How it was built

Michael Malis started pgrust in April. His first post says “coding agents, mainly Codex, have been a massive accelerant” and reports 250,000 lines in two weeks. A July write-up says that after four attempts and $100,000, the team “succeeded in using Opus”, running “up to 40 concurrent subagents for days” to produce 1.8 million lines of Rust. The post’s title says the result passed 100% of the Postgres regression suite. Those figures are the author’s; we have not counted them.

The ClickBench removal

On 22 September, ClickHouse’s Alexey Milovidov merged ClickBench PR #2165, titled “Remove pgrust: the entry is cheating (per-query kernels and result caching)”. It says the submitted v0.4-preview binary embedded source files named after ClickBench query numbers, that hot runs replayed cached answers, and that the v0.4-preview source was not published. The PR reached Hacker News on 23 October.

Malis answered there the same day. He says the per-query code was a probe of the new IR that “does not run during the benchmark”, and that the cache holds which predicates match dictionary strings, not query results, a kind of intermediate cache he says ClickBench allows. He concedes that the preview’s new engine handled little SQL beyond the 43 benchmark queries. Luu writes that “the accusations of nefarious cheating seem to be false” and that Malis acknowledged the submission was premature. The v0.4-beta post still leads with the ClickHouse comparison, which is pgrust’s own number, not a ClickBench listing.

Bugs found in Postgres itself

Luu’s case rests partly on testing. He links five reports to the pgsql-bugs list, copied to Malis’s pgrust address. BUG #19600, logged by Malis on 3 August against 19beta1, shows pgcrypto’s crypt() returning text with a broken UTF-8 sequence that the server would reject as input but stores anyway. BUG #19586 is a money division overflow. A footnote in the essay notes that some of the reports were not considered bugs. Luu also says Malis now uses Lean to verify parts “too complex to do in Kani”.

Why it matters

pgrust is among the most ambitious agent-written Rust projects we have covered, after smaller efforts such as php-rust. It also shows the open problem with such projects: speed claims get checked by outsiders, and a premature benchmark entry can follow a project around. Malis’s own request in the release post is for bug reports on any query where pgrust diverges from Postgres.