multiprocessing.Queue deadlocks after "reader" process death

Question

I've been playing with multiprocessing package and noticed that queue can be deadlocked for reading when:

1. The "reader" process is using get with timeout > 0:

   self.queue.get(timeout=3)
   

2. "reader" dies while get is blocking due to timeout.

After that queue is locked forever.

Application demonstrating the problem

I create two child processes "Worker" (putting into queue) and "Receiver" (getting from queue). Also parent process periodically checks if his children are alive and starts new child if needed.

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import multiprocessing
import procname
import time

class Receiver(multiprocessing.Process):
    ''' Reads from queue with 3 secs timeout '''

    def __init__(self, queue):
        multiprocessing.Process.__init__(self)
        self.queue = queue

    def run(self):
        procname.setprocname('Receiver')
        while True:
            try:
                msg = self.queue.get(timeout=3)
                print '<<< `{}`, queue rlock: {}'.format(
                    msg, self.queue._rlock)
            except multiprocessing.queues.Empty:
                print '<<< EMPTY, Queue rlock: {}'.format(
                    self.queue._rlock)
                pass


class Worker(multiprocessing.Process):
    ''' Puts into queue with 1 sec sleep '''

    def __init__(self, queue):
        multiprocessing.Process.__init__(self)
        self.queue = queue

    def run(self):
        procname.setprocname('Worker')
        while True:
            time.sleep(1)
            print 'Worker: putting msg, Queue size: ~{}'.format(
                self.queue.qsize())
            self.queue.put('msg from Worker')


if __name__ == '__main__':
    queue = multiprocessing.Queue()

    worker = Worker(queue)
    worker.start()

    receiver = Receiver(queue)
    receiver.start()

    while True:
        time.sleep(1)
        if not worker.is_alive():
            print 'Restarting worker'
            worker = Worker(queue)
            worker.start()
        if not receiver.is_alive():
            print 'Restarting receiver'
            receiver = Receiver(queue)
            receiver.start()

How processes tree looks like in ps

bash
 \_ python queuetest.py
     \_ Worker
     \_ Receiver

Console output

$ python queuetest.py
Worker: putting msg, Queue size: ~0
<<< `msg from Worker`, queue rlock: <Lock(owner=None)>
Worker: putting msg, Queue size: ~0
<<< `msg from Worker`, queue rlock: <Lock(owner=None)>
Restarting receiver                        <-- killed Receiver with SIGTERM
Worker: putting msg, Queue size: ~0
Worker: putting msg, Queue size: ~1
Worker: putting msg, Queue size: ~2
<<< EMPTY, Queue rlock: <Lock(owner=SomeOtherProcess)>
Worker: putting msg, Queue size: ~3
Worker: putting msg, Queue size: ~4
Worker: putting msg, Queue size: ~5
<<< EMPTY, Queue rlock: <Lock(owner=SomeOtherProcess)>
Worker: putting msg, Queue size: ~6
Worker: putting msg, Queue size: ~7

Is there any way to bypass this? Using get_nowait combined with sleep seems to be some kind of workaround but it does not read the data "as it comes".

System information

$ uname -sr
Linux 3.11.8-200.fc19.x86_64

$ python -V
Python 2.7.5

In [3]: multiprocessing.__version__
Out[3]: '0.70a1'

"it just works" solution

While writing this question I came up with some silly modification to Receiver class:

class Receiver(multiprocessing.Process):

    def __init__(self, queue):
        multiprocessing.Process.__init__(self)
        self.queue = queue

    def run(self):
        procname.setprocname('Receiver')
        while True:
            time.sleep(1)
            while True:
                try:
                    msg = self.queue.get_nowait()
                    print '<<< `{}`, queue rlock: {}'.format(
                        msg, self.queue._rlock)
                except multiprocessing.queues.Empty:
                    print '<<< EMPTY, Queue rlock: {}'.format(
                        self.queue._rlock)
                    break

But it doesn't seem very good to me.

Answer 1

It's probably because *not_empty.release()* from Queue.get() never happends (the proccess has been killed already). Did you try to catch the TERM signal in Receiver and release the Queue mutex before exiting?