PyPI in the face: running jokes that PyPI download stats can play on you
summary
scikit-learn maintainers try to use PyPI counts and find BigQuery and clickhouse-backed dashboards misleading: automated cloud workflows dominate downloads, installer_type field shows many non-pip sources, CI tags cover ~10–15% only, and 50% from Amazon Linux 2 explains lingering Python 3.7 traffic. Version brownouts show ~3k/day transition-release downloads, and dependency counts reverse expectations (joblib ~20M fewer or more downloads than scikit-learn). The tooling surfaces patterns but not simple human-usage signals, so metrics need careful slicing before driving decisions.
AI-written summary · about
chapters (8)
- 0:00Introduction and background
- 2:34Why PyPI stats are misleading
- 8:20Sklearn package name confusion
- 11:28Python 2 and dependency quirks
- 13:17Human vs automated downloads
- 16:11Dependency download paradoxes
- 21:17The limits of metrics
- 24:48Conclusion and Q&A
related talks
01h 04mLightning Talks
20mMalicious Needle in a Haystack - PyPi Security Pitfalls
52mKeynote - Ee Durbin
32mHow (not) to upgrade a platform
28mWhyPy? Understanding Trends Driving the Usage of Python
18mMalicious Needle in a Haystack – PyPi Security Pitfalls
24mBuilding a Sustainable Python Package Index
34mPython as a Framework for Analytics and Growth Hacking