Option to kill "stuck" workers in a multiprocessing pool

[pfmoore]

BPO	14148
Nosy	@pfmoore, @brianquinlan, @pitrou

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields: ```python assignee = None closed_at = None created_at = labels = ['type-feature', 'library'] title = 'Option to kill "stuck" workers in a multiprocessing pool' updated_at = user = 'https://github.com/pfmoore' ``` bugs.python.org fields: ```python activity = actor = 'bquinlan' assignee = 'none' closed = False closed_date = None closer = None components = ['Library (Lib)'] creation = creator = 'paul.moore' dependencies = [] files = [] hgrepos = [] issue_num = 14148 keywords = [] message_count = 3.0 messages = ['154549', '154573', '154575'] nosy_count = 4.0 nosy_names = ['paul.moore', 'bquinlan', 'pitrou', 'neologix'] pr_nums = [] priority = 'normal' resolution = None stage = None status = 'open' superseder = None type = 'enhancement' url = 'https://bugs.python.org/issue14148' versions = ['Python 3.3', 'Python 3.4'] ```

[pfmoore]

I have an application which fires off a number of database connections via a multiprocessing pool. Unfortunately, the database software occasionally gets "stuck" and a connection request hangs indefinitely. This locks up the whole process doing the connection, and cannot be interrupted except by killing the process.

It would be useful to have a facility to restart "stuck" workers in this case.

As an interface, I would suggest an additional argument to the AsyncResult.get method, kill_on_timeout. If this argument is true, and the get times out, the worker servicing the result will be killed and restarted.

Alternatively, provide a method on an AsyncResult to access the worker process that is servicing the request. I could then wait on the result and kill the worker manually if it does not respond in time.

Without a facility like this, there is a potential for the pool to get starved of workers if multiple connections hang.

[pitrou]

The problem is that queues and other synchronization objects can end up in an inconsistent state when a worker crashes, hangs or gets killed. That's why, in concurrent.futures, a crashed worker makes the ProcessPoolExecutor become "broken". A similar thing should be done for multiprocessing.Pool but it's a more complex object.

[pfmoore]

As an alternative, maybe leave the "stuck" worker, but allow the pool to recognise when a worker has not processed new messages for a long period and spawn an extra worker to replace it. That would avoid the starvation issue, and the stuck workers would die when the pool is terminated.