Recognizing wrapped classes as supporting a Protocol

[a-recknagel]

from dataclasses import dataclass
from typing import Dict

from typing_extensions import Protocol

@dataclass
class A:
    pass

class IsDataclass(Protocol):
    __dataclass_fields__: Dict

def dataclass_only(x: IsDataclass):
    pass

dataclass_only(A())

When running mypy on this code, I'd expect no warning, but I get

test.py:16: error: Argument 1 to "dataclass_only" has incompatible type "A"; expected "IsDataclass".

While checking for __dataclass_fields__ might not look like the best thing to go for, it's what dataclasses.is_dataclass checks for as well, so this seems like the most reliable condition right now.

I'd assume it happens because the dataclass-wrapper adds this attribute, and mypy only checks the class definition itself. Can this be fixed?

[ilevkivskyi]

This can be supported by just adding __dataclass_fields__ Var to the symbol table in the plugin. This is quite exotic use case, so setting priority to low. On the other hand, if you have time, PRs are welcome, this shouldn't be hard to fix.

[a-recknagel]

I'm not 100% sure what you mean with adding the symbol to the plugin table, do you mean this?

@dataclass
class A:
    __dataclass_fields__: Dict

Because that does fix it without side effects, since the dataclass wrapper just overrides it. Then again, once I start adding fields to my dataclasses only to fulfill a protocol I might as well inherit from a classic no-op interface class.

Low prio is what I think as well, I'll look if I can manage to write a proper fix.

[ilevkivskyi]

I'm not 100% sure what you mean with adding the symbol to the plugin table, do you mean this?

You misread what I wrote, I meant fixing this in mypy plugin for dataclasses (dataclasses in mypy are supported by a bundled/builtin plugin).

[a-recknagel]

I see, thanks for the info =)

[viniciusd]

So we are talking about mypy/plugins/dataclasses.py here, right @ilevkivskyi ?

I would like to have that fixed because I see no other reliable way of type hinting data classes.
Have never touched mypy source code, but I can give it a try.

[ilevkivskyi]

So we are talking about mypy/plugins/dataclasses.py here

Yes.

Have never touched mypy source code, but I can give it a try.

Please go ahead! If you feel stuck, feel free to ask here or on Gitter.

[viniciusd]

I will certainly have some questions, thanks :)

For now, in order to leave it commented here, I managed to track the type checking to the subtypes submodule. Today, I am simply trying to get the gist of what is mypy's flow.

The function call that isn't returning what we expect in order to validate the protocol is at subtypes.py(481):

subtype = find_member(member, left, left)

Where member is the attribute we are checking (i.e.,__dataclass_fields__), and left is A class.

find_member consists of:

	    def get_method(self, name: str) -> Optional[FuncBase]:
	        for cls in self.mro:
	            if name in cls.names:
	                node = cls.names[name].node
	                if isinstance(node, FuncBase):
	                    return node
	                else:
	                    return None

Where:

self = <TypeInfo foo.A>
self.mro = [<TypeInfo foo.A>, <TypeInfo builtins.object>]
name = '__dataclass_fields__'

Given self.names is an empty SymbolTable, name is never in cls.names. Tweaking with dataclasses' symbol table should make __dataclass_fields__ visible here, then making A compliant with the protocol IsDataclass.

Thoughts here so far?

[ilevkivskyi]

Tweaking with dataclasses' symbol table should make __dataclass_fields__ visible here, then making A compliant with the protocol IsDataclass.

Yes, we already add __init__ and other things to the symbol table in plugins/dataclasses.py, just add a Var there with a type like Dict[str, Any].

[viniciusd]

It seems I was following the wrong thread there 😅 Ended up at the semanal.analyze_class class, which analyzes the definitions (defn.defs.body). Still, that was very instructive :)

Gonna go for the plugin now, thanks

[viniciusd]

How does Var relate to TypeInfo? Creating a Var via Var(random_name, Dict[Any, Any] initializes it with VAR_NO_INFO, for example.
I see it is assigned at some places, like semanal.make_name_lvalue_var

[ilevkivskyi]

How does Var relate to TypeInfo?

For instance/class attributes, Var.info should be set to the class where it was defined (can be get as ClassDefContext.cls.info).

[viniciusd]

That makes sense, thanks!

Aren't we supposed to have some sort of Dict node-type at mypy/types.py? I see TypedDict there, which doesn't seem to be the right fit here