Consider removing hook from CPython PathFinder

[jaraco]

In GitLab by @jaraco on Sep 11, 2019, 06:26

Given the complicated interactions between importlib_metadata and importlib.PathFinder, revealed by #86 and also by the recent complications I discovered today where cherry-picked changes to importlib.PathFinder from CPython master no longer apply cleanly to the 3.8 branch (due to the generated C header file from make regen-importlib), I'd like to consider:

• The least invasive approach would be to replace the code in PathFinder.find_distributions with a small, unchanging hook that calls into importlib.metadata. This change would address the latter issue (with cherry picks) but not the interaction issue.
• A more invasive approach would be to strip the importlib.PathFinder functionality out altogether and install a MetaPathFinder to sys.meta_path just as importlib_metadata does. This approach would substantially reduce the interactions between importlib_metadata and importlib.metadata, although the two would still interact (as both would install DistributionFinders into sys.meta_path).
• Even more aggressive might be to simply disallow the interaction of importlib_metadata and importlib.metadata, either by only letting one import in a given session, or by discouraging install of importlib_metadata on Python 3.8+, or perhaps another way.
• The most aggressive approach could be to remove importlib.metadata altogether and rely on pip-installed importlib_metadata.

I'm not fond of any of these approaches, except maybe the first.

@warsaw Do you have any thoughts on it?

[jaraco]

In GitLab by @jaraco on Sep 11, 2019, 06:35

mentioned in commit 83b48ec

[jaraco]

In GitLab by @jaraco on Sep 11, 2019, 06:41

mentioned in merge request !92

[jaraco]

In GitLab by @jaraco on Sep 11, 2019, 06:43

changed the description

[jaraco]

In GitLab by @jaraco on Sep 11, 2019, 10:50

In !92, I've enacted option one. Additionally, I've updated the wiki accordingly.

[jaraco]

In GitLab by @jaraco on Oct 26, 2019, 18:36

As this issue is partially addressed and as #91 captures the remaining interaction issue, I'm closing this issue.

[jaraco]

In GitLab by @jaraco on Oct 26, 2019, 18:36

closed