The PEP 544 specifies that:

A protocol can be used as a second argument in isinstance() and issubclass() only if it is explicitly opt-in by @runtime_checkable decorator.

It is not specified exactly whether this should be enforced by static type checkers or at runtime. Currently it is enforced in both cases: mypy flags this as error, and a TypeError is raised at runtime.

There is however a problem with current runtime implementation: abc and functools modules may call issubclass() on non-runtime checkable protocols if they appear as explicit superclasses. This is currently solved by a sys._getframe() hack in  typing module.

The problem is that current solution is incomplete. For example, the TypeError is not raised in the case of non-method protocols:

>>> class P(Protocol):
...     x: int
... 
>>> 
>>> class C: ...
... 
>>> isinstance(C(), P)
False  # should be TypeError

I tried to fix it this morning but after an hour of attempts to tweak the existing hack I gave up. It looks like there are only two reasonable solutions:

* Don't enforce @typing.runtime_checkable at runtime, make it a type-checker-only feature (like @typing.final).
* Add a little helper to abc module that would skip classes in MRO for which C._is_protocol is True but C._is_runtime_protocol is False.

Any thoughts/preferences?

If the "little helper in abc" solves it, let's do that. The sys._getframe() hack has always been a bit smelly -- I assume we can get rid of that then?

It seems like the straightforward, minimal fix would be to just add

if (getattr(cls, '_is_protocol', False) and
    not getattr(cls, '_is_runtime_protocol', False) and
    not _allow_reckless_class_cheks()):
    raise TypeError(...)

to _ProtocolMeta.__instancecheck__. Does that fail on some edge case (that the current implementation works on)? It's a little weird that _ProtocolMeta.__instancecheck__ doesn't explicitly check that the protocol is runtime-checkable.

Hi Kevin,

If you want to work on this, could you come up with some test cases that
try to explore edge cases for this behavior? Ideally some that already
pass, and some that currently don't pass (because of the limitation
mentioned by Ivan). It's possible that the tests for typing.py already
contain some suitable passing tests. You could then investigate whether
your proposal seems to work. If it does, you can submit a PR and we can lay
this issue to rest!

--Guido

Hi Guido,

I decided to help with this issue as far as Kevin did not respond.

I checked and this bug is not present at 3.11 version. So I simply added test to cover case mentioned by Ivan.

Please forget my previous msg, this bug still present at 3.11 version. PR contains fix proposed by Kevin.

Sorry for making to much noise.

New changeset a2d94a0a9b8ae95d7d2b7fc34b501da5242ec22c by Miss Islington (bot) in branch '3.10':
bpo-38908: Fix issue when non runtime_protocol failed to raise TypeError (GH-26067)
https://github.com/python/cpython/commit/a2d94a0a9b8ae95d7d2b7fc34b501da5242ec22c

New changeset 88136bbd0500b688c05e914be031cd3c243e42d8 by Ken Jin in branch '3.9':
[3.9] bpo-38908: Fix issue when non runtime_protocol does not raise TypeError (GH-26067) (GH-26075)
https://github.com/python/cpython/commit/88136bbd0500b688c05e914be031cd3c243e42d8

Yurii and Kevin, thanks for pushing the patch forward! Thanks for the review too Guido.

I'm closing this issue as all bugfix PRs have landed to bugfix branches. This will make its way into the next versions of Python for those respective branches.

One point of contention is whether to introduce this in 3.9.6. This will cause code previously working in 3.9.5 to throw an error if people were using it incorrectly. So it might be better to only enforce this in 3.10 onwards.

I created GH-26077 just in case the decision is made to revert. Sorry for any inconvenience caused everyone!

New changeset 9b90ce68503f4861ce4e9ac9444d9a82b3d943a5 by Ken Jin in branch '3.9':
[3.9] Revert "[3.9] bpo-38908: Fix issue when non runtime_protocol does not raise TypeError (GH-26067)" (GH-26077)
https://github.com/python/cpython/commit/9b90ce68503f4861ce4e9ac9444d9a82b3d943a5

New changeset 09696a3e218404e77f8c1fbf187ca29a4a357a9d by Miss Islington (bot) in branch '3.10':
[3.10] bpo-38908: [docs] Add changes to 3.10 whatsnew and fix some minor inaccuracies in news (GH-26096) (GH-26337)
https://github.com/python/cpython/commit/09696a3e218404e77f8c1fbf187ca29a4a357a9d