is_generic_type(List) is False in 3.9

[JelleZijlstra]

In [7]: sys.version                                                                                                                                                                  
Out[7]: '3.9.0b1 (v3.9.0b1:97fe9cfd9f, May 18 2020, 20:39:28) \n[Clang 6.0 (clang-600.0.57)]'

In [8]: from typing import List                                                                                                                                                      

In [9]: typing_inspect.is_generic_type(List)                                                                                                                                         
Out[9]: False

Apparently in 3.9 List became an instance of _SpecialGenericAlias instead of _GenericAlias, which is what typing_inspect checks for.

[thatch]

This same bug has far-reaching effects and is the cause of Instagram/LibCST#305 as well.

[thatch]

This is a little more interesting, and mainly appears to affect unsubscripted types that ought to be considered generic. In 3.8, these have __args__ with TypeVar; in 3.9 there's only _nargs.

If we apply the trivial fix to consider unsubscripted types to be generic as well, we get a predictable AttributeError. If we return empty-tuple, that's a behavior change, and I'd like to have others weigh in on whether this is a bug in upstream 3.9, and whether get_args returning empty-tuple is good enough. That would work for the current implementation in LibCST, since it just skips further validation.

[thatch]

Jelle, I had a conversation with Łukasz and I don't feel like we have buy-in on this being a cpython bug. I don't think there's a way to fix this in typing_inspect, can you help me with considering fixes and escalating?

[ilevkivskyi]

FWIW, a potential fix can be hard-coding some special cases. Technically, typing_inspect uses a bunch of private APIs to provide a bit deeper level of introspection than the one provided by typing, so I can see why this is not an upstream bug.

I may not have time to work on this soon, but if someone has ideas for a robust fix, I will be glad to merge a PR.

[JelleZijlstra]

Similar to what Ivan said, I'm not sure why this can't be fixed in typing-inspect, for example by looking for the _SpecialGenericAlias class inside the implementation of is_generic_type().

I ran into this doing some runtime type inspection in pyanalyze and worked around it with some 3.9-specific code: https://github.com/quora/pyanalyze/blob/master/pyanalyze/annotations.py#L225.

Is there a bugs.python.org issue about this?

[baldurmen]

FWIW, this means I can't build typing_inspect in Debian anymore :(

https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=972803

Hopefully someone can implement a fix soonish, since it won't move from unstable to testing until this is fixed.