gracefully handle missing files/directories in cache storage

[paveldikov]

Not sure if this is a common pattern, but we like to store the cache directory on /var/tmp as the usual home directory may be on a network share (and extremely slow I/O as a result).

This works great for us, but it also happens that /var/tmp is affected by a periodic cleanup process that deletes old files/directories. Once this happens, we get a bunch of error messages along the lines of:

  Caused by: Failed to install: packaging-24.1-py3-none-any.http.whl (packaging==24.1)
  Caused by: failed to read directory `/var/tmp/pavel/.cache/uv/archive-v0/REDACTED`
  Caused by: No such file or directory (os error 2)

rm -rf'ing the cache dir (or uv cache clean -- but not uv cache prune) seems to work around this problem fine, but it's quite a severe thing to do.

[charliermarsh]

Agree, I think we should be robust to this. I don't know if we can be robust to /var/tmp being cleaned up while uv is running though.

[paveldikov]

I think catching ENOENT and friends (at arbitrary depth?) and re-caching the distribution from scratch should be robust enough? I am trying to think of the various types of race conditions in there, but assuming that the hardlinking/copying into venv comes right at the end, this should be reasonably complete.

(symlink mode will not be helped by this, but that is probably fine?)