`GenericModel` calling `get_args()` which is not hashable

[samuelcolvin]

See #4482 (comment)

From @pepastach

We recently bumped Pydantic from 1.10.1 to 1.10.2 and our tests started failing. After some lengthy investigation I found this change to be the root cause. I'd like to ask you for your opinions.

We have a Pydantic model with a field such as

f: MyGenericModel[Callable[[Task, Dict[str, Any]], Iterable[Result]]] = Field(...)

Just importing Python module with this code gives me:

cls = <class 'foo.MyGenericModel'>
params = typing.Callable[[foo.Task, typing.Dict[str, typing.Any]], typing.Iterable[foo.Result]]

    def __class_getitem__(cls: Type[GenericModelT], params: Union[Type[Any], Tuple[Type[Any], ...]]) -> Type[Any]:
        """Instantiates a new class from a generic class `cls` and type variables `params`.
    
        :param params: Tuple of types the class . Given a generic class
            `Model` with 2 type variables and a concrete model `Model[str, int]`,
            the value `(str, int)` would be passed to `params`.
        :return: New model class inheriting from `cls` with instantiated
            types described by `params`. If no parameters are given, `cls` is
            returned as is.
    
        """
    
        def _cache_key(_params: Any) -> Tuple[Type[GenericModelT], Any, Tuple[Any, ...]]:
            return cls, _params, get_args(_params)
    
>       cached = _generic_types_cache.get(_cache_key(params))
E       TypeError: unhashable type: 'list'

The problem seems to be that the function GenericModel._cache_key() now calls get_args() which in turns calls Python's typing.get_args() -> and this function returns a tuple with a list in it. And that makes it unhashable.

def get_args(tp):
    """Get type arguments with all substitutions performed.

    For unions, basic simplifications used by Union constructor are performed.
    Examples::
        get_args(Dict[str, int]) == (str, int)
        get_args(int) == ()
        get_args(Union[int, Union[T, int], str][int]) == (int, str)
        get_args(Union[int, Tuple[T, int]][str]) == (int, Tuple[str, int])
        get_args(Callable[[], T][int]) == ([], int)
    """
    if isinstance(tp, _GenericAlias) and not tp._special:
        res = tp.__args__
        if get_origin(tp) is collections.abc.Callable and res[0] is not Ellipsis:
            res = (list(res[:-1]), res[-1])             # <======== this list is a problem
        return res
    return ()

If I skip the callable typing and only define the field as

f: MyGenericModel[Callable] = Field(...)

then it runs okay.

Thanks in advance!

---

Reply from @sveinugu

@pepastach Hmm... Seems we did not think of that. Surprised that typing uses lists and not tuples for Callable arguments, but there is probably a good reason for this (or it might be just that it would be too much a nuisance to switch between brackets and parentheses...). In any case, there might be other examples of unhashable return values from get_params, so a simple solution could be to just catch TypeError and in that case default to the previous _cache_key generation.

I am new as contributor to pydantic, but I assume it would be cleanest if you could create another issue for this.

[samuelcolvin]

Fixed by #4653