TypedDict subtypes ignore any other metaclasses in 3.9+

[sobolevn]

BPO	44919
Nosy	@ambv, @serhiy-storchaka, @sobolevn, @uriyyo
PRs	python/cpython#28057

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields: ```python assignee = None closed_at = None created_at = labels = ['type-bug', 'library', '3.9', '3.10', '3.11'] title = 'TypedDict subtypes ignore any other metaclasses in 3.9+' updated_at = user = 'https://github.com/sobolevn' ``` bugs.python.org fields: ```python activity = actor = 'lukasz.langa' assignee = 'none' closed = False closed_date = None closer = None components = ['Library (Lib)'] creation = creator = 'sobolevn' dependencies = [] files = [] hgrepos = [] issue_num = 44919 keywords = ['patch'] message_count = 2.0 messages = ['399615', '402781'] nosy_count = 4.0 nosy_names = ['lukasz.langa', 'serhiy.storchaka', 'sobolevn', 'uriyyo'] pr_nums = ['28057'] priority = 'normal' resolution = None stage = 'patch review' status = 'open' superseder = None type = 'behavior' url = 'https://bugs.python.org/issue44919' versions = ['Python 3.9', 'Python 3.10', 'Python 3.11'] ```

[sobolevn]

Some context. I have a User class defined as a TypedDict:

from typing import TypedDict

class User(TypedDict):
    name: str
    registered: bool

Now, I want to check if some dict is an instance of User like so: isinstance(my_dict, User). But, I can't. Because it raises TypeError('TypedDict does not support instance and class checks').

Ok. Let's change __instancecheck__ method then. We can only do this in a metaclass:

from typing import _TypedDictMeta

class UserDictMeta(_TypedDictMeta):
    def __instancecheck__(cls, arg: object) -> bool:
        return (
            isinstance(arg, dict) and
            isinstance(arg.get('name'), str) and
            isinstance(arg.get('registered'), bool)
        )

class UserDict(User, metaclass=UserDictMeta):
    ...

It looks like it should work. It used to work like this in Python3.7 and Python3.8.

But since Python3.9 it does not work. Let's try to debug what happens:

print(type(UserDict))  # <class 'typing._TypedDictMeta'>
print(UserDict.__instancecheck__(UserDict, {}))  # TypeError: TypedDict does not support instance and class checks

It looks like my custom UserDictMeta is completely ignored. And I cannot change how __instancecheck__ behaves.

I suspect that the reason is in these 2 lines: https://github.com/python/cpython/blob/ad0a8a9c629a7a0fa306fbdf019be63c701a8028/Lib/typing.py#L2384-L2385

What's the most unclear in this behavior is that it does not match regular Python subclass patterns. Simplified example of the same behavior, using only primite types:

class FirstMeta(type):
    def __instancecheck__(cls, arg: object) -> bool:
        raise TypeError('You cannot use this type in isinstance() call')

class First(object, metaclass=FirstMeta):
    ...

# User space:

class MyClass(First): # this looks like a user-define TypedDict subclass
    ...

class MySubClassMeta(FirstMeta):
    def __instancecheck__(cls, arg: object) -> bool:
        return True  # just an override example

class MySubClass(MyClass, metaclass=MySubClassMeta):
    ...

print(isinstance(1, MySubClass))  # True
print(isinstance(1, MyClass))  # TypeError

As you can see our MySubClassMeta works perfectly fine this way.

I suppose that this is a bug in TypedDict, not a desired behavior. Am I correct?

[ambv]

The change to disallow this in Python 3.9 was deliberate, see BPO-40187 and its description.

Your attempt to make isinstance() work with TypedDict subclasses goes directly against this design. In fact, PEP-589 explicitly says that:

• Methods are not allowed, since the runtime type of a TypedDict object will always be just dict (it is never a subclass of dict).
• Specifying a metaclass is not allowed.

I'm -1 to allow this but I'll also wait for Serhiy to voice his opinion.

[JelleZijlstra]

Agree with @ambv, let's close this as rejected.