whybig

"Why is this so big?" — point it at a Docker image, a Python wheel, a model checkpoint, a JS bundle, or a git repo. Get a treemap plus ranked culprits and concrete fixes.

$ whybig api-image.tar --html report.html

whybig — api-image.tar
kind: docker-image   total: 812.4 MB
layers: 14

Top culprits:
    340.2 MB   41%  root/.cache/pip
    121.7 MB   14%  usr/lib/python3.11/site-packages/torch
     48.9 MB    6%  app/.git

Findings (3, ~401.5 MB reclaimable):
  [WB001]   340.2 MB  pip cache shipped
           340.2 MB of pip cache at root/.cache/pip
           fix: pip install --no-cache-dir, or RUN --mount=type=cache,target=/root/.cache/pip
  [WB004]    48.9 MB  .git directory shipped
           ...

Zero dependencies, pure stdlib, Python 3.9+. Nothing is executed or unpickled — checkpoints are analyzed from their headers only.

Install

pip install whybig        # or just copy the directory; it's stdlib-only
whybig <path> [--html report.html] [--json] [--kind KIND] [--top N]

What it understands

--html writes a self-contained treemap (no CDN, works offline, dark/light): click to zoom, breadcrumb to go back, findings listed below the map.

Rules

Contributing

The contributor unit is deliberately small:

• One analyzer per artifact type — whybig/analyzers/<type>.py exposing KIND and analyze(path) -> Report. Register it in analyzers/__init__.py and detect.py. Wanted: .deb/.rpm, .apk (Android), .jar, ONNX, conda envs, npm packages.
• One size heuristic per rule — whybig/rules/<name>.py exposing RULE (next free WBnnn), TITLE, APPLIES (set of kinds), and check(report) -> [Finding]. Auto-discovered; no registration needed. Wanted: locales/i18n data, conda pkgs/ cache, .DS_Store/thumbnail litter, debug symbols in .so, duplicate vendored deps.

Every rule needs a test that builds its fixture programmatically (see tests/) — no binary fixtures in the repo.

python3 -m pytest tests/ -q

License

MIT