(In very specific circumstances)

Bug report

Due to some changes that have been made to the implementation oftyping.Protocol in Python 3.12, whether or notisinstance([], collections.abc.Mapping) evaluates toTrue orFalse now depends on precisely when and if garbage collection happens.

Minimal repro:

>>>import gc, collections.abc, typing>>> gc.disable()>>>try:...classFoo(collections.abc.Mapping,typing.Protocol):pass...exceptTypeError:...pass...>>>isinstance([], collections.abc.Mapping)True

Why does this happen?! It's becauseFoo is an illegal class, soTypeError is raised during class initialisation here:

TypeError being raised here means thatFoo never has__protocol_attrs__ set on it.But, the error is raisedafter class construction, meaning that if garbage collection happens at the wrong time, a reference toFoo is still retained incollections.abc.Mapping.__subclasses__(). This then creates problems in theisinstance() check, because of some behaviour in theabc module where it iterates through all of the subclasses ofcollections.abc.Mapping in order to determine whether an object is an instance ofcollections.abc.Mapping.

>>>import gc, collections.abc, typing>>> gc.disable()>>>try:...classFoo(collections.abc.Mapping,typing.Protocol):pass...exceptTypeError:...pass...>>> x= collections.abc.Mapping.__subclasses__()>>> x[<class 'collections.abc.MutableMapping'>, <class '__main__.Foo'>]>>> x[-1].__protocol_attrs__set()

The fix is to raiseTypeError before the class has actually been created, rather than during class initialisation.

Why does this matter?

This might seem like an absurdly specific bug report, but it would be good to see it fixed. It was causing bizarre test failures in thetyping_extensions backport. The issue was that we did this in one test method:

deftest_protocols_bad_subscripts(self):T=TypeVar('T')S=TypeVar('S')withself.assertRaises(TypeError):classP(typing.Mapping[T,S],Protocol[T]):pass

...And that finalP classwasn't being cleaned up by the garbage collector immediately after that test having been run. That then caused an unrelated test elsewhere to fail due to the fact thatisinstance([], collections.abc.Mapping) was evaluating toTrue, and it wasvery hard to figure out why.

Linked PRs

• gh-105280: Ensureisinstance([], collections.abc.Mapping) always evaluates toFalse #105281
• [3.12] gh-105280: Ensureisinstance([], collections.abc.Mapping) always evaluates toFalse (GH-105281) #105318