Python 3.12+ breaks backwards compatibility for logging QueueHandler with some Queue classes

[shmilee]

Bug report

Bug description:

Related bug: #111615
Related pull: #111638
Related codes:

cpython/Lib/logging/config.py

Lines 789 to 804 in bd0d97c

	if 'queue' in config:
	from multiprocessing.queues import Queue as MPQueue
	qspec = config['queue']
	if not isinstance(qspec, (queue.Queue, MPQueue)):
	if isinstance(qspec, str):
	q = self.resolve(qspec)
	if not callable(q):
	raise TypeError('Invalid queue specifier %r' % qspec)
	q = q()
	elif isinstance(qspec, dict):
	if '()' not in qspec:
	raise TypeError('Invalid queue specifier %r' % qspec)
	q = self.configure_custom(dict(qspec))
	else:
	raise TypeError('Invalid queue specifier %r' % qspec)
	config['queue'] = q

reproducible example using Python 3.12.3

import logging.config


def main(q):
    config = {
        'version': 1,
        'handlers': {
            'sink': {
                'class': 'logging.handlers.QueueHandler',
                'queue': q,
            },
        },
        'root': {
            'handlers': ['sink'],
        },
    }
    logging.config.dictConfig(config)


if __name__ == '__main__':
    import multiprocessing as mp
    main(mp.Manager().Queue())
    #import asyncio
    #main(asyncio.Queue())  # broken too

error:

Traceback (most recent call last):
  File "/usr/lib/python3.12/logging/config.py", line 581, in configure
    handler = self.configure_handler(handlers[name])
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/logging/config.py", line 801, in configure_handler
    raise TypeError('Invalid queue specifier %r' % qspec)
TypeError: Invalid queue specifier <AutoProxy[Queue] object, typeid 'Queue' at 0x7f8b07189d90>

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/home/xxx/logging-queue-handler-bug.py", line 33, in <module>
    main(mp.Manager().Queue())
  File "/home/xxx/logging-queue-handler-bug.py", line 29, in main
    logging.config.dictConfig(config)
  File "/usr/lib/python3.12/logging/config.py", line 914, in dictConfig
    dictConfigClass(config).configure()
  File "/usr/lib/python3.12/logging/config.py", line 588, in configure
    raise ValueError('Unable to configure handler '
ValueError: Unable to configure handler 'sink'

Queue classes to check for logging QueueHandler

• queue.Queue(), <queue.Queue at 0x7fb86eeef170>, works.
• multiprocessing
  • multiprocessing.Queue(), <multiprocessing.queues.Queue at 0x7fb871790500>, works.
  • multiprocessing.Manager().Queue() <AutoProxy[Queue] object, typeid 'Queue' at 0x7fb86ef4a840>, broken.
    Its class is multiprocessing.managers.AutoProxy[Queue], a subclass of multiprocessing.managers.BaseProxy.
• asyncio.queues.Queue() <Queue at 0x7fb86f0be4e0 maxsize=0>, broken.
• any other Queue?

discuss: how to fix

Check all Queue classes mentioned above in

cpython/Lib/logging/config.py

Line 792 in bd0d97c

if not isinstance(qspec, (queue.Queue, MPQueue)):

or just focus on handling special cases: str and dict, while leaving other cases un-checked (e.g., queue.Queue, multiprocessing.queues.Queue).

CPython versions tested on:

3.12

Operating systems tested on:

Linux

Linked PRs

• gh-119819: Fix regression to allow logging configuration with multipr… #120030
• [3.12] gh-119819: Fix regression to allow logging configuration with multipr… (GH-120030) #120034
• [3.13] gh-119819: Fix regression to allow logging configuration with multipr… (GH-120030) #120035
• gh-119819: Update test to skip if _multiprocessing is unavailable. #120067
• [3.12] gh-119819: Update test to skip if _multiprocessing is unavailable. (GH-120067) #120071
• [3.13] gh-119819: Update test to skip if _multiprocessing is unavailable. (GH-120067) #120072
• gh-119819: Update logging configuration to support joinable multiproc… #120090
• [3.12] gh-119819: Update logging configuration to support joinable multiproc… (GH-120090) #120092
• [3.13] gh-119819: Update logging configuration to support joinable multiproc… (GH-120090) #120093
• gh-119819: Conditional skip of logging tests that require multiprocessing subprocess support #120476
• [3.13] gh-119819: Conditional skip of logging tests that require multiprocessing subprocess support (GH-120476) #120531
• [3.12] gh-119819: Conditional skip of logging tests that require multiprocessing subprocess support (GH-120476) #120532

[picnixz]

I'm not really sure about this but would it actually make sense to only check for protocol-based queues and not the actual types of queues? strictly speaking, we only need a queue-like interface right? If so, I can work on that one.

[shmilee]

After reading the codes above and this pull request, I tend to not check the queue or queue-like classes, as the main purpose of the code from lines 789 to 804 is to handle the special cases of isinstance(qspec, str) and isinstance(qspec, dict) and then update config['queue'] to be a valid queue instance.
Other cases are considered that config['queue'] is already a queue, like an instance of queue.Queue or any other Queue type, which is the common use case in the previous version of Python.

Perhaps we can ask @vsajip for some advice. Is there anything else we should consider?

[vsajip]

Can you explain the use cases which require multiprocessing.Manager().Queue() (the AutoProxy)? Also, since logging isn't async, why would you want to pass an asyncio.queues.Queue()?

[shmilee]

I have only used multiprocessing.Manager().Queue(). The asyncio.queues.Queue() mentioned is just a possible use case I imagined after searching the standard library for queues. So I can only talk about the former.

Using Python 3.12 on Linux, I change manager.Queue() (manager is a Manager instance ) to multiprocessing.Queue(), in my codes logging also works. So, currently, it seems that using manager.Queue() is not mandatory.
The reason for choosing it before is that in previous versions of Python, multiprocessing.Queue() may cause some errors related to multiprocessing Pool, spawn, and fork things, on Linux or Windows.

Sorry for leaking an example code with logging, because tests on Windows, and installing previous versions of Python, may take a lot of time. However, some information about using use a queue created with a manager to avoid some potential issues can be provided below after I searched:

1. python-multiprocessing-queue-vs-multiprocessing-manager-queue
2. https://docs.python.org/3.12/library/multiprocessing.html#pipes-and-queues

   If they really bother you then you can instead use a queue created with a manager.

3. google keywords: RuntimeError: Queue objects should only be shared between processes through inheritance.
   • Example_of_Error_When_Sharing_a_Queue_with_a_Pool
   • Example_of_Sharing_a_Queue_with_a_Pool_Via_a_Manager

[vsajip]

Do you have an actual use case where you're running into problems? Or is it just a theoretical exercise? If an actual use case, please give details.

[shmilee]

Here is an actual use case，and quickly show the problem:

cd /tmp/
git clone --depth 1 --branch v0.8.3 https://github.com/shmilee/gdpy3.git
cd gdpy3/
python3.12 -m unittest src/processors/tests/test_multiprocessor.py

Then I will get errors from more and more ForkPoolWorker-XXX, XXX is an increasing number.

Process ForkPoolWorker-210:
Traceback (most recent call last):
  File "/usr/lib/python3.12/logging/config.py", line 581, in configure
    handler = self.configure_handler(handlers[name])
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/logging/config.py", line 801, in configure_handler
    raise TypeError('Invalid queue specifier %r' % qspec)
TypeError: Invalid queue specifier <AutoProxy[Queue] object, typeid 'Queue' at 0x7f5250423d70>

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/usr/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/usr/lib/python3.12/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "/usr/lib/python3.12/multiprocessing/pool.py", line 109, in worker
    initializer(*initargs)
  File "/tmp/gdpy3/src/glogger.py", line 167, in __call__
    logging.config.dictConfig(self.glogger_config)
  File "/usr/lib/python3.12/logging/config.py", line 914, in dictConfig
    dictConfigClass(config).configure()
  File "/usr/lib/python3.12/logging/config.py", line 588, in configure
    raise ValueError('Unable to configure handler '
ValueError: Unable to configure handler 'queue'

[shmilee]

1. Using multiprocessing.Queue() can fix it. see here

2. I also test this patch on logging/config.py. The error also disappears.
   (cannot find class AutoProxy[Queue] to use in managers module, so just test with BaseProxy)

--- config.py	2024-05-31 10:16:24.905893630 +0800
+++ config.py-fix3.12	2024-05-31 13:18:07.223555882 +0800
@@ -786,8 +786,11 @@
                     # raise ValueError('No handlers specified for a QueueHandler')
                 if 'queue' in config:
                     from multiprocessing.queues import Queue as MPQueue
+                    from multiprocessing.managers import BaseProxy as MPBaseProxy                    
                     qspec = config['queue']
-                    if not isinstance(qspec, (queue.Queue, MPQueue)):
+                    print('============', isinstance(qspec, MPBaseProxy))
+                    print(qspec, type(qspec))
+                    if not isinstance(qspec, (queue.Queue, MPQueue, MPBaseProxy)):

[vsajip]

No, that's not the use case I meant. What I meant is, do you have a real problem where you have to use the proxy and can't use multiprocessing.Queue?

[shmilee]

No. For me, using logging with multiprocessing.Queue also works fine.

[shmilee]

And I think using the proxy "manager.Queue()" should not be excluded / droped, it was working fine before 3.12.
Even in version 3.12, there is still a way to bypass the existing restrictions (Line-792), using the much more powerful dict qspec configuration. About the listener part, I haven't figured out how to set up a "manager.Queue()" together with handler yet.

import logging.config
import multiprocessing as mp
import asyncio

def main(i, qd):
    print('%d. qspec dict: ' % i, qd)
    config = {
        'version': 1,
        'handlers': {
            'sink': {
                'class': 'logging.handlers.QueueHandler',
                'queue': qd,
            },
        },
        'root': {
            'handlers': ['sink'],
        },
    }
    logging.config.dictConfig(config)
    l = logging.getLogger()
    s = l.handlers[0]
    print('USE: ', id(s.queue), s.queue, type(s.queue))
    print('------')

def get_mp_queue(manager=None, maxsize=None):
    q = manager.Queue(maxsize)
    print('GET: ', id(q), q)
    return q

def get_asy_queue(maxsize=None):
    q = asyncio.Queue(maxsize)
    print('GET: ', id(q), q)
    return q

if __name__ == '__main__':
    m = mp.Manager()
    main(1, {'()': get_mp_queue, 'manager': m, 'maxsize': 10})
    main(2, {'()': m.Queue, 'maxsize': 20})
    main(3, {'()': get_asy_queue, 'maxsize': 10})  # not async, so the user is responsible for his choice.

Output:

1. qspec dict:  {'()': <function get_mp_queue at 0x7fd645687100>, 'manager': <multiprocessing.managers.SyncManager object at 0x7fd644d2c440>, 'maxsize': 10}
GET:  140558252776848 <queue.Queue object at 0x7fd644bccbf0>
USE:  140558252776848 <queue.Queue object at 0x7fd644bccbf0> <class 'multiprocessing.managers.AutoProxy[Queue]'>
------
2. qspec dict:  {'()': <bound method BaseManager.register.<locals>.temp of <multiprocessing.managers.SyncManager object at 0x7fd644d2c440>>, 'maxsize': 20}
USE:  140558252622480 <queue.Queue object at 0x7fd644bccf50> <class 'multiprocessing.managers.AutoProxy[Queue]'>
------
3. qspec dict:  {'()': <function get_asy_queue at 0x7fd645532200>, 'maxsize': 10}
GET:  140558252955872 <Queue maxsize=10>
USE:  140558252955872 <Queue maxsize=10> <class 'asyncio.queues.Queue'>
------

[vsajip]

And I think using the proxy "manager.Queue()" should not be excluded / droped

Why should it not be excluded, since multiprocessing.Queue () is available to be used? You can't check the class easily as AutoProxy could be used to proxy other things, not just queues.