inspect not capturing type annotations created by __class_getitem__

[rhettinger]

BPO	45438
Nosy	@gvanrossum, @rhettinger, @ambv, @serhiy-storchaka, @ilevkivskyi, @miss-islington, @Fidget-Spinner, @martinitus
PRs	bpo-45438: format of inspect.Signature with generic builtins #29212 [3.10] bpo-45438: format of inspect.Signature with generic builtins (GH-29212) #29253 [3.9] bpo-45438: format of inspect.Signature with generic builtins (GH-29212) #29254

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2021-10-29.08:43:15.006>
created_at = <Date 2021-10-11.22:05:54.548>
labels = ['type-bug', 'library', '3.9', '3.10', '3.11']
title = 'inspect not capturing type annotations created by __class_getitem__'
updated_at = <Date 2021-10-29.08:43:15.004>
user = 'https://github.com/rhettinger'

bugs.python.org fields:

activity = <Date 2021-10-29.08:43:15.004>
actor = 'serhiy.storchaka'
assignee = 'none'
closed = True
closed_date = <Date 2021-10-29.08:43:15.006>
closer = 'serhiy.storchaka'
components = ['Library (Lib)']
creation = <Date 2021-10-11.22:05:54.548>
creator = 'rhettinger'
dependencies = []
files = []
hgrepos = []
issue_num = 45438
keywords = ['patch']
message_count = 21.0
messages = ['403692', '403697', '403759', '403760', '403826', '404961', '404966', '404967', '404970', '404974', '404976', '404993', '405034', '405128', '405132', '405133', '405138', '405161', '405163', '405283', '405292']
nosy_count = 8.0
nosy_names = ['gvanrossum', 'rhettinger', 'lukasz.langa', 'serhiy.storchaka', 'levkivskyi', 'miss-islington', 'kj', 'martinitus']
pr_nums = ['29212', '29253', '29254']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue45438'
versions = ['Python 3.9', 'Python 3.10', 'Python 3.11']

[rhettinger]

In the example below, __annotations__ is correct but not the corresponding Signature object.

-----------------------------------------------------------------------

from typing import List

def f(s: List[float]) -> None: pass

def g(s: list[float]) -> None: pass

>>> inspect.signature(f)
<Signature (s: List[float]) -> None>

>>> inspect.signature(g)
<Signature (s: list) -> None>

g.__annotations__
{'s': list[float], 'return': None}

[gvanrossum]

Raymond, the bug must be in the Python code in inspect.py. Could you dig a little deeper there? I don't know much about it. Specifically I think the problem may just be in the repr() of class Parameter:

>>> inspect.signature(g).parameters['s']             
<Parameter "s: list">
>>> inspect.signature(g).parameters['s'].annotation
list[float]
>>>

[rhettinger]

It looks like the error is in inspect.formatannotation().

For instances of type, that function incorrectly returns only annotation.__qualname__.

Instead, it should return repr(annotation) which would include the args.

=================================================================

def formatannotation(annotation, base_module=None):
    if getattr(annotation, '__module__', None) == 'typing':
        return repr(annotation).replace('typing.', '')
    if isinstance(annotation, type):                # <== Erroneous case
        if annotation.__module__ in ('builtins', base_module):
            return annotation.__qualname__
        return annotation.__module__+'.'+annotation.__qualname__
    return repr(annotation)

[gvanrossum]

Sure. Waiting for your PR. :-)

[serhiy-storchaka]

The cause is that isinstance(list[int], type) returns True. It can cause bugs in other parts of the code which test for instance of type. For example:

>>> types.resolve_bases((typing.List[int],))
(<class 'list'>, <class 'typing.Generic'>)
>>> types.resolve_bases((list[int],))
(list[int],)

>>> types.prepare_class('A', (int,), {'metaclass': typing.Type[int]})
(typing.Type[int], {}, {})
>>> types.prepare_class('A', (int,), {'metaclass': type[int]})
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/serhiy/py/cpython/Lib/types.py", line 125, in prepare_class
    meta = _calculate_meta(meta, bases)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/serhiy/py/cpython/Lib/types.py", line 139, in _calculate_meta
    if issubclass(base_meta, winner):
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: issubclass() argument 2 cannot be a parameterized generic

>>> @functools.singledispatch
... def g(a): pass
... 
>>> @g.register
... def g2(a: typing.List[int]): pass
... 
Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
  File "/home/serhiy/py/cpython/Lib/functools.py", line 863, in register
    raise TypeError(
    ^^^^^^^^^^^^^^^^
TypeError: Invalid annotation for 'a'. typing.List[int] is not a class.
>>> @g.register(list[int])
... def g2(a): pass
... 
>>> @g.register
... def g3(a: typing.List[int]): pass
... 
Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
  File "/home/serhiy/py/cpython/Lib/functools.py", line 863, in register
    raise TypeError(
    ^^^^^^^^^^^^^^^^
TypeError: Invalid annotation for 'a'. typing.List[int] is not a class.
>>> @g.register
... def g3(a: list[int]): pass
... 

And many other examples, too many to show here.

Was it mistake to make isinstance(list[int], type) returning True?

[martinitus]

Can confirm for 3.9.7 as well.

[martinitus]

I just created a PR and signed the contributor agreement. Waiting for it to update :-) Comments welcome!

[serhiy-storchaka]

There are two ways to fix the larger issue.

1. Make issubclass(types.GenericAlias, type) returning True, and also make isinstance(typing.List[int], type) returning True and issubclass(typing._GenericAlias, type) returning True, and analyze every place in the interpreter and the stdlib which calls isinstance(..., type) or issubclass(..., type) and ensure that they work with types.GenericAlias and typing._GenericAlias and their instances as well as with ordinary types, fix them and add tests for them. And perhaps do the same for types.UnionType, typing._UnionGenericAlias, typing.TypeVar, typing.NewType, etc, etc.

2. Make isinstance(list[int], type) returning False. It would be nice to ad also tests for the same cases as in option 1, but it is not so urgent, and I expect that in most cases the current behavior which matches the status quo is expected.

I tried to implement option 1, but it is just too much places, so it would take a large amount of time which I do not have right now. First than invest my time in this I want to make sure which option is desirable in long term.

Guido, Ivan, Ken Jin, what would you say?

[ilevkivskyi]

Was it mistake to make isinstance(list[int], type) returning True?

What was the motivation for this? At first glance returning True looks wrong.

[gvanrossum]

I really don't recall if we even seriously considered what isinstance(list[int], type) should return. PEP-585 doesn't mention it. I presume it falls out of the way it's being tested and the way list[int] passes most attribute requests on to the origin (i.e., to list).

Given that this has returned True for two releases now I'm very reluctant to changing this now. So our best option is pursuing Serhiy's option 1, at least partially. The fix for the OP should be simple enough, right?

[serhiy-storchaka]

Two years is not so long for a bug. We fixed 8-year and 12-year bugs.

The issue is that this feature is internally inconsistent (isinstance() is not consistent with issubclass()), the C implementation of list[int] is not consistent with the Python implementation of List[int], and a lot of code was not tested with this feature, so we perhaps have a lot of bugs in the stdlib and in the third-party libraries. Are we going to backport changes for making code working with GenericAlias as a type to 3.9 and 3.10? If not, we will actually add new features in 3.11 and keep 3.9 and 3.10 broken. If yes, these changes can have larger effect than just making isinstance(list[int], type) returning False, and they can break more code.

Note also that isinstance(List[int], type) was True before 3.7, and we intentionally made it False in 3.7 (it was required significant rewriting of the typing module and introducing __mro_entries__). Do we want to revert this decision?