Can't copy.copy or deepcopy any type that inherits from _FinalTypingBase

[ento]

Appears to be since #283, according to git bisect.

To repro:

$ python --version
Python 3.6.0b2
$ python -c 'import typing, copy; copy.copy(typing.Any)'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/usr/local/Cellar/python3/3.6.0b2_3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/copy.py", line 106, in copy
    return _reconstruct(x, None, *rv)
  File "/usr/local/Cellar/python3/3.6.0b2_3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/copy.py", line 274, in _reconstruct
    y = func(*args)
  File "/usr/local/Cellar/python3/3.6.0b2_3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/copyreg.py", line 88, in __newobj__
    return cls.__new__(cls, *args)
  File "/usr/local/Cellar/python3/3.6.0b2_3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/typing.py", line 191, in __new__
    raise TypeError("Cannot instantiate %r" % cls)
TypeError: Cannot instantiate typing.Any

Motivation for copying a type: deepcopying a function's __annotations__ makes it possible to do manipulations on the annotations to get only the data you want, like type annotations for just the function arguments.

annotations = copy.deepcopy(func.__annotations__)
annotations.pop('return', None)

[gvanrossum]

Hm, there are tests for pickling, I guess they are incomplete. :-( @ilevkivskyi thoughts?

[ilevkivskyi]

@gvanrossum Yes, this could be fixed simply by adding a trivial __reduce__ to things that are intended to be singletons. I tried this locally and it works with both pickle and deepcopy. I will make a PR for this right after I finish my work on #301 #299 #298 (it is almost ready, will make PR today or tomorrow).

By the way, I think we could not make subscripted generics like List[int] pickleable, since pickle identifies classes by __name__. Although, subscripted generics are deepcopyable.

[gvanrossum]

We typically allow pickles to reference classes by (fully-qualified) name, so I don't see why we shouldn't allow that here.

[ilevkivskyi]

@gvanrossum The problem is not with being qualified. X and X[int] have equal names (also qualified names). For some reason pickle hates this:
_PickleError: can't pickle typing.X[int]: it's not the same object as typing.X
Even if try to __reduce__ (probably this is the reason why there are no subscripted generics in tests). Do you know how to overcome this limitation?

Anyway, there are no problems with copy, deepcopy; and unsubscripted generics and instances are pickleable, so that the original issue could be fixed.

[gvanrossum]

IIRC there are other APIs to override pickling. See the pickle docs, I haven't used them in ages. :-(

[gvanrossum]

(Or maybe you'd need to add pickle support to the metaclass? I'm reaching here, out of time, and I'm not sure it's worth it trying to support pickling of List[int]. (The OP just has concerns with deepcopy().)

[ento]

I (== OP) personally won't mind waiting for the right fix if supporting both pickling and copying is an entangled issue, and if it doesn't make sense to support just copying. Currently I can workaround this issue without much impact (don't type-annotate this arg, only copy after doing that, etc.).

[mitar]

I do not think this was really solved: https://github.com/python/typing/pull/311/files#r132876422