python multiprocessing: some functions do not return when they are complete (queue material too big)

Question

I am using multiprocessing's Process and Queue. I start several functions in parallel and most behave nicely: they finish, their output goes to their Queue, and they show up as .is_alive() == False. But for some reason a couple of functions are not behaving. They always show .is_alive() == True, even after the last line in the function (a print statement saying "Finished") is complete. This happens regardless of the set of functions I launch, even it there's only one. If not run in parallel, the functions behave fine and return normally. What kind of thing might be the problem?

Here's the generic function I'm using to manage the jobs. All I'm not showing is the functions I'm passing to it. They're long, often use matplotlib, sometimes launch some shell commands, but I cannot figure out what the failing ones have in common.

def  runFunctionsInParallel(listOf_FuncAndArgLists):
    """
    Take a list of lists like [function, arg1, arg2, ...]. Run those functions in parallel, wait for them all to finish, and return the list of their return values, in order.   
    """
    from multiprocessing import Process, Queue

    def storeOutputFFF(fff,theArgs,que): #add a argument to function for assigning a queue
        print 'MULTIPROCESSING: Launching %s in parallel '%fff.func_name
        que.put(fff(*theArgs)) #we're putting return value into queue
        print 'MULTIPROCESSING: Finished %s in parallel! '%fff.func_name
        # We get this far even for "bad" functions
        return

    queues=[Queue() for fff in listOf_FuncAndArgLists] #create a queue object for each function
    jobs = [Process(target=storeOutputFFF,args=[funcArgs[0],funcArgs[1:],queues[iii]]) for iii,funcArgs in enumerate(listOf_FuncAndArgLists)]
    for job in jobs: job.start() # Launch them all
    import time
    from math import sqrt
    n=1
    while any([jj.is_alive() for jj in jobs]): # debugging section shows progress updates
        n+=1
        time.sleep(5+sqrt(n)) # Wait a while before next update. Slow down updates for really long runs.
        print('\n---------------------------------------------------\n'+ '\t'.join(['alive?','Job','exitcode','Func',])+ '\n---------------------------------------------------')
        print('\n'.join(['%s:\t%s:\t%s:\t%s'%(job.is_alive()*'Yes',job.name,job.exitcode,listOf_FuncAndArgLists[ii][0].func_name) for ii,job in enumerate(jobs)]))
        print('---------------------------------------------------\n')
    # I never get to the following line when one of the "bad" functions is running.
    for job in jobs: job.join() # Wait for them all to finish... Hm, Is this needed to get at the Queues?
    # And now, collect all the outputs:
    return([queue.get() for queue in queues])

Reena

Complete shot in the dark: Do the ones that hang return a value? (literally, do they have return in them?) – Morganica

All functions, good and bad, return a single (long) string. – Reena

However, if I eliminate the use of Queues, the problem goes away. So... a queue has been filled. I can look at it, and it looks fine, but somehow the job is not finishing when there's an associated queue (and only for "bad" functions). – Reena

It may be due to the size of what I'm putting in the queue. #10029309 If I put a large fixed string in the queue, the job won't end. If it's small, it does. I don't understand how to get around this yet... – Reena

Try bumping up the queue size. You could also have your subprocesses essentially mimic the queue blocking/timeout functionality by having them write to the queue with an explicit timeout, then catching the Full exception and printing, then repeating the put. Try chunking the string writes to the queue too. – Cutlet

That sounds complex. Is there really not a simpler way to buffer these things? Anyway, from my experimenting, it seems that I don't need to chunk the input (wouldn't know how, since the kind of return value is arbitrary). Even though the queue is not necessarily full when functions were hanging, just doing some .get()s on the other end pulls stuff through and allows the processes to finish. I've now implemented that in my proposed answer. – Reena

Upping the queue size solves the same problem. It's a simple argument to the Queue constructor. – Cutlet

Thanks. I don't think I'd know how big to make it in advance. My general-function tool requires a more general solution. – Reena

Answer 1

Alright, it seems that the pipe used to fill the Queue gets plugged when the output of a function is too big (my crude understanding? This is an unresolved/closed bug? http://bugs.python.org/issue8237). I have modified the code in my question so that there is some buffering (queues are regularly emptied while processes are running), which solves all my problems. So now this takes a collection of tasks (functions and their arguments), launches them, and collects the outputs. I wish it were simpler /cleaner looking.

Edit (2014 Sep; update 2017 Nov: rewritten for readability): I'm updating the code with the enhancements I've made since. The new code (same function, but better features) is here: https://gitlab.com/cpbl/cpblUtilities/blob/master/parallel.py

The calling Description is also below.

def runFunctionsInParallel(*args, **kwargs):
    """ This is the main/only interface to class cRunFunctionsInParallel. See its documentation for arguments.
    """
    return cRunFunctionsInParallel(*args, **kwargs).launch_jobs()

###########################################################################################
###
class cRunFunctionsInParallel():
    ###
    #######################################################################################
    """Run any list of functions, each with any arguments and keyword-arguments, in parallel.
The functions/jobs should return (if anything) pickleable results. In order to avoid processes getting stuck due to the output queues overflowing, the queues are regularly collected and emptied.
You can now pass os.system or etc to this as the function, in order to parallelize at the OS level, with no need for a wrapper: I made use of hasattr(builtinfunction,'func_name') to check for a name.
Parameters
----------
listOf_FuncAndArgLists : a list of lists 
    List of up-to-three-element-lists, like [function, args, kwargs],
    specifying the set of functions to be launched in parallel.  If an
    element is just a function, rather than a list, then it is assumed
    to have no arguments or keyword arguments. Thus, possible formats
    for elements of the outer list are:
      function
      [function, list]
      [function, list, dict]
kwargs: dict
    One can also supply the kwargs once, for all jobs (or for those
    without their own non-empty kwargs specified in the list)
names: an optional list of names to identify the processes.
    If omitted, the function name is used, so if all the functions are
    the same (ie merely with different arguments), then they would be
    named indistinguishably
offsetsSeconds: int or list of ints
    delay some functions' start times
expectNonzeroExit: True/False
    Normal behaviour is to not proceed if any function exits with a
    failed exit code. This can be used to override this behaviour.
parallel: True/False
    Whenever the list of functions is longer than one, functions will
    be run in parallel unless this parameter is passed as False
maxAtOnce: int
    If nonzero, this limits how many jobs will be allowed to run at
    once.  By default, this is set according to how many processors
    the hardware has available.
showFinished : int
    Specifies the maximum number of successfully finished jobs to show
    in the text interface (before the last report, which should always
    show them all).
Returns
-------
Returns a tuple of (return codes, return values), each a list in order of the jobs provided.
Issues
-------
Only tested on POSIX OSes.
Examples
--------
See the testParallel() method in this module
    """

Reena

"If this doesn't work, maybe the stuff you're returning from your functions is not pickleable, and therefore unable to make it through the Queues properly." Huge help, I had this exact problem and did not know that multiprocessing relies on pickling to pass objects between processes (including returning results). – Knockabout

Just a suggestion, but you should invest a bit of time in making this readable. There's probably some really useful stuff in here, but it's next to impossible to tell. – Tildie

@BrianCline I tried to clean up the code: gitlab.com/cpbl/cpblUtilities/blob/master/parallel.py I don't expect this is PEP-8, but it's all I can muster (and we use this function a tonne, to great effect). Luckily, it's open to improvement! – Reena