Poetry Cache
Poetry keeps two main cache areas on disk: package metadata under its cache directory and downloaded distribution artifacts such as wheel files and source tarballs under artifacts, so it can resolve dependencies and reinstall packages without hitting PyPI every time. Those caches speed up lock, install, and update operations, but they can also accumulate outdated index responses, corrupted wheels, and stale source archives. Kudu removes Poetry's cached metadata and downloaded distributions while leaving your projects, lockfiles, configured environments, and the virtualenvs folder untouched.
Why clean Poetry Cache?
- Stale package index metadata can make Poetry insist an older version is the newest available, so update or add commands fail to see releases you know exist on PyPI
- A corrupted cached wheel or source tarball causes repeat install failures with checksum, unzip, or build errors until the bad artifact is removed
- Large artifacts caches grow from repeated downloads of platform-specific wheels and sdists, and users notice gigabytes consumed under the pypoetry cache folders
- Cached metadata from an earlier repository state can lead to confusing dependency resolution results, where Poetry keeps selecting versions that no longer match current constraints
- Reusing old downloaded archives after a package was republished or yanked can produce inconsistent install behavior, with one machine succeeding from cache while a fresh machine resolves differently
- After clearing the cache, the next install may take longer because Poetry re-downloads metadata and distributions, but it often fixes loops of failed installs caused by poisoned cached files
Cache paths Kudu targets
Windows
%LocalAppData%/pypoetry/Cache/cache |
%LocalAppData%/pypoetry/Cache/artifacts |
macOS
~/Library/Caches/pypoetry/cache |
~/Library/Caches/pypoetry/artifacts |
Linux
~/.cache/pypoetry/cache |
~/.cache/pypoetry/artifacts |
Common questions about Poetry Cache
Download Kudu and reclaim your disk space.
Available on Windows, macOS, and Linux. No account required, no feature gates, no telemetry without consent. All cleaning targets are open source and community-auditable.