GeezerVison

_ProtocolMeta.__instancecheck__ treats a member set to None as "explicitly not implemented" only for callable members:

if val is None and attr not in cls.__non_callable_proto_members__:
    break

_proto_hook applies the same sentinel to every member, callable or not. So a class that sets a non-callable protocol member to None passes isinstance() but is rejected by the subclass hook. It lands in ABCMeta's negative cache instead of the positive one, and every subsequent isinstance() call falls through to the getattr_static loop over all protocol members.

>>> [r() for r in _abc._get_dump(P)[1]], [r() for r in _abc._get_dump(P)[2]]
([<class 'Good'>], [<class 'Bad'>])   # positive, negative

This makes the check O(N) in the number of protocol members, on every call, for a class that conforms.

Performance

Release build of this branch, macOS on Apple silicon, best of 11 runs of 20,000 isinstance() calls, ns/call. P has N property members; Bad differs from Good only in setting the last one to None.

The before column scales at roughly 500 ns per protocol member. After the change Bad is a cache hit like Good, and Good itself is unaffected.

This gives _proto_hook the same rule, so such a class is cached like any other. None-valued method members are still rejected, and a missing attribute still fails.

The only externally visible change is via _allow_reckless_class_checks: an issubclass() originating in abc, functools or _py_abc now returns True where it returned False, which is what isinstance() already answers for the same pair. A direct issubclass() still raises TypeError for protocols with non-method members, so user-facing behaviour is unchanged.

• Issue: typing: one None-valued member makes isinstance() on a runtime_checkable Protocol O(N) per call #156413