In the C implementation of zoneinfo.ZoneInfo, __init_subclass__ is not declared as a classmethod, which prevents it from being subclassed. This was not noticed because the tests for ZoneInfo subclasses in C are actually testing zoneinfo.ZoneInfo, not a subclass, due to a mistake in the inheritance tree: https://github.com/python/cpython/blob/8f192d12af82c4dc40730bf59814f6a68f68f950/Lib/test/test_zoneinfo/test_zoneinfo.py#L465-L487


Originally reported on the backport by Sébastien Eustace: https://github.com/pganssle/zoneinfo/issues/82

The fix in the backport is here: https://github.com/pganssle/zoneinfo/pull/83

New changeset 87d8287865e5c9f137f6b5cf8c34c2c509eb5e9d by Paul Ganssle in branch 'master':
bpo-41025: Fix subclassing for zoneinfo.ZoneInfo (GH-20965)
https://github.com/python/cpython/commit/87d8287865e5c9f137f6b5cf8c34c2c509eb5e9d

New changeset 33d3c64095bcdf9066a3441f6dda5d2e2f4118a8 by Miss Islington (bot) in branch '3.9':
bpo-41025: Fix subclassing for zoneinfo.ZoneInfo (GH-20965) (GH-21876)
https://github.com/python/cpython/commit/33d3c64095bcdf9066a3441f6dda5d2e2f4118a8

Łukasz: Would it be possible to backport / cherry-pick the changes from PR GH-21876 (https://github.com/python/cpython/pull/21876) into the 3.9.0 branch? I think this is a fairly severe issue considering that pendulum is planning to use a zoneinfo subclass.

It was merged as commit 33d3c64095bcdf9066a3441f6dda5d2e2f4118a8.

Thanks!

Marking as release blocker to put it on the checklist. Feel free to demote it if you decide it should be deferred to 3.9.1.

> New changeset 87d8287865e5c9f137f6b5cf8c34c2c509eb5e9d by Paul Ganssle in branch 'master':
> bpo-41025: Fix subclassing for zoneinfo.ZoneInfo (GH-20965)

This change introduced a reference leak. 3.9 and master branch are affected.

$ make && ./python -m test -R 3:3 test_zoneinfo 
(...)
test_zoneinfo leaked [84, 84, 84] references, sum=252
test_zoneinfo leaked [41, 41, 41] memory blocks, sum=123
(...)

ZoneInfoSubclassTest and CZoneInfoSubclassTest test cases leak.

Example of test method which leaks:

test.test_zoneinfo.test_zoneinfo.CZoneInfoSubclassTest.test_folds_and_gaps

There are two refleaks here. One is a reference leaking to the weak cache in `__init_subclass__` (one leak every time a subclass is created), and the other is that when `subclass.clear_cache()` is called, it sets `ZONEINFO_STRONG_CACHE = NULL`, thus causing a reference leak to the parent class's strong cache.

I'll send a PR to fix it shortly.

> test_zoneinfo leaked [84, 84, 84] references, sum=252

Reported as bpo-41568 and fixed there.

Paul: What is the status of this issue? Is it already fixed? It is the release blocker priority.

This is fixed in the 3.9 and master branches, it needs to be cherry-picked into the 3.9.0 release (at Łukasz's discretion, of course).

Łukasz: If you cherry-pick GH-20965/GH-21876 into the 3.9.0 release, please also cherry-pick GH-21907/GH-21912, since that fixes the refleak. (PRs are listed as "master"/"backport" since I don't know your workflow).

I will be cutting RC2 from the head of 3.9.