Description
Description:
The type stub for functools.update_wrapper()
does not permit a descriptor for the wrapper
(first) argument, a legitimate case arising when defining a descriptor-type decorator (like @property
is).
https://github.com/python/typeshed/blob/main/stdlib/functools.pyi#L83
@erictraut While perhaps not exactly a regression, this used to work up until this commit of yours here (last week): 9c4bfd5#diff-02c51ed45b741b6765c365fb5d6578c586510bbbd57b10a1dba10aeba37ae157L70-R84. I say not exactly a regression because the prior constraint introduced here was perhaps overly loose: https://github.com/python/typeshed/pull/4627/files
Example
As an example, the following (pyright) typing error is detected in the lazyproperty
descriptor-type decorator defined here: https://gist.github.com/scanny/255f16339b74b2b59e896d67bd459d6b
lazyproperty.py|65 col 34-38 error| Argument of type "Self@lazyproperty[T@lazyproperty]"
cannot be assigned to parameter "wrapper"
of type "(**_PWrapper@update_wrapper) -> _RWapper@update_wrapper"
in function "update_wrapper"
Type "Self@lazyproperty[T@lazyproperty]" cannot be assigned to
type "(**_PWrapper@update_wrapper) -> _RWapper@update_wrapper"
To Reproduce
The gist referenced above has no dependencies other than the stdlib so can be type-checked directly (using pyright v1.1.296):
$ pyright lazyproperty.py
Analysis
The type definition at work constrains wrapper
to be a Callable[_PWrapper, _RWapper]
. I believe it also needs to permit Descriptor
, although I'm not completely sure what that type definition is. My best guess is something like this:
class HasDunderGet(Protocol):
def __get__(self, obj: Any, type: type = None) -> Any: ...
class HasDunderSet(Protocol):
def __set__(self, obj: Any, value: Any) -> None: ...
class HasDunderDelete(Protocol):
def __delete__(self, obj: Any) -> None: ...
Descriptor = Union[HasDunderGet, HasDunderSet, HasDunderDelete]
Expected behavior
As far as I can tell, this is a legitimate usage of functools.update_wrapper()
and decorators that produce a descriptor (like @property
does) are also legitimate Python. It has certainly worked for a long time and also even typechecked with pyright previously (v1.1.273), maybe six months ago. So I would expect it to be accepted as written.