"Dynamic metaclass not supported" false positive on generics

[DetachHead]

from typing import TypeVar, Generic

T = TypeVar("T")

class FooMetaclass(type, Generic[T]): pass

class Foo(metaclass=FooMetaclass[T]): # error: Dynamic metaclass not supported for "Foo"
    pass

https://mypy-play.net/?mypy=master&python=3.10&gist=d4b169a5ece08ed0607b19fff8e9b749

according to the documentation:

Mypy does not understand dynamically-computed metaclasses, such as class A(metaclass=f()): ...

however this isn't dynamically computed, it's just a generic so mypy should be able to understand it statically

[sobolevn]

I am not sure that generic metaclasses is even a thing 🤔 I've never seen any references in PEPs / docs. Can you please check that it can be used this way?

[DetachHead]

pyright supports it and it works at runtime, so i see no reason these two features shouldn't work together for this to emerge.

from typing import TYPE_CHECKING, Generic, TypeVar

T = TypeVar("T")

class FooMetaclass(type, Generic[T]):
    value: T

class Foo(metaclass=FooMetaclass[int]):  # mypy error, no pyright error
    @classmethod
    def foo(cls) -> None:
        if TYPE_CHECKING:
            reveal_type(cls.value)  # pyright says `int`, mypy says `Any`
        cls.value = 1
        print(cls.value)

Foo.foo()  # prints 1
print(Foo.__orig_class__)  # __main__.FooMetaclass[int]

[erictraut]

Pyright should generate an error in the first example above. The fact that it doesn't is an oversight / bug. FooMetaclass[T] shouldn't be allowed because TypeVar T doesn't have a defined scope in this case.

Pyright does support the use of a specialized metaclass, as in your second example above. I don't see any reason why this shouldn't work from a type-checking standpoint. That said, it's a use case that isn't well tested, so you may run into issues.

There's an interesting case not covered in either of the above two samples:

from typing import TypeVar, Generic

T = TypeVar("T")

class FooMetaclass(type, Generic[T]):
    pass

class Foo(Generic[T], metaclass=FooMetaclass[T]):
    pass

In this variant, TypeVar T does have a well-defined scope (since it's bound to class Foo), so theoretically, this variant could be valid. I'm not sure it's something we'd want to support though.

@DetachHead, is there a real use case here, or are you just playing around with ideas?

[DetachHead]

Pyright does support the use of a specialized metaclass, as in your second example above. I don't see any reason why this shouldn't work from a type-checking standpoint. That said, it's a use case that isn't well tested, so you may run into issues.

yeah i found this issue:

class FooMetaclass(type, Generic[T]):
    metaclass_value: T

class Foo(metaclass=FooMetaclass[int]):
    value: int

# wrong error message:
# Cannot assign member "value" for type "Type[Foo]"
#   Member "value" is unknown
Foo.metaclass_value = "1"

# it should be more like this error instead:
# Cannot assign member "value" for type "Type[Foo]"
#   Expression of type "Literal['1']" cannot be assigned to member "value" of class "Foo"
#     "Literal['1']" is incompatible with "int"
Foo.value = "1"

There's an interesting case not covered in either of the above two samples:

from typing import TypeVar, Generic

T = TypeVar("T")

class FooMetaclass(type, Generic[T]):
    pass

class Foo(Generic[T], metaclass=FooMetaclass[T]):
    pass

In this variant, TypeVar T does have a well-defined scope (since it's bound to class Foo), so theoretically, this variant could be valid. I'm not sure it's something we'd want to support though.

yeah that's what i intended to put as my original example, i just forgot to include Generic[T] in the bases.

however it doesn't seem to work in this case:

class FooMetaclass(type, Generic[T]):
    metaclass_value: T

class Foo(Generic[T], metaclass=FooMetaclass[T]):
    value: T

# should be `int` but is still the unbound generic
reveal_type(Foo[int].metaclass_value)  # Type of "Foo[int].metaclass_value" is "T@Foo"

Foo[int].metaclass_value = 1  # error

Foo[int].value = 1 # no error

@DetachHead, is there a real use case here, or are you just playing around with ideas?

i stubled across this while trying to learn how metaclasses work, but i was basically trying to make a ReifiedGeneric class that looks something like this:

class _ReifiedGenericMetaclass(Generic[T], type):
    __generics__: T
    def __instancecheck__(
        cls,
        obj: object,
    ) -> bool:
        ...

class ReifiedGeneric(Generic[T], metaclass=_ReifiedGenericMetaclass[T]):
    def __class_getitem__(
        cls,
        item: T,
    ) -> type[ReifiedGeneric[T]]:
        ...

[last-partizan]

Those dymanic metaclasses would be really useful for factory-boy. I'm trying to add typing support for it, and i've managed to get if work for some cases, but one unsolved issue remains, here's minimal example based on factory_boy/factory/base.py

from typing import Generic, TypeVar

T = TypeVar("T")

class FactoryMetaClass(type):
    """Factory metaclass for handling ordered declarations."""

    def __call__(cls, **kwargs):
        """Override the default Factory() syntax to call the default strategy.

        Returns an instance of the associated class.
        """
        return cls.create(**kwargs)  # type: ignore

    def __new__(cls, class_name, bases, attrs):
        return super().__new__(cls, class_name, bases, attrs)

class BaseFactory(Generic[T]):
    """Factory base support for sequences, attributes and stubs."""

    @classmethod
    def create(cls, **kwargs) -> T:
        ...

class Factory(BaseFactory[T], metaclass=FactoryMetaClass):
    """Factory base with build and create support.

    This class has the ability to support multiple ORMs by using custom creation
    functions.
    """

# ------------------------------

class Book:
    name: str
    author: str

class BookFactory(Factory[Book]):
    ...

book1 = BookFactory.create()
reveal_type(book1)

book2 = BookFactory()
reveal_type(book2)

Both mypy and pyright correctly handling first case, but i have no idea how to get it to return Book for second case.

mypy basic_example.py
basic_example.py:40:13: note: Revealed type is "basic_example.Book*"
basic_example.py:43:13: note: Revealed type is "basic_example.BookFactory"

@erictraut maybe you have idea how to solve this without generic metaclasses?

[erictraut]

maybe you have idea how to solve this without generic metaclasses?

Yeah, use classes rather than metaclasses. This is the normal way to implement the factory pattern.

from typing import Generic, TypeVar

T = TypeVar("T")

class Factory(Generic[T]):
    def __init__(self, cls: type[T]):
        self.cls = cls

    def create(self, *args, **kwargs):
        return self.cls(*args, **kwargs)

    def __call__(self, *args, **kwargs):
        return self.create(*args, **kwargs)

class Book:
    name: str
    author: str

BookFactory = Factory(Book)

book1 = BookFactory.create()
reveal_type(book1)

book2 = BookFactory()
reveal_type(book2)

[erdnaxeli]

That's not how FactoryBoy works. It works like this:

class BookFactory(factory.Factory):
    ...

book1 = Bookfactory()

You don't instantiate the factory.