pyarrow
pypi package · declared by 46 of 2,343 sampled repositories · graph computed 2026-08-24
Counted from the runtime dependencies public repositories declare — not the resolved tree. A transitive dependency is a consequence; a declared one is a choice. Development dependencies are excluded because they repeat across every project in a language regardless of what is being built.
Declared with
ranked by association, not by count| Declared with | Repositories | vs chance |
|---|---|---|
| pandas | 23 of 2,343 | 0.50 |
| datasets | 11 of 2,343 | 0.45 |
| torch | 14 of 2,343 | 0.39 |
| huggingface-hub | 9 of 2,343 | 0.39 |
| numpy | 24 of 2,343 | 0.38 |
| scikit-learn | 9 of 2,343 | 0.37 |
| transformers | 8 of 2,343 | 0.26 |
| tqdm | 9 of 2,343 | 0.26 |
| pydantic | 16 of 2,343 | 0.24 |
| pyyaml | 11 of 2,343 | 0.22 |
| openai | 8 of 2,343 | 0.19 |
| uvicorn | 8 of 2,343 | 0.17 |
| requests | 8 of 2,343 | 0.13 |
Every pair above is counted from the same evidence: repo manifest.
Used instead
candidates · shares a neighbourhood and refuses to co-occurNo alternative measured. A pair qualifies only if it shares a neighbourhood and then fails to appear together against an expectation that it would — and almost everything that shares a neighbourhood does appear together. That is the test working, not a gap in the data: similarity alone would list this tool’s closest neighbours as alternatives, and most of them are things it is used with.