This issue tracker has been migrated to GitHub, and is currently read-only.
For more information, see the GitHub FAQs in the Python's Developer Guide.

Title: __future__.annotations breaks inspect.signature()
Type: behavior Stage:
Components: Library (Lib) Versions: Python 3.9, Python 3.8
Status: open Resolution:
Dependencies: Superseder:
Assigned To: Nosy List: 1ace, JelleZijlstra, The Compiler, eric.smith, kj, terry.reedy
Priority: normal Keywords:

Created on 2021-03-01 16:02 by 1ace, last changed 2022-04-11 14:59 by admin.

Messages (9)
msg387875 - (view) Author: Eric Engestrom (1ace) Date: 2021-03-01 16:01
We have a pytest that boils down to the following:

#from __future__ import annotations
from inspect import Parameter, signature

def foo(x: str) -> str:
    return x + x

def test_foo():
    expected = (
        Parameter("x", Parameter.POSITIONAL_OR_KEYWORD, annotation=str),

    actual = tuple(signature(foo).parameters.values())

    assert expected == actual

(execute with `pip install pytest && pytest -vv`)

I tried importing 3.10 annotations (so that we can get rid of quotes around the class containing `foo()`, which is omitted here because it isn't necessary to reproduce the bug), but doing so changes the output of `inspect.signature()` but not the output `inspect.Parameter()`, causing a mismatch between the two that breaks the test.

The above passes on 3.7.9, 3.8.7 & 3.9.1, and if I uncomment the first line, it fails on those same versions.
As can be expected, the annotations import is a no-op on 3.10.0a5 and the test passes either way.

I expect `inspect` might have not been correctly updated to support postponed annotations, but I haven't looked at the implementation (I'm not familiar with the CPython codebase at all) so it's just a guess.
msg387937 - (view) Author: Eric V. Smith (eric.smith) * (Python committer) Date: 2021-03-02 16:10
Can you show the values of “expected” and “actual” for both the failures and successes?
msg388018 - (view) Author: Eric Engestrom (1ace) Date: 2021-03-03 13:41
Sure thing.

Putting the reproducing code in `` and running `docker run --rm -it -v $PWD:/code python sh -c 'pip install pytest && pytest -vvv /code/'` yields:

Collecting pytest
  Downloading pytest-6.2.2-py3-none-any.whl (280 kB)
     |████████████████████████████████| 280 kB 516 kB/s
Collecting toml
  Downloading toml-0.10.2-py2.py3-none-any.whl (16 kB)
Collecting py>=1.8.2
  Downloading py-1.10.0-py2.py3-none-any.whl (97 kB)
     |████████████████████████████████| 97 kB 1.6 MB/s
Collecting attrs>=19.2.0
  Downloading attrs-20.3.0-py2.py3-none-any.whl (49 kB)
     |████████████████████████████████| 49 kB 1.8 MB/s
Collecting iniconfig
  Downloading iniconfig-1.1.1-py2.py3-none-any.whl (5.0 kB)
Collecting packaging
  Downloading packaging-20.9-py2.py3-none-any.whl (40 kB)
     |████████████████████████████████| 40 kB 833 kB/s
Collecting pluggy<1.0.0a1,>=0.12
  Downloading pluggy-0.13.1-py2.py3-none-any.whl (18 kB)
Collecting pyparsing>=2.0.2
  Downloading pyparsing-2.4.7-py2.py3-none-any.whl (67 kB)
     |████████████████████████████████| 67 kB 1.5 MB/s
Installing collected packages: pyparsing, toml, py, pluggy, packaging, iniconfig, attrs, pytest
Successfully installed attrs-20.3.0 iniconfig-1.1.1 packaging-20.9 pluggy-0.13.1 py-1.10.0 pyparsing-2.4.7 pytest-6.2.2 toml-0.10.2
============================= test session starts ==============================
platform linux -- Python 3.9.2, pytest-6.2.2, py-1.10.0, pluggy-0.13.1 -- /usr/local/bin/python
cachedir: .pytest_cache
rootdir: /code
collected 1 item

code/ FAILED                                        [100%]

=================================== FAILURES ===================================
___________________________________ test_foo ___________________________________

    def test_foo():
        expected = (
            Parameter("x", Parameter.POSITIONAL_OR_KEYWORD, annotation=str),

        actual = tuple(signature(foo).parameters.values())

>       assert expected == actual
E       assert (<Parameter "x: str">,) == (<Parameter "x: 'str'">,)
E         At index 0 diff: <Parameter "x: str"> != <Parameter "x: 'str'">
E         Full diff:
E         - (<Parameter "x: 'str'">,)
E         ?                 -   -
E         + (<Parameter "x: str">,)

code/ AssertionError
=========================== short test summary info ============================
FAILED code/ - assert (<Parameter "x: str">,) == (<Param...
============================== 1 failed in 0.02s ===============================
msg388023 - (view) Author: Ken Jin (kj) * (Python committer) Date: 2021-03-03 14:41
@eric.smith, here's a summarized version. I hope it helps:

def foo(x: str) -> str: ...

In Python 3.7 - 3.9 with from __future__ import annotations, inspect.signature sees foo's parameter as:

<Parameter "x: 'str'">

Without the future import (and also in Python 3.10):

<Parameter "x: str">

The reason for the difference in 3.10 (IIRC) is that inspect.signature auto converts all strings to typing.ForwardRef internally and then resolves them. This is a result of PEP 563 becoming default (also mentioned here )

Prior to 3.10, the future import treats function annotations as strings and inspect.signature _doesn't_ convert them. I don't know of this is a bug or intentional. Especially considering what PEP 563 has to say:

@1ace, a possible workaround if you want full compatibility regardless of the future import is to use typing.get_type_hints:

(The result doesn't change even with from __future__ import annotations. This also works from 3.7-3.10)
>>> from typing import get_type_hints
>>> get_type_hints(foo)
{'x': <class 'str'>, 'return': <class 'str'>}
msg388025 - (view) Author: Eric Engestrom (1ace) Date: 2021-03-03 15:56
Sorry, I just realized you asked for the "success" value as well.

For python 3.7-3.9 without the `import annotations`, and for python 3.10, the type annotation is a class, ie. `<Parameter "x: str">`.
msg388029 - (view) Author: Eric Engestrom (1ace) Date: 2021-03-03 16:25
(... and I had the window open for too long and didn't notice Ken's reply before mine ^^')

@kj: thanks for the workaround! it's good to be able to change our code and not have to depend on a yet-to-be-released patch on each python version :)

That said, typing.get_type_hints() only returns the type hint, whereas our current test checks for other things (like the default value) which need to match (I think) for cython's use of `__signature__`? (I inherited all this code, so I'm not familiar with it and its history, and I'm beginning to question this test's necessity.)
msg388033 - (view) Author: Eric V. Smith (eric.smith) * (Python committer) Date: 2021-03-03 16:57
I think this is all about: should inspect.signature() resolve string annotations into actual types (via get_type_hints, or whatever)? I don't use expect much, so I can't offer an opinion there.
msg388169 - (view) Author: Terry J. Reedy (terry.reedy) * (Python committer) Date: 2021-03-05 22:19
3.7 only gets security fixes
msg392912 - (view) Author: Jelle Zijlstra (JelleZijlstra) * (Python committer) Date: 2021-05-04 15:41
Python 3.10 will add an eval_str= argument to inspect.signature() that lets you evaluate string annotations. I don't think it makes sense to change anything in the bugfix branches, so I propose that this issue be closed.
Date User Action Args
2022-04-11 14:59:42adminsetgithub: 87521
2021-05-04 15:41:06JelleZijlstrasetnosy: + JelleZijlstra
messages: + msg392912
2021-03-15 18:32:00The Compilersetnosy: + The Compiler
2021-03-05 22:19:41terry.reedysetnosy: + terry.reedy

messages: + msg388169
versions: - Python 3.7
2021-03-03 16:57:55eric.smithsetmessages: + msg388033
2021-03-03 16:25:111acesetmessages: + msg388029
2021-03-03 15:56:461acesetmessages: + msg388025
2021-03-03 14:41:54kjsetnosy: + kj
messages: + msg388023
2021-03-03 13:41:531acesetmessages: + msg388018
2021-03-02 16:10:31eric.smithsetnosy: + eric.smith
messages: + msg387937
2021-03-01 16:02:001acecreate