-
-
Notifications
You must be signed in to change notification settings - Fork 30.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
multiprocessing's default posix start method of 'fork'
is broken: change to `'forkserver' || 'spawn'
#84559
Comments
By default, multiprocessing uses fork() without exec() on POSIX. For a variety of reasons this can lead to inconsistent state in subprocesses: module-level globals are copied, which can mess up logging, threads don't survive fork(), etc.. The end results vary, but quite often are silent lockups. In real world usage, this results in users getting mysterious hangs they do not have the knowledge to debug. The fix for these people is to use "spawn" by default, which is the default on Windows. Just a small sample:
I suggest changing the default on POSIX to match Windows. |
Looks like as of 3.8 this only impacts Linux/non-macOS-POSIX, so I'll amend the above to say this will also make it consistent with macOS. |
Just got an email from someone for whom switching to "spawn" fixed a problem. Earlier this week someone tweeted about this fixing things. This keeps hitting people in the real world. |
Another person with the same issue: https://twitter.com/volcan01010/status/1324764531139248128 |
I just ran into and fixed (thanks to itamarst's blog post) a problem likely related to this. Multiprocessing workers performing work and sending a logging message back with success/fail info. I had a few intermittent deadlocks that became a recurring problem when I sped up the process that skipped tasks which had previously completed (I think this shortened the time between forking and attempting to send messages causing the third process to deadlock). After changing that it deadlocked *every time*. Switching to "spawn" at the top of the main function has fixed it. |
The problem with changing the default is that this will break any application that depends on passing non-picklable data to the child process (in addition to the potentially unexpected performance impact). The docs already contain a significant elaboration on the matter, but feel free to submit a PR that would make the various caveats more explicit: |
This change was made on macOS at some point, so why not Linux? "spawn" is already the default on macOS and Windows. |
The macOS change was required before "fork" simply ceased to work. |
Given people's general experience, I would not say that "fork" works on Linux either. More like "99% of the time it works, 1% it randomly breaks in mysterious way". |
Agreed, but again, changing will break some applications. We could switch to forkserver, but we should have a transition period where a FutureWarning will be displayed if people didn't explicitly set a start method. |
After updating PyPy3 to use Python 3.9's stdlib, we hit very bad hangs because of this — literally compiling a single file with "parallel" compileall could hang. In the end, we had to revert the change in how Python 3.9 starts workers because otherwise multiprocessing would be impossible to use: https://foss.heptapod.net/pypy/pypy/-/commit/c594b6c48a48386e8ac1f3f52d4b82f9c3e34784 This is a very bad default and what's even worse is that it often causes deadlocks that are hard to reproduce or debug. Furthermore, since "fork" is the default, people are unintentionally relying on its support for passing non-pickleable projects and are creating non-portable code. The code often becomes complex and hard to change before they discover the problem. Before we managed to figure out how to workaround the deadlocks in PyPy3, we were experimenting with switching the default to "spawn". Unfortunately, we've hit multiple projects that didn't work with this method, precisely because of pickling problems. Furthermore, they were surprised to learn that their code wouldn't work on macOS (in the end, many people perceive Python as a language for writing portable software). Finally, back in 2018 I've made one of my projects do parallel work using multiprocessing. It gave its users great speedup but for some it caused deadlocks that I couldn't reproduce nor debug. In the end, I had to revert it. Now that I've learned about this problem, I'm wondering if this wasn't precisely because of "fork" method. |
Provide a way for the calling code to specify which "multiprocessing context" to use to spawn subprocesses. See https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods I'm using this to allow us to mock out multiprocessing with multithreading in doctests. This will also let you more easily test differences between "spawn" and "fork" modes. I'm defaulting to using "spawn" because I think "fork" mode was the cause of some mysterious hanging in tests. General consensus seems to be "spawn" is less buggy: python/cpython#84559 I've felt like tests are consistently faster with it. Also uses the `multiprocessing.Manager` as a context manager so it gets cleaned up correctly. This might have been the cause of other hanging in local cluster execution.
Provide a way for the calling code to specify which "multiprocessing context" to use to spawn subprocesses. See https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods I'm using this to allow us to mock out multiprocessing with multithreading in doctests. This will also let you more easily test differences between "spawn" and "fork" modes. I'm defaulting to using "spawn" because I think "fork" mode was the cause of some mysterious hanging in tests. General consensus seems to be "spawn" is less buggy: python/cpython#84559 I've felt like tests are consistently faster with it. Also uses the `multiprocessing.Manager` as a context manager so it gets cleaned up correctly. This might have been the cause of other hanging in local cluster execution.
Another example: Nelson Elhage reports that "as of recently(?) pytorch silently deadlocks (even without GPUs involved at all) using method=fork so that's been fun to debug". Examples he provided:
|
After updating a couple of libraries in a project we are working on, the code would hang without much explanation. After much debugging, I think one of the reasons for our issues is the forking default (this issue). Our business logic does not use multiprocessing, but the underlying execution engine does (in our case Luigi). Turns out that gRPC client (which was buried deep into one of our dependencies) can hang in some cases when forked grpc/grpc#18075. This was the case for us, and was very tricky to debug. |
spawn
general plan:
|
In order to migrate away from unsafe os.fork() usage in threaded processes (python/cpython#84559), add a returnproc parameter that is similar to returnpid, which causes spawn to return Process objects instead of pids. The Process API is a subset of asyncio.subprocess.Process. In the future, spawn will return objects of a different type but with a compatible interface to Process, in order to encapsulate implementation-dependent objects like multiprocessing.Process which are designed to manage the process lifecycle and need to persist until it exits. Trigger a UserWarning when the returnpid parameter is used, in order to encourage migration to returnproc (do not use DeprecationWarning since it is hidden by default). This warning will be temporarily suppressed for portage internals, until they finish migrating to returnproc. There are probably very few if any external consumers of spawn with the returnpid parameter, so it seems safe to move quickly with this deprecation. Bug: https://bugs.gentoo.org/916566 Signed-off-by: Zac Medico <[email protected]>
In order to migrate away from unsafe os.fork() usage in threaded processes (python/cpython#84559), add a returnproc parameter that is similar to returnpid, which causes spawn to return Process objects instead of pids. The Process API is a subset of asyncio.subprocess.Process. In the future, spawn will return objects of a different type but with a compatible interface to Process, in order to encapsulate implementation-dependent objects like multiprocessing.Process which are designed to manage the process lifecycle and need to persist until it exits. Trigger a UserWarning when the returnpid parameter is used, in order to encourage migration to returnproc (do not use DeprecationWarning since it is hidden by default). This warning will be temporarily suppressed for portage internals, until they finish migrating to returnproc. There are probably very few if any external consumers of spawn with the returnpid parameter, so it seems safe to move quickly with this deprecation. Bug: https://bugs.gentoo.org/916566 Signed-off-by: Zac Medico <[email protected]>
In order to migrate away from unsafe os.fork() usage in threaded processes (python/cpython#84559), add a returnproc parameter that is similar to returnpid, which causes spawn to return a single Process object instead of a list of pids. The Process API is a subset of asyncio.subprocess.Process. The returnproc parameter conflicts with the logfile parameter, since the caller is expected to use the fd_pipes parameter to implement logging (this was also true for the returnpid parameter). In the future, spawn will return objects of a different type but with a compatible interface to Process, in order to encapsulate implementation-dependent objects like multiprocessing.Process which are designed to manage the process lifecycle and need to persist until it exits. Trigger a UserWarning when the returnpid parameter is used, in order to encourage migration to returnproc (do not use DeprecationWarning since it is hidden by default). This warning will be temporarily suppressed for portage internals, until they finish migrating to returnproc. There are probably very few if any external consumers of spawn with the returnpid parameter, so it seems safe to move quickly with this deprecation. Bug: https://bugs.gentoo.org/916566 Signed-off-by: Zac Medico <[email protected]>
The article is now https://pythonspeed.com/articles/python-multiprocessing/ |
Use forkserver when available to avoid issues with forking multi-threaded processes. See python/cpython#84559 for more context. This also removes a warning when running on python 3.12: ``` /opt/hostedtoolcache/Python/3.12.2/x64/lib/python3.12/multiprocessing/popen_fork.py:66: DeprecationWarning: This process (pid=1991) is multi-threaded, use of fork() may lead to deadlocks in the child. self.pid = os.fork() ```
Use forkserver when available to avoid issues with forking multi-threaded processes. See [1] for more context. This also removes a warning when running on python 3.12 that can be seen in CI: > /opt/hostedtoolcache/Python/3.12.2/x64/lib/python3.12/multiprocessing/popen_fork.py:66: DeprecationWarning: This process (pid=1991) is multi-threaded, use of fork() may lead to deadlocks in the child. > self.pid = os.fork() [1] python/cpython#84559
Use forkserver when available to avoid issues with forking multi-threaded processes. See [1] for more context. This also removes a warning when running on python 3.12 that can be seen in CI: > /opt/hostedtoolcache/Python/3.12.2/x64/lib/python3.12/multiprocessing/popen_fork.py:66: DeprecationWarning: This process (pid=1991) is multi-threaded, use of fork() may lead to deadlocks in the child. > self.pid = os.fork() [1] python/cpython#84559
Provide a way for the calling code to specify which "multiprocessing context" to use to spawn subprocesses. See https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods I'm using this to allow us to mock out multiprocessing with multithreading in doctests. This will also let you more easily test differences between "spawn" and "fork" modes. I'm defaulting to using "spawn" because I think "fork" mode was the cause of some mysterious hanging in tests. General consensus seems to be "spawn" is less buggy: python/cpython#84559 I've felt like tests are consistently faster with it. Also uses the `multiprocessing.Manager` as a context manager so it gets cleaned up correctly. This might have been the cause of other hanging in local cluster execution.
Provide a way for the calling code to specify which "multiprocessing context" to use to spawn subprocesses. See https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods I'm using this to allow us to mock out multiprocessing with multithreading in doctests. This will also let you more easily test differences between "spawn" and "fork" modes. I'm defaulting to using "spawn" because I think "fork" mode was the cause of some mysterious hanging in tests. General consensus seems to be "spawn" is less buggy: python/cpython#84559 I've felt like tests are consistently faster with it. Also uses the `multiprocessing.Manager` as a context manager so it gets cleaned up correctly. This might have been the cause of other hanging in local cluster execution.
The default of `fork` is known to be problematic. Python itself is changing the default to `spawn`. The new default is expected to be in place for Python 3.14. Python references for the change to the default: * python/cpython#84559 * python/cpython#100618 We also have several places where this option had to be set to `spawn` to make tests work. The AMD code even checks and overrides the value if it's not set to `spawn`. Simplify things for everyone and just default to `spawn`, but leave the option in place just in case, at least for now. Signed-off-by: Russell Bryant <[email protected]>
'fork'
is broken: change to 'spawn'
'fork'
is broken: change to `'forkserver' || 'spawn'
…ver` (GH-101556) Change the default multiprocessing start method away from fork to forkserver or spawn on the remaining platforms where it was fork. See the issue for context. This makes the default far more thread safe (other than for people spawning threads at import time... - don't do that!). Co-authored-by: blurb-it[bot] <43283697+blurb-it[bot]@users.noreply.github.com> Co-authored-by: Hugo van Kemenade <[email protected]>
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:
bugs.python.org fields:
Linked PRs
forkserver
#101556The text was updated successfully, but these errors were encountered: