Clarification about how to implement namespace packages (as in PEP 420) via import hooks for PEP 660 use case

まず、参照されている namespace package の実装の周辺にある Lib/importlib._bootstrap_external.py を読み、その後、PEP 420 とリンクされている Discourse の議論および proof-of-concept gist を比較します。Solutions A または B が仕様と互換性があるかどうか、また import hooks が PEP 660 のユースケースをどのようにサポートすべきかを明らかにして文書化できれば完了です。

This is a follow up of 2 previous discussions in the Python discourse and an issue in pypa/packaging-problems:

• https://discuss.python.org/t/pep-660-and-setuptools/14855/15
• https://discuss.python.org/t/how-to-implement-namespace-packages-as-in-pep-420-via-import-hooks/15192
• https://github.com/pypa/packaging-problems/issues/588

Summary

One of the popular ways of implementing PEP 660 is via import hooks (since other alternatives like symlinking package folders are not available in all platforms).

However the reference implementation editables currently does not support namespaces.
Specifically the dynamic path computation aspect of PEP 420 is a challenge.

After inspecting the implementation for importlib._bootstrap_external, I may have managed to find 2 solutions for this problem, described in the detail in this discourse post and demonstrated as a PoC in this gist:

• Solution A: Create a custom "path class" that tries to emulate the _NamespacePath behaviour.
• Solution B: Create a custom PathEntryFinder and add a “bogus” entry to sys.path just to trigger it (implicitly relying that importlib.machinery.PathFinder will take care of setting __path__ to _NamespacePath object under the hood).

However, the problem is that I don't know if these approaches are reliable in the long run and in accordance with the language specification.

Open Questions

Questions like the following are still not clear to me:

1. Is there a chance that an internal check for namespace packages will fail if __path__ uses a custom class instead of _NamespacePath even if the behaviour described in PEP 420 is observed?
2. Can we rely on the fact that the import machinery will automatically set the namespace loader for specs in the form of ModuleSpec("name", None, is_package=True)?
3. Is pkgutil (and particularly pkgutil.extend_path) still considered first class citizen of the standard library or is it better to avoid it when writing new code?
4. Can we rely on the behaviour of sys.path_hooks + importlib.machinery.PathFinder to create ModuleSpec objects for namespaces that will automatically perform dynamic path computation? Or is this an internal implementation detail that can change in the future?

If the discussed approaches are not recommended, how can we support namespace packages using import hooks and deliver a complete solution for PEP 660?