compas_benchmarks

Benchmarks for the COMPAS core ecosystem.

This repository has one job: produce reproducible numbers about COMPAS, so design decisions in the core libraries are made on measurements rather than on argument. It measures COMPAS from the outside and never modifies it — anything a benchmark needs beyond the public API (such as a format-independent content hash) is implemented here.

What it measures is up to you: uv.lock pins released versions so runs are comparable by default, and any branch, tag, or commit can be installed over the top — locally or on CI — when the point is to measure unreleased work.

Suites

Suite	What it measures
serialization	wire size, serialize/deserialize time, peak memory, and round-trip fidelity across JSON, protobuf (compas_pb), and MessagePack — the measurement instrument for docs/PRD-serialization.md

Install

The project is managed with uv:

uv sync --all-extras

--all-extras adds the optional binary formats (compas_pb, msgspec, zstandard); a plain uv sync gives you COMPAS, JSON, and the tooling, and any format whose dependency is missing is skipped rather than failing the run. uv.lock pins the exact versions every measurement was taken against — CI syncs with --locked, so a run's numbers always name a known environment.

Benchmarking a branch

Any branch, tag, or commit of a subject under test can be installed over the synced environment — measuring unreleased work is the normal case, not an exception:

uv pip install "compas @ git+https://github.com/compas-dev/compas.git@some-branch"
uv pip install "compas_pb @ git+https://github.com/gramaziokohler/compas_pb.git@some-branch"

# ... and for an unpushed local checkout:
uv pip install /path/to/compas

Two things to know about the override. First, use uv run --no-sync afterwards (or export UV_NO_SYNC=1): a plain uv run re-syncs to uv.lock and would silently restore the released version, leaving you benchmarking something other than what you asked for. Second, uv pip install targets $VIRTUAL_ENV when one is active, unlike uv sync and uv run, which always find this project's .venv — so deactivate any unrelated environment first, or you will install into it by mistake.

uv sync --all-extras --locked puts everything back to the locked versions.

On CI, the same thing is a workflow input: compas_ref / compas_pb_ref accept a PyPI specifier (==2.15.1) or a git ref (git:some-branch).

Run

# Quick baseline — writes results/baseline_quick.{csv,html} + results/samples/
uv run python -m compas_benchmarks.serialization.run

# Full corpus (large; slow and memory-hungry)
uv run python -m compas_benchmarks.serialization.run --preset full --out results/baseline_full.csv

# Tests and style
uv run pytest
uv run ruff check . && uv run ruff format .

Run from the repository root: results land in results/, relative to the working directory. Every run writes a CSV (the machine record) and a self-contained HTML report next to it (the human one). results/ is git-ignored — benchmark output is reproduced, not committed.

A run reports each batch on stderr as it completes, so a long run shows its progress:

Benchmarking 3 formats over 4 batches (2 timed runs each): json, compas_pb, compas_pb_zstd
[2/4] mesh @ 10,000
         json                     864.1KB     34ms
         compas_pb                320.3KB     27ms
         compas_pb_zstd           168.6KB     29ms

Progress goes to stderr and the results table to stdout, so redirecting the table keeps the progress on screen. Every line is printed between measurements, never inside one — the timed regions are confined to metrics.measure — so this cannot influence the numbers. --quiet turns it off.

See docs/serialization.md for the corpus, the formats under test, and the findings so far.

On CI

The serialization-benchmark workflow runs the quick preset on every push to main and every pull request. The output is an interactive HTML report, so it is published to GitHub Pages rather than transcribed into a job summary: pushes to main and manual runs deploy the site (index + report + CSV + encoded samples), while pull requests upload the same site as a downloadable artifact and leave the published site showing main.

Use Actions → serialization-benchmark → Run workflow to choose the preset, repeat count, formats, and the compas / compas_pb versions to benchmark. Treat CI timings as indicative only — size and losslessness numbers are deterministic, timings on a shared runner are not.

Publishing needs Pages enabled once, in Settings → Pages → Build and deployment → Source: GitHub Actions. Each deploy replaces the site with that run's results; earlier runs stay available as workflow artifacts.

Build the same site locally with:

uv run python -m compas_benchmarks.site results --out site

Adding a benchmark suite

Add a subpackage under src/compas_benchmarks/, runnable as uv run python -m compas_benchmarks.<suite>.run, with its notes in docs/<suite>.md.

License

MIT — see LICENSE.