asv_bench_tachyon

Real ASV benchmark plugin (same shape as asv_bench_memray): a new metric type discovered by asv_runner, stored in results, plotted and compared across commits.

Uses Python 3.15 Tachyon (profiling.sampling) to estimate self-time per call from stack samples.

This is a metric plugin

Install

# into the same env ASV uses for the project (Python >= 3.15)
pip install asv_bench_tachyon

asv.conf.json:

{
  "matrix": {
    "req": {
      "pip+asv_bench_tachyon": [""]
    }
  }
}

Write a benchmark

class Suite:
    def sample_hot_loop(self):
        s = 0
        for i in range(200_000):
            s += i * i
        return s

    sample_hot_loop.sample_mode = "cpu"       # wall | cpu | gil
    sample_hot_loop.sample_duration = 0.25    # seconds of sampling

asv run --bench sample_hot_loop

Result unit: seconds (estimated exclusive time per call from sample counts × interval).

How it works

1. A background thread samples this process via profiling.sampling.sample(pid, ...).
2. The main thread runs your function in a loop for sample_duration.
3. Direct samples whose function name matches the benchmark are converted to time and divided by call count.

Same discovery contract as memray: export_as_benchmark = [SampleTimeBenchmark].

Optional CLI

asv-tachyon doctor

License

MIT.