This pull request removes the pgrust entry from the ClickBench benchmark. The entry's reported results come from machinery written specifically for the 43 ClickBench queries, and its hot runs replay cached answers from the previous run. Both violate the benchmark rules, which require query result caches and caches near the end of the pipeline to be disabled. The entry is also tagged as untuned while it is tuned for these exact queries.
The author verified the claims against the published v0.4-preview release binary and the public v0.3 source. The aarch64 binary embeds source paths for one hand-written kernel per benchmark query, and error messages named after the benchmark appear in its plan matchers. The v0.3 source describes hand-lowered physical plans for all 43 queries, and several queries are answered from per-part statistics without scanning data.
Local measurements on a 96-core Graviton machine show that hot-run times reflect replayed verdicts rather than real execution. A substring query that took 10.8 seconds cold ran in about 5 milliseconds when repeated. The engine's true warm speed for such a scan is a few hundred milliseconds, close to ClickHouse on the same hardware. Disabling the cache settings in the configuration does not stop the repeated-query speedup.
Queries outside the 43 known shapes are refused, and the engine has no general execution path for the columnar table. The removal is intended to keep the benchmark's results comparable and honest.
Source: Hacker News · Summarized by HeadlinesBriefing