Interpret benchmark results
Latest stable Based on Leader release 0.4.0
Use the 0.4.0 JMH data for controlled comparisons, not infrastructure purchasing decisions.
What was measured
Section titled “What was measured”The benchmark module uses kotlinx-benchmark with JMH. The recorded cross-backend runs use one fork, one thread, two warmups, and three one-second measurements on one machine. PostgreSQL, MySQL, Kubernetes, Redis extension, history recorder, and selected repeat runs have separate raw artifacts and constraints.
Read direction and uncertainty
Section titled “Read direction and uncertainty”Higher throughput and lower average time are better only within comparable rows. Error bounds are broad for several container-backed results; overlapping intervals do not justify a winner. Local and H2 rows measure in-process or local SQL overhead and are intentionally separated from distributed backend charts.
Use in engineering
Section titled “Use in engineering”Repeat the exact command before and after a code change on the same machine. Preserve environment, raw JSON, warmup, fork, and error data. For backend selection, benchmark the deployed topology and actual action; the release results cannot model your network, durability, or pool contention.
Release sources
Section titled “Release sources”benchmark/README.mddocs/benchmarks/2026-05-21-leader-cross-backend-baseline.mddocs/benchmarks/2026-05-29-issue-405-rdb-backend-throughput.json