ACENLY Bench

Track Python function performance across commits. Catch regressions before they ship.

  ACENLY Bench  ·  main@a3f9c1b2  ·  Python 3.12.3

  Function                  Median        P95    vs last commit
  ──────────────────────────────────────────────────────────────
  deduplicate_users         1.24 ms    1.51 ms   ▼ 38.1% faster
  build_index              84.21 ms   97.13 ms   ▲ 12.4% slower  ⚠
  filter_records            0.88 ms    0.92 ms   ~ +0.2%

What it does

• Precise timing — runs each function in an isolated subprocess with warmup, batch timing, and outlier trimming
• Regression detection — compares against the last stored result and flags slowdowns
• Git hook integration — blocks pushes automatically when a regression exceeds your threshold
• History tracking — stores all runs in a local SQLite database so you can see trends over time

No external services. No accounts. Runs entirely on your machine.

Install

pip install pyyaml

Then copy bench.py and bench_db.py into your project root.

Quick start

# Benchmark a specific function right now
python3 bench.py myfile.py::my_function

# Show history for tracked functions
python3 bench.py --history

# High-precision mode (batch timing, 3s measurement window)
python3 bench.py --precise

Track functions automatically

Create acenly.yml in your project root (see acenly.example.yml):

benchmark:
  track:
    - file: mymodule/utils.py
      function: process_batch
    - file: mymodule/search.py
      function: find_duplicates

  regression_warn:  0.10   # warn if 10% slower
  regression_block: 0.25   # block push if 25% slower
  noise_floor:      0.05   # ignore changes smaller than 5%
  trials: 5
  warmup: 2

Then install the git hook:

python3 bench.py --install-hooks

From now on, every git push runs the benchmark automatically. If a function regresses past the block threshold, the push is stopped.

python3 bench.py --skip-hooks   # bypass when needed
python3 bench.py --uninstall-hooks

Options

How timing works

Each benchmark runs in a separate subprocess to avoid interference from the parent process. In normal mode, each function is called trials times with warmup discarded runs first. In --precise mode:

1. Adaptive warmup — keeps running until timing variance drops below 3% (CPU caches settled)
2. Batch calibration — finds a batch size so each measurement window takes ~50ms, then divides — this eliminates OS scheduler jitter from individual timings
3. 3-second window — collects ~60 batch measurements, reports the minimum (least OS interference)

Results are stored in bench.db alongside your project.

License

MIT