Rely on session context when building sdist and wheel

[jaraco]

Summary of changes

Closes #3059

Pull Request Checklist

• Changes have tests
• News fragment added in changelog.d/.
  (See documentation for details)

[jaraco]

Github has messed up this PR. The diff is wrong and there are too many commits. If I look at the branch, it shows correctly that this branch is one commit ahead of main:

That's not pertinent to the failures, but it surely doesn't help in the analysis.

[jaraco]

Aha. I did some searching and found this article, which looks suspiciously similar, so I suspect its where the inspiration was drawn from, and confirms my misunderstanding about session fixtures. They're not automatically synchronized across workers in xdist, so each worker attempts to generate their own.

[jaraco]

I distilled the two issues I found with that article.

[abravalheri]

Hi @jaraco, thank you very much for the deep investigation.
Indeed, I did not realise the shortcoming of this technique...

I understand that this PR will solve the problem if xdist is not used, but when xdist is being used, things remain more or less the same. Is that correct?

Does it make sense / is it worthy to manually remove the created directory after the yield inside the context?

setuptools/setuptools/tests/contexts.py

Lines 120 to 125 in 253ecb0

	locked_dir = shared_dir / name
	with FileLock(locked_dir.with_suffix(".lock")):
	# ^-- prevent multiple workers to access the directory at once
	locked_dir.mkdir(exist_ok=True, parents=True)
	yield locked_dir

and transform setuptools_wheel and setuptools_sdist in "yield fixtures"? Would that solve the problem when using xdist?

[jaraco]

I understand that this PR will solve the problem if xdist is not used, but when xdist is being used, things remain more or less the same. Is that correct?

Yes. That's right. As I was investigating, I learned that tmp_path_factory.getbasetemp() changes depending on whether xdist is enabled, and it was only in the case that xdist was disabled or using only one process that the state was getting stored outside the session dir.

Does it make sense / is it worthy to manually remove the created directory after the yield inside the context?

Originally, I was concerned, but once I figured out the basedir mismatch and also learned that pytest normally leaves files behind to clean them up in the future, so I'm okay with it now that it gets cleaned up with or without xdist.