Before this PR:
sys.gettotalrefcount: 14288
sys.gettotalrefcount: 18042
sys.gettotalrefcount: 21796
sys.gettotalrefcount: 25550
sys.gettotalrefcount: 29304
sys.gettotalrefcount: 33058
sys.gettotalrefcount: 36812
sys.gettotalrefcount: 40566
sys.gettotalrefcount: 44320
sys.gettotalrefcount: 48074

After this PR:
sys.gettotalrefcount: 10537
sys.gettotalrefcount: 10540
sys.gettotalrefcount: 10543
sys.gettotalrefcount: 10546
sys.gettotalrefcount: 10549
sys.gettotalrefcount: 10552
sys.gettotalrefcount: 10555
sys.gettotalrefcount: 10558
sys.gettotalrefcount: 10561
sys.gettotalrefcount: 10564

Merged since it seems to fix multiple issues: #21902 (comment)

Does it fix your issue with encoding names and encodings._aliases?

cc @ncoghlan @ericsnowcurrently: Python finalization is getting better ;-) Yet another GC collection at exit!

Merged since it seems to fix multiple issues: #21902 (comment)

Does it fix your issue with encoding names and encodings._aliases?

Yes, this PR have fix my issue. there have no remaining encodings.aliases in refdumps.
Thanks a million, victor. I have learned much in this issue.

I'm not sure that the GC is still fully functionnal after PyInterpreterState is cleared, but at least, it seems like there were a bunch of reference cycles.

cc @pablogsal

Just curious, what are the three leaks.

Just curious, what are the three leaks.

You found this detail. I checked that some str object own the changed refcount in each iteration(not check the details).