Unexhaustive pattern match on dataclass field union

[ducdetronquito]

Hello !

Here is a minimal case of pattern matching on a dataclass field that is the union of two types:

from dataclasses import dataclass

@dataclass
class Classic:
    pass


@dataclass
class Admin:
    pass


@dataclass
class User:
    name: str
    role: Classic | Admin


some_user = User("pluto", role=Admin())

match some_user:
    case User(_, role=Classic()):
        pass
    case User(_, role=Admin()):
        pass

I would expect that the pattern match is exhaustive here but pyright (1.1.386) disagree:

error: cases within match statement do not exhaustively handle all values
    Unhandled type: "User"

Pyright only considers that the pattern match is exhaustive when I add the following branch:

match some_user:
    case User(_, role=Classic()):
        pass
    case User(_, role=Admin()):
        pass
    case User():
        pass

Do you have any idea why this 3rd branch is required ?

Thanks in advance, and have a nice day 🌴

EDIT: Here is a link to the same code on pyright's playground

[erictraut]

Exhaustion detection uses type narrowing to determine whether you have handled all possible types. In this case, the subject type starts as User. The subsequent case statements do not narrow the type further. The type of the subject remains User after the first and the second. There is no notion in the Python type system of the type User where the "role" attribute refers to a "Classic" object.

If you want static analysis exhaustion to work, you'll need to change your code somewhat.

Code sample in pyright playground

# pyright: strict

from dataclasses import dataclass

@dataclass
class Classic:
    pass

@dataclass
class Admin:
    pass

@dataclass
class User[T: Classic | Admin]:
    name: str
    role: T

def func(some_user: User[Classic] | User[Admin]):
    match some_user:
        case User(_, role=Classic()):
            pass
        case User(_, role=Admin()):
            pass

[ducdetronquito]

Thanks a lot for your detailed answer :)