There's a known issue with shutil.rmtree on Windows, in that it fails intermittently.

The issue is well known (https://mail.python.org/pipermail/python-dev/2013-September/128353.html), and the agreement is that it cannot be cleanly solved inside shutil and should instead be solved by the calling app. Specifically, python devs themselves faced it in their test suite and solved it by retrying delete.

However, what to do about tempfile.TemporaryDirectory? Is it considered the calling app, and therefore should retry delete when it calls shutil.rmtree in its cleanup method?

I don't think tempfile is protected by the same argument that shutil.rmtree is protected, in that it's too messy to solve it in the standard library. My rationale is that while it's very easy for the end user to retry shutil.rmtree, it's far more difficult to fix the problem with tempfile.TempDirectory not deleting itself - how would the end user retry the cleanup method (which is called from weakref.finalizer)?

So perhaps the retry loop should be added to cleanup.

A simpler approach would be to simply ignore the error when it occurs in TemporaryDirectory._cleanup. There are no absolute guarantees about tempfile always cleaning up -- just a best effort. The complexity (and potential delays) of a retry loop seem more risky than simply occasionally not cleaning up -- that happens anyways, for a variety of reasons.

In addition to transient failures, this can also occur when, for example, files opened in the temporary directory (perhaps by library or application code not under direct control of the caller) haven't been properly cleaned up and their file handles don't get closed, resulting in permissions errors trying to delete them (particularly on platforms like Windows, that automatically lock files when opening them).

This case came up in e.g. this recent PR and rendered tempfile.TemporaryDirectory unusable for such use cases, forcing a reversion to the lower-level tempfile.mkdtemp without the cleaner, more robust and easier to interpret high-level interface that the former provides. Wrapping a with statement in a try-finally is syntactically ugly and semantically incongruous, and requires a second shutil.rmtree(..., ignore_errors=True)call to clean up in a best-effort manner, while when manually callingcleanup()` in a try-except, the finalizer still gets executed at a a non-deterministic later time when Python exits, raising an error.

Therefore, in the spirit of Guido's statements above in terms of providing a "best-effort" cleanup, I propose (and am willing to submit a PR to implement) adding an ignore_errors bool parameter (defaulting to False, of course, for backward compat--and should it be keyword only like errors to TemporaryFile?) to the tempfile.TemporaryDirectory constructor, which gets passed to shutil.rmtree on cleanup. This would not only address both cases here, but also one of the two discussed by Anthony Sotitle on BPO-25024, in a cleaner and simpler fashion that would take advantage of existing tempfile.TemporaryDirectory functionality and behavior.

Would a PR be accepted on this? If so, any specific guidance on tests and whether to mention it in What's New, etc., would be appreciated. Thanks!

SGTM

Thanks! Submitted as PR bpo-24793 #24793

New changeset bd2fa3c by CAM Gerlach in branch 'master':
bpo-29982: Add "ignore_cleanup_errors" param to tempfile.TemporaryDirectory() (GH-24793)
bd2fa3c