Title: Threads within multiprocessing Process terminate early
Type: enhancement Stage: resolved
Components: Library (Lib) Versions: Python 3.7
Status: closed Resolution: fixed
Dependencies: Superseder:
Assigned To: Nosy List: davin, eryksun, neologix, pietvo, pitrou, r.david.murray, sbt, tim.peters
Priority: normal Keywords:

Created on 2013-09-08 01:20 by pietvo, last changed 2017-08-16 18:54 by pitrou. This issue is now closed.

File name Uploaded Description Edit pietvo, 2013-09-08 01:20 Test program
Pull Requests
URL Status Linked Edit
PR 3111 merged pitrou, 2017-08-16 18:27
Messages (17)
msg197210 - (view) Author: Piet van Oostrum (pietvo) Date: 2013-09-08 01:20
When a process started as a multiprocessing Process spawns a thread, it doesn't wait until the thread terminates. It terminates the thread early when the main thread of the process terminates, as if the thread would be daemonic (it isn't).

It may sound a bit weird to start a Thread within multiprocessing, but it isn't prohibited. Neither is this behavior documented.

In the attached program the thread doesn't complete. However when the mythread.join() statement is uncommented it does run to completion.
msg197293 - (view) Author: Antoine Pitrou (pitrou) * (Python committer) Date: 2013-09-08 17:40
That's because multiprocessing exits child processes with os._exit(), not sys.exit().

The fix would be trivial (call threading._shutdown() before os._exit()), but I don't know if that's something we want to do. After all there are many things in the Python shutdown procedure that we may want to similarly replicate in multiprocessing children, such as calling atexit handlers. This screams for a more general solution, IMHO.
msg270338 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2016-07-13 20:05
This came up again today as bug 27508.  In the absence of "fixing it", we should add docs to multiprocessing explaining the high-level consequences of skipping "normal" exit processing (BTW, I'm unclear on why it's skipped).

I've certainly mixed threads with multiprocessing too, but I always explicitly .join() my threads so never noticed this.  It's sure a puzzle when it happens ;-)
msg270343 - (view) Author: Eryk Sun (eryksun) * Date: 2016-07-13 20:30
In 3.4+ it works correctly with the "spawn" start method. This uses multiprocessing.spawn.spawn_main, which exits the child via sys.exit(exitcode). "fork" and "forkserver" exit the child via os._exit(code), respectively in multiprocessing.popen_fork.Popen._launch and multiprocessing.forkserver.main.
msg270351 - (view) Author: Davin Potts (davin) * (Python committer) Date: 2016-07-14 01:54
It is a general rule that when a process terminates, it likewise terminates all its threads (unless a thread has been explicitly detached).  How it goes about doing so is complicated.

Remember that POSIX threads have no concept of parent/child among themselves and all threads are viewed as a single pool.  The section "No parents, no children" at offers us motivation for why waiting on a pthread should be explicitly requested and not assumed as a default behavior.

There are numerous differences between what a UNIX-style process and a win32 process does at termination.  Though an older post from Microsoft, a strong sense of how complicated the process-termination-begets-thread-termination truly is can be had from reading which also helps reinforce the sentiment above (needs explicit instructions on what to do, no general solution can exist).  Whereas the prior provided some sense of motivation, this link walks us through ugly complications and consequences that arise.

The short of it is that the current use of os._exit() is most appropriate in multiprocessing.  Threads should be signaled that the process is terminating but we are not generally expected to wait on those threads.

These and many other reference-worthy links help support the call for atexit-like functionality to be exposed on multiprocessing.Process.  There have been multiple issues opened on the bug tracker ultimately requesting this enhancement (do a search for multiprocessing and atexit).  I think it's a very sensible enhancement (Antoine did too and perhaps still does) and worth taking the time to pursue.  As an aside, I wonder if an equivalent to pthread_cleanup_push should also be exposed on threading.Thread.

When it comes to documentation, I am of two minds.  There seem to be an increasing number of people coming to Python without much prior exposure to the concepts of threads and processes and so it would be wrong for us to ignore this reality.  On the flip side, attempting to convey all the core concepts of threads and processes and how they interact would result in a large documentation effort that ultimately few people would eagerly read to completion.  Adding a one-sentence caveat hiding somewhere in the docs won't do much to help.  Given this topic and a few other issues that have come up very recently, I suggest that a concise paragraph be added on the topic of using threads and processes together -- likely placed at the beginning of the docs on the Process class.  I think I'm up for taking a crack at that but I'd very much appreciate critical eyes to review it with me.

Per Eryk's point about the difference in multiprocessing's behavior when using spawn vs. fork, the explanation for why it's done that way is also described in the DeveloperWorks article I mentioned above.

Finally, per the original post from pietvo for future readers, not only is it *not* weird to start a Thread within a Process, it's a popular and useful technique.
msg270352 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2016-07-14 02:35
Devin, a primary point of `` is to provide a sane alternative to the cross-platform thread mess.  None of these reports are about making it easier for threads to go away "by magic" when the process ends.  It's the contrary:  they're asking for multiprocessing to respect's default behavior of making it the programmer's responsibility to shut down their threads cleanly.  "Shut down your threads, or we refuse to let the process end."

It doesn't matter that native OS threads may behave differently. very deliberately makes _its_ thread abstraction "non-daemonic" by default, and advertises that behavior for all platforms.  So it's at best surprising that threading.Thread's default semantics get turned inside out when multiprocessing creates a process.  I still see no reason to believe that's "a feature".

As to docs, if it boils down to the difference between `sys.exit()` and `os._exit()`, then _those_ are the places to put details, and multiprocessing docs just need to point out when a created process will use one or the other.

As is, the docs don't contain the slightest clue anywhere that a threading.Thread may violate its own docs (with respect to process-exit behavior) when created by a process launched by multiprocessing (or also by a concurrent.futures.ProcessPoolExecutor?  I didn't check).
msg270353 - (view) Author: Eryk Sun (eryksun) * Date: 2016-07-14 02:50
> Per Eryk's point about the difference in multiprocessing's behavior
> when using spawn vs. fork, the explanation for why it's done that
> way is also described in the DeveloperWorks article I mentioned above.

Please spell this out for me. Why can't the "fork" and "forkserver" variations call sys.exit(), and thus Py_Finalize? I don't see why it's OK to call the _shutdown function in Lib/ from a spawned child but not a forked child.
msg270355 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2016-07-14 03:24
About ""No parents, no children", that's fine so far as it goes.  But Python isn't C, a threading.Thread is not a POSIX thread, and _does_ have a concept of "the main thread".  There's no conceptual problem _in Python_ with saying "the main thread" waits to .join() other non-daemon threading.Threads at process exit.  No parent/child relationships are implied by that either - it's just the concept that one thread is distinguished.
msg270375 - (view) Author: Antoine Pitrou (pitrou) * (Python committer) Date: 2016-07-14 06:42
I agree with Tim. Regardless of what OS threads do, Python tries to enforce predictable semantics of its own. There's no reason (apart from historical baggage) to not join Python threads (and only Python threads, of course, not other OS threads) at the shutdown of a child process.

I don't exactly remember why using os._exit() rather than sys.exit() is required in child processes. Presumably it is because we don't want the child to clobber any resources shared with the parent (open files?). This doesn't have to be a binary thing, though: it may as well be os._exit() + a bunch of cleanup steps we know are safe to perform.
msg270382 - (view) Author: Charles-Fran├žois Natali (neologix) * (Python committer) Date: 2016-07-14 08:33
One reason for not calling sys.exit() is because on Linux, the default
implementation uses fork(), hence the address space in the chilren is
a clone of the parent: so all atexit handlers, for example, would be
called multiple times.
There's also the problem that fork() isn't MT-safe, making the
probability of screwups/deadlocks in various destructors/stack
unwinding greater.
msg270415 - (view) Author: Davin Potts (davin) * (Python committer) Date: 2016-07-14 15:15
Tim:  Totally agreed about threading.Thread not being a POSIX thread.  It was not my intent to suggest that they were equivalent -- apologies for the confusion.

Instead I was attempting to describe a mentality of processes and their common behavior across multiple platforms at termination.  The behavior of child processes via multiprocessing currently appears to follow this common mentality of signal the threads then exit quickly.  (To avoid confusion, I am making an observation here.)

Whereas threading.Thread is attempting to provide something homogeneous across platforms, achieving a similar goal in multiprocessing.Process is complicated by the concepts of fork vs. spawn and their availability on various OSes (a source of real confusion for some).  This further opens the question of what should the mentality be for multiprocessing.Process?  The notion that a process can die in such a way that not all of its threads were given time to clean up does not strike me as a foreign concept.  The notion that a threading.Thread should always be (or at least be attempted to be) joined makes sense.  The notion of categorically refusing to let a process end perhaps overreaches in certain situations.

I believe the more general solution exists in offering atexit handlers on multiprocessing.Process.
msg270419 - (view) Author: Eryk Sun (eryksun) * Date: 2016-07-14 15:58
> all atexit handlers, for example, would be called multiple times.

Davin is (I think) proposing a multiprocessing atexit facility, which can be used to ensure threading._shutdown is called. But could Python's regular atexit handling be reset in the child, allowing Py_Finalize to be called? In other words, can atexit can be integrated into the PyOS_AfterFork (Modules/signalmodule.c) sequence? multiprocessing could set a sys flag that forces atexit to clear its registered handlers, and for Py_AtExit, reset the static nexitfuncs variable in Python/pylifecycle.c. Or is that just opening a can of worms that will cause Py_Finalize to crash in various scenarios?

> There's also the problem that fork() isn't MT-safe

This issue is about joining Python threads created in the child, which has a clean slate via PyOS_AfterFork, PyEval_ReInitThreads (Python/ceval.c), and threading._after_fork.
msg270421 - (view) Author: R. David Murray (r.david.murray) * (Python committer) Date: 2016-07-14 16:36
As far as muliprocessing's "mentality" goes, it aims to provide the *same* API as Threading, so it is logical that it should preserve threading's behavior with respect to child threads in a process, rather than violating threading's model.  Anything else is counter-intuitive to a python programmer, as demonstrated by this issue and Tim's comments :)
msg270422 - (view) Author: R. David Murray (r.david.murray) * (Python committer) Date: 2016-07-14 16:37
Note, however, that fixing this will be a backward compatibility issue, since there are doubtless programs relying on this behavior, probably mostly unintentionally.
msg270433 - (view) Author: Tim Peters (tim.peters) * (Python committer) Date: 2016-07-14 18:46
About:  "The notion of categorically refusing to let a process end perhaps overreaches in certain situations." addressed that all along:  if the programmer _wants_ the process to exit without waiting for a particular threading.Thread, that's fine, they ask the Thread constructor for a "daemon" thread.  Whether a threading.Thread does or doesn't prevent process exit before it's .join()'ed has always been the programmer's choice.  Python never attempted to guess their intent (except for "the main thread", which is necessarily non-daemonic).  That's why it's especially surprising that multiprocessing can silently overrule what had always been an explicit choice about process-exit threading.Thread behavior.

About compatibility, yup, that's potentially painful.  I will note that compability was already broken on Windows with no apparent angst, or subsequent complaints (the program in issue 27508 is an example:  "runs forever" under 3.5.2 but "ends very quickly" under 2.7.11; "runs forever" is what the programmer wanted, matching how they expected non-daemon threading.Threads to work).
msg300385 - (view) Author: Antoine Pitrou (pitrou) * (Python committer) Date: 2017-08-16 18:53
New changeset ee84a608587b930176d37303afae8a4358e15990 by Antoine Pitrou in branch 'master':
bpo-18966: non-daemonic threads created by a multiprocessing.Process should be joined on exit (#3111)
msg300386 - (view) Author: Antoine Pitrou (pitrou) * (Python committer) Date: 2017-08-16 18:54
This is now fixed in git master.  Thank you for the report!
Date User Action Args
2017-08-16 18:54:02pitrousetstatus: open -> closed
resolution: fixed
messages: + msg300386

stage: patch review -> resolved
2017-08-16 18:53:31pitrousetmessages: + msg300385
2017-08-16 18:35:47pitrousetstage: needs patch -> patch review
versions: + Python 3.7, - Python 3.6
2017-08-16 18:27:28pitrousetpull_requests: + pull_request3150
2016-07-14 18:46:35tim.peterssetmessages: + msg270433
2016-07-14 16:37:57r.david.murraysetmessages: + msg270422
2016-07-14 16:36:43r.david.murraysetnosy: + r.david.murray
messages: + msg270421
2016-07-14 15:58:12eryksunsetmessages: + msg270419
2016-07-14 15:15:15davinsetmessages: + msg270415
2016-07-14 08:33:23neologixsetmessages: + msg270382
2016-07-14 06:42:23pitrousetstage: needs patch
type: behavior -> enhancement
versions: - Python 2.7, Python 3.5
2016-07-14 06:42:06pitrousetmessages: + msg270375
2016-07-14 03:24:32tim.peterssetmessages: + msg270355
2016-07-14 02:50:34eryksunsetmessages: + msg270353
2016-07-14 02:35:58tim.peterssetmessages: + msg270352
2016-07-14 01:54:46davinsetmessages: + msg270351
2016-07-13 22:28:09ned.deilysetnosy: + davin
2016-07-13 20:30:59eryksunsetnosy: + eryksun

messages: + msg270343
versions: + Python 3.5, Python 3.6, - Python 3.3, Python 3.4
2016-07-13 20:05:05tim.peterssetnosy: + tim.peters
messages: + msg270338
2016-07-13 19:58:10tim.peterslinkissue27508 superseder
2013-09-08 17:40:56pitrousetnosy: + neologix, pitrou, sbt

messages: + msg197293
versions: + Python 3.4
2013-09-08 01:20:27pietvocreate