asv_bench_tachyon

ASV benchmark type plugin (same contract as asv_bench_memray).

Defines sample_* / Sample* benchmarks. asv run stores a number (estimated exclusive time per call from Python 3.15 Tachyon samples) in results for history, compare, and graphs.

There is no separate CLI package. Install this into the benchmark environment and write sample_* functions.

Related packages (metric plugin vs web UI)

asv_bench_tachyon is not the asv-tachyon web UI. Installing this package does not serve or replace the results site. It only registers a new benchmark type so sample_hot_loop is measured via Tachyon sampling and written into the usual ASV result JSON. Pair with asv-tachyon only if you want the modern dashboard over asv publish output.

Install

pip install asv_bench_tachyon

Python >= 3.15 (needs stdlib profiling.sampling). The package must be importable inside each ASV benchmark environment, not only on the host that runs asv. Put it on the ASV matrix:

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

Editable / local checkout during development:

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

If an environment was created before the matrix change:

rm -rf .asv/env
asv run --bench sample_

Benchmark attributes

Name the function sample_* or Sample* so the type is selected. Optional attributes (also accepted on the class or module, first wins via asv_runner):

sample_mode

Invalid values raise ValueError at run time.

sample_duration

How long the sampler and the call loop run. Longer duration → more stack samples → stabler exclusive-time estimate, at the cost of wall time per benchmark. Values below 0.05 are raised to 0.05.

Metric returned

Estimated seconds per call:

(self_samples * sample_interval_s) / n_calls

Self samples are counted for the benchmark function name when present; otherwise all collector direct-call counts are summed. Unit is seconds, type is time, so stock ASV history / compare / graphs work unchanged.

Example

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

A full mini-suite (two sample_* functions, matrix install, asv.conf.json) lives under examples/.

Discovery contract

asv_runner loads external types the same way as memray:

1. Distribution / package name starts with asv_bench
2. Package has a benchmarks/ subpackage
3. Each submodule may export export_as_benchmark = [SomeBenchmark, ...]

This package:

• Distribution name: asv_bench_tachyon
• Module: asv_bench_tachyon.benchmarks.sampletime
• Export: export_as_benchmark = [SampleTimeBenchmark]
• Name regex: ^(Sample[A-Z_].+)|(sample_.+)$

No asv profile GUI, no standalone CLI, no HTML UI. Metrics only.

License

MIT.