Executive summary
One-line health per category. Green = within budget. Amber = watch. Red = investigate.
Loading…
Median latency by benchmark
Horizontal bars, log-scale, shortest first. Hover to see raw ops/sec.
Category breakdown
All benchmarks grouped by subsystem. Click a category to jump to the full detail table.
Loading…
Full benchmark table
Sortable, searchable. All the raw numbers behind the summary above.
| Category | Benchmark | Median latency / RSS growth | Ops/sec | σ (%) |
|---|---|---|---|---|
| Loading benchmark data… | ||||
Methodology & disclaimers
How the numbers are produced
- Runner: GitHub-hosted
ubuntu-latest. Shared-runner load and allocator reuse can affect individual measurements. - Command:
pytest bench/ --benchmark-only --benchmark-min-rounds=5against a clean checkout. - Latency: median wall-time, ops/sec, standard deviation. The chart includes latency measurements only.
- Memory: median sampled peak RSS growth in bytes, sampled every 5 ms, with a fixed per-workload budget. Legacy runs without RSS byte samples show “not measured”.
- Cadence: every push to
master. Thebench-pages.ymlworkflow rebuilds this page automatically.
How to read the numbers
- Median < standard deviation means an unstable measurement — do not draw conclusions from a single run.
- Log-scale bars in the chart above compress the range so you can see µs and ms benchmarks side by side.
- Category KPIs pick the representative benchmark for that subsystem. See
bench/for the source. - Compare the same workload version and units on comparable hardware. Memory sampling can miss peaks shorter than the sampling interval.