Benchmarks

Review the recorded SDK, ingest and recovery results and reproduce the tests.

Run these commands from a clone of the repository with Bun and Docker installed.

SDK logging

The comparison measured 10,000 scalar calls per tool. Rundial used a local accepting transport and disk spool. The evidence does not record the host model.

Tool	Version / mode	p99 per call
Rundial run.log_metric()	Local transport	10.52 µs
W&B	0.27.2, offline	193.478 µs
MLflow	3.13.0, SQLite	2,761.424 µs

bun run bench:sdk:comparative:competitors:docker

See the raw results and methodology.

Ingest and queries

The recorded compose-backed run ingested 5,000 points at 124,293 points/s, with chart query p95 of 60.32 ms across two queries. Results depend on the host and workload.

bun run bench:ui:responsiveness:compose

See the raw results and methodology.

Recovery

The deterministic SIGKILL, simulated API outage and simulated ClickHouse restart scenarios each recorded zero accepted points lost. Compose-backed tests also recorded zero accepted points lost after an API container kill and a ClickHouse restart.

bun run bench:durability
bun run bench:durability:compose

See the deterministic results, compose results, and methodology.

Next steps

• Log a training run
• Limits