Why do concurrent.futures.ProcessPoolExecutor and multiprocessing.pool.Pool fail with super in Python?

问题

Why does the following Python code using the concurrent.futures module hang forever?

import concurrent.futures


class A:

    def f(self):
        print("called")


class B(A):

    def f(self):
        executor = concurrent.futures.ProcessPoolExecutor(max_workers=2)
        executor.submit(super().f)


if __name__ == "__main__":
    B().f()

The call raises an invisible exception [Errno 24] Too many open files (to see it, replace the line executor.submit(super().f) with print(executor.submit(super().f).exception())).

However, replacing ProcessPoolExecutor with ThreadPoolExecutor prints "called" as expected.

Why does the following Python code using the multiprocessing.pool module raise the exception AssertionError: daemonic processes are not allowed to have children?

import multiprocessing.pool


class A:

    def f(self):
        print("called")


class B(A):

    def f(self):
        pool = multiprocessing.pool.Pool(2)
        pool.apply(super().f)


if __name__ == "__main__":
    B().f()

However, replacing Pool with ThreadPool prints "called" as expected.

Environment: CPython 3.7, MacOS 10.14.

回答1:

concurrent.futures.ProcessPoolExecutor and multiprocessing.pool.Pool uses multiprocessing.queues.Queue to pass the work function object from caller to worker process, Queue uses pickle module to serialize/unserialize, but it failed to proper processing bound method object with child class instance:

f = super().f
print(f)
pf = pickle.loads(pickle.dumps(f))
print(pf)

outputs:

<bound method A.f of <__main__.B object at 0x104b24da0>>
<bound method B.f of <__main__.B object at 0x104cfab38>>

A.f becomes B.f, this effectly creates infinite recursive calling B.f to B.f in the worker process.

pickle.dumps utilize __reduce__ method of bound method object, IMO, its implementation, has no consideration of this scenario, which does not take care of the real func object, but only try to get back from instance self obj (B()) with the simple name (f), which resulting B.f, very likely a bug.

good news is, as we know where the issue is, we could fix it by implementing our own reduction function that tries to recreate the bound method object from the original function (A.f) and instance obj (B()):

import types
import copyreg
import multiprocessing

def my_reduce(obj):
    return (obj.__func__.__get__, (obj.__self__,))

copyreg.pickle(types.MethodType, my_reduce)
multiprocessing.reduction.register(types.MethodType, my_reduce)

we could do this because bound method is a descriptor.

ps: I have filed a bug report.

来源：https://stackoverflow.com/questions/56609847/why-do-concurrent-futures-processpoolexecutor-and-multiprocessing-pool-pool-fail

标签

python

pickle

python-multiprocessing

process-pool