Generic class atributes aren't fully instantiated

T = TypeVar('T')
class A(Generic[T]):
    def f(self, x: T) -> None: ...
class B(Generic[T]):
    x = None  # type: A[T]

B[int].x.f(0)  # E: Argument 1 to "f" of "A" has incompatible type "int"; expected "T"

Discovered while reviewing #2873.

@ilevkivskyi

I don't think that the current semantics actually support that. Consider this:

from typing import TypeVar, Generic

T = TypeVar('T')

class A(Generic[T]):
    x = None  # type: T

A[int].x = 1
A[str].x = 'a'
print(A[int].x)  # 1
print(A[str].x)  # a

a = A()  # type: A[int]
print(a.x)  # None
print(type(a).x)  # None
print(A[int]().x)  # None

The value of x depends on how we access it, which seems pretty bad. This looks like a problem in typing -- @ilevkivskyi do you agree?

Mypy could work around this even with the current semantics by rejecting code like A[int].x -- the only use supported for A[int] outside a type context would be A[int](...) where we don't have this problem.

Also, attributes that use class type variables should only be accessible through an instance, not the class. So A().x would be okay but A.x would be rejected, since the value of the type variable is conceptually bound to an instance, not the type object. There's effectively only a single type object A (though the implementation sometimes behaves like this is not the case).

@ilevkivskyi

The value of x depends on how we access it, which seems pretty bad. This looks like a problem in typing -- @ilevkivskyi do you agree?

I agree this looks strange. One possible solution (first that comes to my mind) is to tweak descriptor __dict__ of GenericMeta so that all assignments and look-ups on subscripted classes (A[int] etc ) will be transferred to the original class object A. But I am not comfortable making this change without discussion with Guido, since this is a change in "user facing API".

Also, attributes that use class type variables should only be accessible through an instance, not the class. So A().x would be okay but A.x would be rejected, since the value of the type variable is conceptually bound to an instance, not the type object. There's effectively only a single type object A (though the implementation sometimes behaves like this is not the case).

I think I agree. Effectively this means that in the proposed implementation of ClassVar #2873 we need to prohibit generic class variables.

But what should we do with look-up of instance variables (those not wrapped in ClassVar) on classes in case if these variables are generic? Just leave it as it is, or set all type variables to Any on such look-up, or something else? (I think just prohibiting this always is not a good solution).

But what should we do with look-up of instance variables (those not wrapped in ClassVar) on classes in case if these variables are generic? Just leave it as it is, or set all type variables to Any on such look-up, or something else? (I think just prohibiting this always is not a good solution).

Prohibiting it seems to be an easy and safe solution. If we give it any normal type, this would make it possible to assign to the default:

A.x = something

This would break x used as an instance attribute, since the value can't be compatible with all possible type variable values (unless the value is None and we aren't using --strict-optional, but that doesn't seem very useful).

However, if we make it a read-only attribute when accessed through the class, we could sometimes use the upper bound type of the type variable, but that wouldn't work for types list List[T] that are invariant -- we'd need an existential type which we don't have.

I can't think of any use reasonable case where it would be important to access the default value of a instance variable that has a generic type variables in it's type through a type object. If you can come up with an example, it could be easier to see what we should do, other than rejecting it.

@JukkaL

@JukkaL

I can't think of any use reasonable case where it would be important to access the default value of a instance variable that has a generic type variables in it's type through a type object. If you can come up with an example, it could be easier to see what we should do, other than rejecting it.

My main concern is that the user who are not aware of instance/class variable distinction will be confused if their perfectly working code is rejected. The problem is finding a good error message.
The best I can think now is something like:

class C(Generic[T]):
    x: T
y = C.x # E: Access to generic instance variables via class is ambiguous

(Btw the runtime issue in typing is now fixed)

Implement ClassVar semantics (fixes #2771) (#2873)

Clarify generic class attribute access

Update generic class attribute and type variable usage (#239)

Navigation Menu

Repository navigation

Generic class atributes aren't fully instantiated #2878

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Generic class atributes aren't fully instantiated #2878

Description

Activity

JukkaL commented on Feb 22, 2017

ilevkivskyi commented on Feb 22, 2017

JukkaL commented on Feb 24, 2017

ilevkivskyi commented on Feb 24, 2017

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions