PyPI in the face: running jokes that PyPI download stats can play on you
summary
Analysis uses the public bigquery dataset of PyPI logs and concrete tools like Pepy and pip stats to show which requests inflate counts: mirrors, request-based downloads, CI and cloud automation. Surprising findings include a four‑year‑old scikit‑learn release (1.0.2) dominating installs due to Amazon Linux 2 users, joblib vs scikit‑learn download inversion, and ~20% of installs hitting the historical sklearn package name; debugging reveals interactions with installer_type, caching, and pip backtracking, and a one‑year brownout was deployed to force errors for legacy installs.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:09PyPI stats misleading
- 6:31Caveats in download metrics
- 8:59Inflated stats examples
- 11:48Dependency anomalies
- 14:18Package name quirks
- 16:44CI effects on stats
- 18:51Dangers of metrics
- 22:03Responsible stats use
- 22:47Q&A
- 23:51Q&A
- 25:37Q&A
- 27:19Q&A
- 28:42Summary
related talks
39mThe JupyterLab Extension Ecosystem- Trends from PyPI +GitHub
01h 04mLightning Talks
44mNobody Expects the Python Packaging Authority
30mThe Evolutionary Saga of Python Packaging
48mDeveloper Insights: ML and analytics on src/
28mBeyond the Benchmark: Why the “Best” Python Dependency Manager Might Not Be Best for You
31mA Recent History of Python Through Stack Overflow Questions and Answers
31mDevpi for teams