@functools.cache destroys the function signature

A function decorated with @functools.cache loses its signature:

@functools.cache
def my_func(foo: int | list) -> str:
    return str(foo)

my_func('completely', 'incorrect', 'arguments')
# Success: no issues found in 1 source file

There is an error if you try to pass an argument that isn't hashable:

my_func([])
# error: Argument 1 to "__call__" of "_lru_cache_wrapper" has
# incompatible type "list[Never]"; expected "Hashable"  [arg-type]

But this completely ruins Intellisense:

I understand that the type system isn't powerful enough to express the correct semantics, which would be this:

@functools.cache
def my_func(foo: int | list) -> str:
    return str(foo)

reveal_type(my_func)  # Callable[[int], str]

That said, I would much rather have functioning Intellisense than a warning about unhashable inputs.

This should now be expressible using ParamSpec.

See also:

I think the conclusion we reached in previous attempts was that ParamSpec is unsufficient to solve this, due to the complex interaction with classmethod.

I think the best bet for getting stuff like this working is a potential future Intersection construct, to extend the Callable that's passed in with the additional attributes, rather than try to implement a descriptor that works transparently in regular functions, methods and classmethods, which does not appear to be possible.

I'm confused. Is there something about functools.cache that makes it particularly difficult to type, or does every function decorator have this problem?

I noticed functools.wraps also isn't annotated correctly:

class Demo1:
    def foo(self) -> None: ...

    @functools.wraps(foo)
    def bar(self): ...

Demo1().bar()  # Argument missing for parameter "self"

But contextlib.contextmanager is correct:

class Demo2:
    @contextlib.contextmanager
    def foo(self) -> Iterator[int]: ...

with Demo2().foo() as output:
    reveal_type(output)  # int

What gives? Why are some decorators correct and some incorrect/impossible?

Also, allow me to repeat/rephrase what I said earlier:

If a perfect solution isn't possible, I'd strongly prefer a solution that gives me useful IntelliSense.

contextlib.contextmanager takes a callable and returns another callable that is in no way distinguished from most other callables. This makes it pretty easy to type:

typeshed/stdlib/contextlib.pyi

Line 76 in c51de8e

    
           def contextmanager(func: Callable[_P, Iterator[_T_co]]) -> Callable[_P, _GeneratorContextManager[_T_co]]: ...

functools.lru_cache and functools.cache, on the other hand, take a callable and return a custom callable object (an instance of functools._lru_cache_wrapper) that also has several special methods present on it (cache_info(), cache_clear() and cache_parameters()):

typeshed/stdlib/functools.pyi

Line 201 in c51de8e

def cache(__user_function: Callable[..., _T]) -> _lru_cache_wrapper[_T]: ...

Unfortunately, this does indeed make it quite a bit harder to type these particular decorators than contextmanager. Ideally we'd just have _lru_cache_wrapper be generic over a ParamSpec as well as the return type, but this isn't as easy as it sounds. There are ambiguities in the spec regarding how ParamSpecs should be captured with decorated methods that mean that this results in false positives on user code. See #6356.

If a perfect solution isn't possible, I'd strongly prefer a solution that gives me useful IntelliSense.

I agree the current situation is deeply unsatisfactory (I think we all do); it's just a question of whether we can find a solution that doesn't cause an unacceptable number of false positives when type checkers are run on user code.

FWIW I think #7771 came close, and I'd be interested in looking at that approach again (though I don't have time right now)

Wait, contextmanager returns a Callable? But it can still be used on methods? So Callable doesn't mean "it's callable", it actually means "it's callable and is a descriptor that returns a bound method"? That's... odd.

@Aran-Fey

@Aran-Fey Callable is special cased inside a class, so foo: Callable[[int], None] behaves (almost) the same as if you had written def foo(self, x: int, /) -> None: ...

The almost concerns calling foo from the class instead of an instance, where the two differ, the former will expect you to pass in an instance of the class, while the latter will not.

But things are yet again different when you use a decorator and return a Callable with the same signature from that, in that case it will actually be equivalent to the def case.

All of this special casing is also part of what makes it so difficult to write a descriptor that transparently works the same way.

I see, thanks! But boy, what a mess.

You can fix this by running:

import functools

def my_function() -> None:
    pass

my_function = functools.wraps(my_function)(functools.cache(my_function))

Instead of:

import functools

@functools.cache
def my_function() -> None:
    pass

While this fixes the incorrect signature, you lose the extra attributes that functools.cache introduces this way, so it's at best trading one problem for another. But it can be a reasonable workaround in code that does not care about the presence of cache_clear/cache_info/cache_parameters.

While this fixes the incorrect signature, you lose the extra attributes that functools.cache introduces this way, so it's at best trading one problem for another.

Those all still work for me...

I assumed this would fall under no type reassignments, but it looks like there's special casing to make this equivalent to decorating the function, but it also looks like your fix doesn't actually preserve the function signature, at least not in mypy or pyright:
https://mypy-play.net/?mypy=latest&python=3.12
https://pyright-play.net/?pythonVersion=3.12&code=JYWwDg9gTgLgBAMwK4DsDGMIQDYGcCwAUEQCYCmCcIAngPrLozAQoAUAlHALQB8cAcizIAuInHFwwAQ1wFihGvVQZmKOAF5EyzDlwA6AO5QpYXK0UMVLdq0s68etFLQALMubp3V7dkShkANzIpbFoYajB3C21vP0Dg0PDIjyVGVT1aWidsUNp2IA

In fact there's no way to have both the extra methods and the correct signature at the same time with the current type stubs, since _lru_cache_wrapper hard codes them to *args: Hashable, **kwargs: Hashable and I postulate there's no way to frame this as a descriptor using ParamSpec either, due to the special-casing for classmethod, so intersection remains our only hope for solving this at this point, unless someone can demonstrate a solution that works in every possible case.

In pyright/pylance you actually have the worst of both worlds with your fix, because you both erase the function signature and _lru_cache_wrapper. So you don't get the correct signature and you also can't access the extra methods.

Perhaps the reason you were fooled into thinking this was a fix for the issue, is due to the fact that Jedi or any LSP built on top of it would show the correct function signature this way. But type checkers still infer the wrong signature.

Also just so we're clear, when I say "there's no way to make this work" I purely am talking about from a type checking perspective. At runtime this will work, but runtime isn't the issue here. Also you can work around this with your own custom descriptor that doesn't need to be able to handle functions, methods and classmethods all at the same time, if you split these into separate descriptors you can make it work right now, but we don't have this luxury with the standard library.

@Aran-Fey

I strongly agree with @Aran-Fey that until this is solved, it would be much better to preserve the function signature at the cost of not checking that parameters are hashable than what is currently there.

If you decorate a function that has non-hashable parameters, you'll have a runtime error the first time you run your code.
With the current state, you can pass wrong parameter types to your function possibly resulting in much subtler/harder to notice bugs.

@Aran-Fey

I strongly agree with @Aran-Fey that until this is solved, it would be much better to preserve the function signature at the cost of not checking that parameters are hashable than what is currently there.

Unfortunately that is not the only cost. The additional cost and the main reason this is annotated the way it currently is, is that we will see false positives anywhere people try to call cache_clear/cache_info/cache_parameters. This will generate a lot of noise and confusion. Typeshed generally prefers false negatives over false positives.

It would be easy to preserve the function signature if we didn't need to make sure those methods are available.

@Daverball

In pyright/pylance you actually have the worst of both worlds with your fix, because you both erase the function signature and _lru_cache_wrapper. So you don't get the correct signature and you also can't access the extra methods.

Perhaps the reason you were fooled into thinking this was a fix for the issue, is due to the fact that Jedi or any LSP built on top of it would show the correct function signature this way. But type checkers still infer the wrong signature.

@Daverball My issue was that functools.cache disabled Vscode's type and docstring hints, which my fix reenables. I use Pylance not Jedi and have not experienced any issues. However, it seems that this would not suffice for your use cases and regardless I would like to see this fixed without the need for workarounds.

I'm aware there's a (slightly stale) discussion about changing ParamSpec to support this but I got a version of cache that works as desired in the examples I came up with. Mypy doesn't error with an unhashable argument for both my version and the builtin version but seems to give correct errors in all the other examples I had. It'll have false positives if someone has a method or classmethod without self or cls as the first argument, or vice versa. I'm ok with that. To get lru_cache working the int | None overload needs to return a protocol that has 3 __call__ overloads that are basically the same as the cache overloads below.

from __future__ import annotations

from typing import overload, Any, Protocol, Self, TYPE_CHECKING
from collections.abc import Callable
from functools import cache as _cache
from typing import overload, Any, Concatenate, Hashable, override, assert_type


class ProtoMethod[T, **Pc, Rc](Protocol):
    def __call__(__self, self: T,  *args: Pc.args, **kwds: Pc.kwargs) -> Rc:
        ...


class MethodDescriptor[T, **P, R](Protocol):
    def __call__(__self, self: T, *args: P.args, **kwds: P.kwargs) -> R:
        ...
    @overload
    def __get__(self, instance: None, owner: type[T]) -> Self:
        ...

    @overload
    def __get__(self, instance: T, owner: type[T]) -> FunctionDescriptor[P, R]:
        ...

    def __get__(self, instance: T | None, owner: type[T]) -> Any:
        ...
    def cache_info(self) -> Any: ...
    def cache_clear(self) -> None: ...
    def cache_parameters(self) -> Any: ...


class ProtoClassmethod[T, **Pc, Rc](Protocol):
    def __call__(__self, cls: type[T],  *args: Pc.args, **kwds: Pc.kwargs) -> Rc:
        ...


class ClassmethodDescriptor[T, **P, R](Protocol):
    def __call__(self, cls: type[T], *args: P.args, **kwds: P.kwargs) -> R:
        ...
    @overload
    def __get__(self, instance: None, owner: type[T]) -> FunctionDescriptor[P, R]:
        ...
    @overload
    def __get__(self, instance: T, owner: type[T]) -> FunctionDescriptor[P, R]:
        ...
    def __get__(self, instance: T | None, owner: type[T]) -> FunctionDescriptor[P, R]:
        ...
    def cache_info(self) -> Any: ...
    def cache_clear(self) -> None: ...
    def cache_parameters(self) -> Any: ...


class ProtoFunction[**Pc, Rc](Protocol):
    def __call__(
            __self,
            *args: Pc.args,
            **kwds: Pc.kwargs
    ) -> Rc:
        ...

class FunctionDescriptor[**Pc, Rc](Protocol):
    def __call__(
            __self,
            *args: Pc.args,
            **kwds: Pc.kwargs
    ) -> Rc:
        ...
    def cache_info(self) -> Any: ...
    def cache_clear(self) -> None: ...
    def cache_parameters(self) -> Any: ...

class FnHashable[R](Protocol):
    def __call__(self, *args: Hashable, **kwds: Hashable) -> R:
        ...
    def cache_info(self) -> Any: ...
    def cache_clear(self) -> None: ...
    def cache_parameters(self) -> Any: ...

class ClsHashable[T, **P, R](Protocol):
    def __call__(__self, cls: type[T], *args: Hashable, **kwds: Hashable) -> R:
        ...
    def cache_info(self) -> Any: ...
    def cache_clear(self) -> None: ...
    def cache_parameters(self) -> Any: ...
    @overload
    def __get__(self, instance: None, owner: type[T]) -> FunctionDescriptor[P, R]:
        ...
    @overload
    def __get__(self, instance: T, owner: type[T]) -> FunctionDescriptor[P, R]:
        ...
    def __get__(self, instance: T | None, owner: type[T]) -> FunctionDescriptor[P, R]:
        ...


@overload
def cache[T, **P, R](fn: ProtoMethod[T, P, R]) -> MethodDescriptor[T, P, R] | FnHashable[R]:
    ...


@overload
def cache[T, **P, R](fn: ProtoClassmethod[T, P, R]) -> ClassmethodDescriptor[T, P, R] | ClsHashable[T, P, R]:
    ...


@overload
def cache[**P, R](fn: ProtoFunction[P, R]) -> FunctionDescriptor[P, R] | FnHashable[R]:
    ...


def cache(fn: Any) -> Any:
    return _cache(fn)


class CFnCls:

    @cache
    def fn(self, arg: int) -> int:
        print("method fn called")
        return arg

    @classmethod
    @cache
    def cls_fn(cls, arg: int) -> int:
        print("class fn called")
        return arg

    @staticmethod
    @cache
    def st_fn(arg: int) -> int:
        print("static fn called")
        return arg

cfn_inst = CFnCls()
cfn_inst.fn(1)
CFnCls.st_fn(1)
CFnCls.cls_fn(1)
cfn_inst.fn(1)
CFnCls.st_fn(1)
CFnCls.cls_fn(1)
assert_type(cfn_inst.fn(1), int)
assert_type(CFnCls.st_fn(1), int)
assert_type(CFnCls.cls_fn(1), int)
CFnCls().fn.cache_clear()
CFnCls.fn.cache_clear()
CFnCls.st_fn.cache_clear()
CFnCls.cls_fn.cache_clear()
if TYPE_CHECKING:
    CFnCls().fn(1, 1) # type error - correct
    CFnCls.st_fn(1, 1) # type error - correct
    CFnCls.cls_fn(1, 1) # type error - correct

@cache
def fn(arg: int) -> int:
    return arg

fn(1)
assert_type(fn(1), int)
if TYPE_CHECKING:
    fn(1, 2) # type error - correct
fn.cache_clear()


@overload
@cache
def fn_overload(arg: int) -> int:
    ...
@overload
@cache
def fn_overload(arg: str) -> str:
    ...
@cache
def fn_overload(arg: int | str) -> int | str:
    return arg

fn_overload(1)
fn_overload("1")
# behaves same as builtin cache, merges overloads
assert_type(fn_overload(1), int | str)
assert_type(fn_overload("1"), int | str)
fn_overload.cache_clear()
if TYPE_CHECKING:
    fn_overload({1,2}) # type error - correct

class Unhashable:
    @override
    def __eq__(self, value: object) -> bool:
        return False

@cache
def no_cache(arg: Unhashable, arg2: int) -> None:
    pass

@_cache
def bi_no_cache(arg: Unhashable, arg2: int) -> None:
    pass



if TYPE_CHECKING:
    # type error in pyright - correct
    # no type error in mypy?
    no_cache(Unhashable(), 2)
    bi_no_cache(Unhashable(), 2)

When mypy does corrrectly error it gives a slightly strange error, eg:

main.py:164: error: Too many arguments for "__call__" of "FunctionDescriptor"  [call-arg]
main.py:164: note: Looks like the first special argument in a method is not named "self", "cls", or "mcs", maybe it is missing?

But otherwise seems to work, if this is of interest to others.

It still doesn't preserve the docstring in intellisense but it avoids any obvious type errors at least.

My intellisense autocomplete suggests kwds as a function parameter. Not sure why. But I wouldn't call that a blocker here, considering the status quo. The built in cache version results in it suggesting kwargs similarily. But with my version it also suggests the correct arguments (and errors if kwds is used even though it's suggested).

If someone can find an example where this doesn't work lemme know. Or I can rewrite it using the older TypeVar syntax if a PR would be of interest.

theory: change _get_chords_from_notes notes parameter type to Iterobla

Use ParamSpec _lru_cache_wrapper

I am going to close this as "inexpressible" for now. There have been multiple tries to fix this, but none have been successful. It seems that we need changes to Python's type system before we can solve this.

theory: change _get_chords_from_notes notes parameter type to Iterobla

Add tests for @functools.cache (#14768)

The types shipped with cachetools as of tkem/cachetools#398 seem to solve this situation, so maybe this is a good time to revisit this?

In all honesty I didn't take the time to dig into how functools.cache is annotated, but it does basically the same thing as cacheutils.func.lru_cache, so if it's possible to express one, then it should now also be possible to express the other. Maybe.

(I also didn't deeply think about whether the cacheutils one is fully correct, but it seems to work OK on some simple tests)

An exploratory PR (with tests for the corner cases described here and in the related issues and PRs) would certainly be welcome.

Navigation Menu

Repository navigation

@functools.cache destroys the function signature #11280

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

@functools.cache destroys the function signature #11280

Description

Activity

srittau commented on Jan 16, 2024

AlexWaygood commented on Jan 16, 2024

Daverball commented on Jan 18, 2024

Aran-Fey commented on Jan 18, 2024

AlexWaygood commented on Jan 18, 2024

Aran-Fey commented on Jan 18, 2024

Daverball commented on Jan 18, 2024

Aran-Fey commented on Jan 18, 2024

umarbutler commented on Mar 11, 2024

Daverball commented on Mar 11, 2024

umarbutler commented on Mar 11, 2024

Daverball commented on Mar 11, 2024

ldorigo commented on Mar 13, 2024

Daverball commented on Mar 13, 2024

umarbutler commented on Mar 17, 2024

alwaysmpe commented on Nov 3, 2024

srittau commented on Sep 23, 2025

zoltan-axle commented on May 26, 2026

srittau commented on May 26, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions