Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

_pickle.PicklingError on 3.14.0a1+ where it didn't before due to the start method change from fork to forkserver #125714

Open
progval opened this issue Oct 18, 2024 · 11 comments
Assignees
Labels
3.14 new features, bugs and security fixes docs Documentation in the Doc dir stdlib Python modules in the Lib dir topic-multiprocessing type-bug An unexpected behavior, bug, or error

Comments

@progval
Copy link

progval commented Oct 18, 2024

Bug report

Bug description:

Code dating back to Python 2 had to do super(OwnClassName, self) instead of super(). Modules designed to be reloadable could not use the super(OwnClassName, self) syntax in their non-constructor method, as OwnClassName was re-binded to the class of the new module, rather than the class that had the name when the object was constructed (which differs when the object was constructed before a reload). Therefore, such modules defined classes like this:

class Subclass(Superclass):
    def __init__(self):
        self.__parent = super(Subclass, self)

    def method1(self, arg):
        self.__parent.method1(arg + 2)

Python 3.14.0a1+ breaks existing code using this trick on subclasses of multiprocessing.Process.

For example:

import multiprocessing

class MyProcess(multiprocessing.Process):
    def __init__(self, target=None, args=(), kwargs={}):
        self.__parent = super(MyProcess, self)
        self.__parent.__init__(target=target, args=args, kwargs=kwargs)

    def run(self):
        self.__parent.run()
    
if __name__ == "__main__":
    p = MyProcess()
    p.start()

worked fine on previous versions, but now errors with:

Traceback (most recent call last):
  File "/home/dev-irc/Limnoria/repro.py", line 12, in <module>
    p.start()
    ~~~~~~~^^
  File "/home/dev-irc/.local-py-git/lib/python3.14/multiprocessing/process.py", line 121, in start
    self._popen = self._Popen(self)
                  ~~~~~~~~~~~^^^^^^
  File "/home/dev-irc/.local-py-git/lib/python3.14/multiprocessing/context.py", line 224, in _Popen
    return _default_context.get_context().Process._Popen(process_obj)
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
  File "/home/dev-irc/.local-py-git/lib/python3.14/multiprocessing/context.py", line 300, in _Popen
    return Popen(process_obj)
  File "/home/dev-irc/.local-py-git/lib/python3.14/multiprocessing/popen_forkserver.py", line 35, in __init__
    super().__init__(process_obj)
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
  File "/home/dev-irc/.local-py-git/lib/python3.14/multiprocessing/popen_fork.py", line 20, in __init__
    self._launch(process_obj)
    ~~~~~~~~~~~~^^^^^^^^^^^^^
  File "/home/dev-irc/.local-py-git/lib/python3.14/multiprocessing/popen_forkserver.py", line 47, in _launch
    reduction.dump(process_obj, buf)
    ~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^
  File "/home/dev-irc/.local-py-git/lib/python3.14/multiprocessing/reduction.py", line 60, in dump
    ForkingPickler(file, protocol).dump(obj)
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^
_pickle.PicklingError: first argument to __newobj__() must be <class 'super'>, not <class '__main__.MyProcess'>
when serializing super object
when serializing dict item '_MyProcess__parent'
when serializing MyProcess state
when serializing MyProcess object

Full version string: Python 3.14.0a1+ (heads/main:c8fd4b12e3d, Oct 18 2024, 22:51:39) [GCC 12.2.0] on linux

CPython versions tested on:

3.14, CPython main branch

Operating systems tested on:

Linux

Linked PRs

@progval progval added the type-bug An unexpected behavior, bug, or error label Oct 18, 2024
@JelleZijlstra
Copy link
Member

Your repro case also fails for me on 3.11 through 3.13. Simpler code that fails similarly:

import pickle
class X: pass
s = super(X, X())
pickle.dumps(s)

@progval
Copy link
Author

progval commented Oct 18, 2024

Strange, my repro case fails passes on 3.11.2; and it's based on a codebase that passed on 3.12 and 3.13 so far.

Yours does fail on 3.11.2 too, though.

@progval
Copy link
Author

progval commented Oct 18, 2024

Can you try this?

import multiprocessing

def f():
    def g():
        print("hello")

    multiprocessing.Process(target=g).start()
    
if __name__ == "__main__":
    f()

For me it passes on 3.11.2, but fails on 3.14.0a1+, even though function g is unpicklable no matter the version.

@JelleZijlstra
Copy link
Member

Oh, this is because I am on MacOS. The relevant change is listed in https://docs.python.org/3.14/whatsnew/3.14.html#deprecated: the default start method moved from fork to forkserver on Linux. With fork, I guess we didn't need things to be picklable.

cc @gpshead as the author of this change, but I think this will be a won't fix: We made that change after careful consideration, and this sort of issue was a known consequence. If you want the old behavior, you can either explicitly use the fork start method or refactor your code so the relevant objects can be pickled.

There might be room for improvements in how super objects are pickled, though.

@progval
Copy link
Author

progval commented Oct 19, 2024

Ah, I understand. Thanks!

I feel there is some room for improvement regarding documentation here, as this is a breaking change and not a deprecation as far as I can tell. And neither the changelog nor "Contexts and start methods" mentions only fork supports unpicklable Process attributes or targets

@gpshead
Copy link
Member

gpshead commented Oct 19, 2024

Agreed, this is mostly a documentation issue. But I don't have a good feel for exactly what to state. Care to draft up some text in a PR @progval? I think we'd want this to be clarified in both what's new 3.14 and multiprocessing docs.

@progval
Copy link
Author

progval commented Oct 20, 2024

Done in #125750

@furkanonder
Copy link
Contributor

CC: @serhiy-storchaka

@JelleZijlstra
Copy link
Member

@progval If you're interested, feel free to open an issue specifically about pickling super objects. I couldn't find any existing issue about that, and the code for super doesn't have anything geared towards pickling support. It's possible that we could fix this to make super objects properly picklable.

>>> import pickle
>>> class X: pass
... 
>>> pickle.dumps(super(X, X()))
Traceback (most recent call last):
  File "<python-input-6>", line 1, in <module>
    pickle.dumps(super(X, X()))
    ~~~~~~~~~~~~^^^^^^^^^^^^^^^
_pickle.PicklingError: first argument to __newobj__() must be <class 'super'>, not <class '__main__.X'>
when serializing super object

@serhiy-storchaka
Copy link
Member

Created a separate issue for pickling super: #125767.

@gpshead
Copy link
Member

gpshead commented Nov 14, 2024

Given the user reports (now plural) on this one, we should also catch and improve the error message within multiprocessing when pickle fails to at least point people in the right direction.

@gpshead gpshead changed the title _pickle.PicklingError on 3.14.0a1+ when starting a process that references its super() _pickle.PicklingError on 3.14.0a1+ where it didn't before due to the start method change from fork to forkserver Nov 14, 2024
@gpshead gpshead added the stdlib Python modules in the Lib dir label Nov 14, 2024
@gpshead gpshead self-assigned this Nov 14, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
3.14 new features, bugs and security fixes docs Documentation in the Doc dir stdlib Python modules in the Lib dir topic-multiprocessing type-bug An unexpected behavior, bug, or error
Projects
Status: Todo
Development

No branches or pull requests

6 participants