functools.wraps can pollute a function's global namespace

[pierreglaser]

using functools.wraps

functools.wraps(f)
def g():
 pass

changes a g's __module__ attribute to f.__module__, but not its __globals__. Because we retrieve globals from state['module'] at unpickling time, this will pollute a depickled version of g with the globals of f.__module__

As an illustration:

import functools
import pickle
import cloudpickle


@functools.wraps(pickle._dumps)
def f():
 pass


depickled_f = pickle.loads(cloudpickle.dumps(f))
assert depickled_f.__globals__ is vars(pickle)

passes.

This is tricky:
On one hand one can spot if a function is wrapping another by checking for a __wrapped__ attribute. On the other hand, once wraps is used, there is no way back, and f's attributes are lost forever.
The side effects are not life threatening, but this may be worth adding a warning at pickling time. I am thinking about (roughly):

if hasattr(func, '__wrapped__'):
 msg = ("This function seems decorated by functools.wraps. This"
 " may cause {}'s __globals__ to be populated with"
 " unnecessary global variables at unpickling time."
 " To fix it, change the assigned argument of"
 " functools.wraps so that is does not include __module__ ")
 raise UserWarning(msg.format(func.__code__.co_name))

Inside extract_func_data.

[pierreglaser]

This is solved by #240. Closing.