Datasets:
Download validation/tool_comparison/NOTES.md from pushkar87dg/pypi-api-drift: direct link, hf CLI and curl.
- Browser
- Download file 9.75 kB
-
https://huggingface.co/datasets/pushkar87dg/pypi-api-drift/resolve/main/validation/tool_comparison/NOTES.md
- Command line
-
hf download hf://datasets/pushkar87dg/pypi-api-drift/validation/tool_comparison/NOTES.md
-
curl -L -o NOTES.md https://huggingface.co/datasets/pushkar87dg/pypi-api-drift/resolve/main/validation/tool_comparison/NOTES.md
Tool comparison run: griffe and AexPy on 30 release pairs
Date: 2026-10-20. Machine: 2 CPUs, about 7 GB RAM, shared with other jobs. Only one install ran at a time, and every per-pair environment was deleted after use.
Scripts: scripts/fetch.py, scripts/run_griffe.py, scripts/run_aexpy.py. The scratch work dir lived outside the repo and was deleted.
Nothing under apidrift/code, release, out, pkg or raw, and nothing elsewhere in validation/, was read or used.
Versions
- Host python 3.13.16; uv 0.11.32; system pip 24.0 (AexPy calls
/usr/bin/pip downloaditself) - griffe 2.4.0, in a
uv venv --python 3.12venv - AexPy 0.4.3 (bb37ad3, 2024-03-13; the latest on PyPI), run with Python 3.12.13, mypy 2.4.0, pydantic 2.14.1, pip 26.2.2
griffe
uv venv --python 3.12 griffe-venv && VIRTUAL_ENV=griffe-venv uv pip install griffe
griffe-venv/bin/python -I scripts/run_griffe.py <workdir> [ids]
For each release, fetch.py downloads one artifact from the PyPI JSON API. It prefers a py3-none-any wheel, then a cp312 manylinux wheel, then any wheel, then an sdist. In practice all 60 releases had pure wheels. It unpacks the artifact into its own directory and lists the top-level importable names (directories that contain .py/.pyi files, plus top-level .py modules; tests/test excluded).
Each release gets a griffe.GriffeLoader(search_paths=[<that dir>], allow_inspection=False). The loader loads every top-level name, including private ones such as _griffe, so that public re-exports resolve. It then calls resolve_aliases(implicit=False, external=False), and griffe.find_breaking_changes(old_module, new_module) runs for each public top-level name present in both releases.
The analysis is purely static: no dependencies are installed and nothing is imported.
Timing (seconds in runs.csv) is wall-clock for the whole pair: download, unpack, load and diff. notes splits out fetch_s and analysis_s.
griffe classes every result as breaking, so breaking=yes on every row. The tool_label column holds the BreakageKind name. Kinds were mapped to change_type as follows:
| griffe BreakageKind | change_type |
|---|---|
| OBJECT_REMOVED | removed |
| PARAMETER_REMOVED | param_removed (vararg_removed / varkw_removed when the old param is *args / **kwargs) |
| PARAMETER_ADDED_REQUIRED | required_param_added |
| PARAMETER_CHANGED_REQUIRED (default removed, param became required) | default_removed |
| PARAMETER_MOVED | positional_reordered |
| PARAMETER_CHANGED_KIND | became_keyword_only / became_positional_only by the new kind, otherwise other |
| PARAMETER_CHANGED_DEFAULT, RETURN_CHANGED_TYPE, OBJECT_CHANGED_KIND, ATTRIBUTE_CHANGED_TYPE, ATTRIBUTE_CHANGED_VALUE, CLASS_REMOVED_BASE | other |
For parameter breakages, path is the function and param is the parameter name. griffe has no rename detection: a rename shows up as a removal plus an addition, or as nothing if the added parameter is optional. So param_renamed never occurs.
Fix during the run: build (pair 4) first produced an empty package list, because the sdist exclusion list contained build. That exclusion now applies to sdists only, and pair 4 was re-run (along with 22 and 27, which had suspiciously low analysis times; their results were unchanged). Only the latest row per tool and id is kept in runs.csv.
AexPy
uv venv --python 3.12 env && uv pip install --python env/bin/python aexpy==0.4.3 pip
echo <project>@<version> | PYTHONUTF8=1 env/bin/aexpy extract - <side>.json -r --no-temp
aexpy diff old.json new.json diff.json # run from a separate aexpy tool venv
(python3 -I scripts/run_aexpy.py <workdir> 0.4.3 [ids])
AexPy's default --temp mode needs micromamba or conda, which this machine does not have. Docker exists, but the README's documented --no-temp mode (use the current environment) works, so a fresh uv venv was built for every release. That venv holds AexPy and pip; AexPy then downloads the release wheel and runs pip install <wheel> with dependencies into the venv, imports it, and runs its mypy-based extractors. The venv is deleted right after extraction.
AexPy 0.4.3 requires Python>=3.12, so a Python 3.11 fallback is impossible. The fallback the runner originally had produced only an install error and was removed.
pip must be installed in the venv explicitly, because uv venv leaves it out. Without it, AexPy fails with "No module named 'pip'".
Timing (seconds) is wall-clock for the whole pair: both venv builds and installs, both extractions, and the diff. notes holds extract_old_s, extract_new_s and diff_s, plus a count of every diff entry by AexPy rank.
AexPy assigns each diff entry a BreakingRank: Unknown=-1, Compatible=0, Low=30, Medium=60, High=100. Its text report lists ranks of Low and above under "Breakings". In aexpy_changes.csv:
breaking=yesmeans rank High.breaking=potentialmeans rank Medium or Low. Low is AexPy's downgrade of a breaking change on a private API (privateRank). Thetool_labelis<Kind>|<Rank>, so Low rows can be filtered out.- Ranks Compatible and Unknown are not written. Unknown is mostly ChangeExternalAlias, as in
collections.abc.Callable -> typing.Callable. Their counts are in runs.csv notes, and the full diffs (messages only) are inaexpy_raw/<id>_<project>_diff.jsonl.
| AexPy kind | change_type |
|---|---|
| RemoveModule/Class/Function/Method/Attribute/InstanceAttribute/Alias/ExternalAlias | removed |
| RemoveRequiredParameter, RemoveOptionalParameter, Remove*Candidate | param_removed, refined to became_keyword_only / became_positional_only when a parameter with the same name exists in the new signature as keyword-only (kind 3) / positional-only (kind 0). AexPy itself reports positional-to-keyword-only as removal of the positional parameter. |
| RemoveVarPositional / RemoveVarKeyword | vararg_removed / varkw_removed |
| AddRequiredParameter, AddRequiredCandidate | required_param_added |
| RemoveParameterDefault | default_removed |
| MoveParameter | positional_reordered |
| everything else (ChangeParameterDefault, ChangeParameterType, ChangeReturnType, ChangeAttributeType, RemoveBaseClass, ChangeMethodResolutionOrder, ...) | other |
For alias entries, AexPy attaches the change to the containing module or class, so path is that container plus .<alias name>. For all other entries, path is the old entry's id, or the new entry's id if there is no old one.
AexPy has no rename detection either.
Failures and fixes (AexPy)
- 13 gitpython failed at first with
Different name between release and dist-info. AexPy asserts that the release id matches the METADATANameexactly, and here that name isGitPython. The runner now passes PyPIinfo.nameas the release id (GitPython@3.1.56), and the pair was re-run: ok. - 27 requirements-parser failed at first with
No top level module extracted. Version 0.10.2 importspkg_resourcesbut does not declare setuptools, and a fresh 3.12 venv has no setuptools. Fix: pre-installsetuptools<81into both extraction envs for this project only, then re-run: ok. This was the only per-project environment tweak. - 26 googleapis-common-protos extracted fine but failed in
aexpy diff. The sandbox memory cgroup OOM-killed the process (rc=-9; dmesg shows about 4.2 GB anon RSS). Each API json is about 700 MB, for two reasons. First, thegooglenamespace also covers the installed protobuf dependency (google._upb,google.protobuf), so AexPy analysed protobuf too. Second, AexPy stores huge per-entryaliaslists. I retried once with manually re-extracted files (same sizes); a diff alone on those files was not attempted. I did not work around the failure, because stripping fields would alter the tool's input. It might succeed on a machine with more than 8 GB free. - 20 yarl and 30 rich-toolkit were re-run as sanity checks, because both produced zero diff entries. Their API entry counts were non-trivial and identical across versions (yarl 9/8/70/124 modules/classes/functions/attributes), and the zero-entry result held.
- runs.csv keeps only the latest row per (tool, id). Rows in runs.csv written before the patch lack the
api=counts.
Caveats for comparison
- Scope differs. griffe analyses the wheel's own top-level packages, statically and without dependencies. AexPy imports the installed package, so for namespace packages (googleapis-common-protos) it can include the namespace content of dependencies. It also counts private APIs (rank Low) and external aliases.
- trio 0.22.2 to 0.23.0 accounts for 299 of AexPy's 473 rows. Many are
RemoveVarPositional/RemoveVarKeyword/AddRequiredParameteron High, probably from wrapper functions with*args, **kwargsin 0.22 that gained real signatures. These rows are reported as the tool gave them. - griffe flags
ATTRIBUTE_CHANGED_VALUE(e.g.__version__, constants), which AexPy does not report. The phonenumbers metadata constants are one case: 15 griffe rows against 0 AexPy rows. - Timing: griffe seconds include the download. AexPy seconds include two venv builds and dependency installs, which AexPy needs because it imports the package. The machine was shared with other CPU-heavy jobs (C++ builds, coverage runs), so the times are noisy.
Results
| tool | version | pairs ok | change rows | pairs with >=1 row | median s/pair (ok) |
|---|---|---|---|---|---|
| griffe | 2.4.0 | 30/30 | 96 (all breaking=yes) | 19 | 1.3 |
| AexPy | 0.4.3 | 29/30 (26 OOM in diff) | 473: 178 yes (High) + 295 potential (69 Medium, 226 Low) | 14 (13 with >=1 High) | 62.7 |
The full per-entry AexPy diffs, including Compatible and Unknown entries, are in aexpy_raw/ (one jsonl per pair, messages only).