BenchmarksSupported

The benchmarks dashboard starts with the 9 September 2026 optimization measurements, comparing baseline 9c8c206f with the changes integrated in 4c9cf2f4. It includes native JSON serialization, generated numeric validation, HTTP handling on Node/Bun/Deno, and eager/deferred nested PostgreSQL relation loading.

Serialization took 53–55% less time on the measured inputs. Mean HTTP throughput rose 6–20% across the four workloads; repository loading throughput rose about 10%. These are short before/after experiments: numeric validation differences were small, deferred-loading results varied, and neither ORM lane establishes a consistent p99 improvement. The dashboard includes all measured cases, sample ranges, runtime versions and downloadable per-run measurements.

The full-text indexes are in the shared database seed for every ORM participant. The corrected 13-route fixture passed 27 response-parity checks across zmdb, Drizzle and Kysely. No competitor was reranked using an older unindexed run against a newly indexed zmdb run. The earlier comparative investigation and profiles retain their original revision and measurement conditions.

The engineering panel retains the earlier measurements for editor/compiler operations, builds, packaging, installed consumers, startup and PostgreSQL replay. Each metric includes its unit, sample count, median and min–max range, with the captured revision, tool versions, date and machine details.

Download the summary and per-metric samples or complete raw runs and command output. The baseline script defines the metrics; the reproduction guide describes setup and individual workloads.

Percentile rows summarize per-run percentiles, not a pooled latency distribution. Installed size includes TypeScript and installation tooling. OS caches are not flushed, cached builds retain filesystem/dependency caches, and PostgreSQL shares the machine with the load generator.

The older upstream competitor results remain separately labelled historical. Their PostgreSQL 16, validation and HTTP peer results were not rerun with either the engineering capture or the latest optimization measurements. The historical HTTP panel also identifies its separate dated Node refresh.