在joblib`Parallel`上下文中对`matlab`对象进行腌制时出错

Question

I'm running some Matlab code in parallel from inside a Python context (I know, but that's what's going on), and I'm hitting an import error involving matlab.double . The same code works fine in a multiprocessing.Pool , so I am having trouble figuring out what the problem is. Here's a minimal reproducing test case.

import matlab
from multiprocessing import Pool
from joblib import Parallel, delayed

# A global object that I would like to be available in the parallel subroutine
x = matlab.double([[0.0]])

def f(i):
    print(i, x)

with Pool(4) as p:
    p.map(f, range(10))
    # This prints 1, [[0.0]]\n2, [[0.0]]\n... as expected

for _ in Parallel(4, backend='multiprocessing')(delayed(f)(i) for i in range(10)):
    pass
# This also prints 1, [[0.0]]\n2, [[0.0]]\n... as expected

# Now run with default `backend='loky'`
for _ in Parallel(4)(delayed(f)(i) for i in range(10)):
    pass
# ^ this crashes.

So, the only problematic one is the one using the 'loky' backend. The full traceback is:

exception calling callback for <Future at 0x7f63b5a57358 state=finished raised BrokenProcessPool>
joblib.externals.loky.process_executor._RemoteTraceback: 
'''
Traceback (most recent call last):
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/process_executor.py", line 391, in _process_worker
    call_item = call_queue.get(block=True, timeout=timeout)
  File "~/miniconda3/envs/myenv/lib/python3.6/multiprocessing/queues.py", line 113, in get
    return _ForkingPickler.loads(res)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/mlarray.py", line 31, in <module>
    from _internal.mlarray_sequence import _MLArrayMetaClass
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/_internal/mlarray_sequence.py", line 3, in <module>
    from _internal.mlarray_utils import _get_strides, _get_size, \
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/_internal/mlarray_utils.py", line 4, in <module>
    import matlab
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/__init__.py", line 24, in <module>
    from mlarray import double, single, uint8, int8, uint16, \
ImportError: cannot import name 'double'
'''

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

Traceback (most recent call last):
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/_base.py", line 625, in _invoke_callbacks
    callback(self)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 309, in __call__
    self.parallel.dispatch_next()
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 731, in dispatch_next
    if not self.dispatch_one_batch(self._original_iterator):
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 759, in dispatch_one_batch
    self._dispatch(tasks)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 716, in _dispatch
    job = self._backend.apply_async(batch, callback=cb)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/_parallel_backends.py", line 510, in apply_async
    future = self._workers.submit(SafeFunction(func))
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/reusable_executor.py", line 151, in submit
    fn, *args, **kwargs)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/process_executor.py", line 1022, in submit
    raise self._flags.broken
joblib.externals.loky.process_executor.BrokenProcessPool: A task has failed to un-serialize. Please ensure that the arguments of the function are all picklable.
joblib.externals.loky.process_executor._RemoteTraceback: 
'''
Traceback (most recent call last):
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/process_executor.py", line 391, in _process_worker
    call_item = call_queue.get(block=True, timeout=timeout)
  File "~/miniconda3/envs/myenv/lib/python3.6/multiprocessing/queues.py", line 113, in get
    return _ForkingPickler.loads(res)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/mlarray.py", line 31, in <module>
    from _internal.mlarray_sequence import _MLArrayMetaClass
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/_internal/mlarray_sequence.py", line 3, in <module>
    from _internal.mlarray_utils import _get_strides, _get_size, \
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/_internal/mlarray_utils.py", line 4, in <module>
    import matlab
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/matlab/__init__.py", line 24, in <module>
    from mlarray import double, single, uint8, int8, uint16, \
ImportError: cannot import name 'double'
'''

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

Traceback (most recent call last):
  File "test.py", line 20, in <module>
    for _ in Parallel(4)(delayed(f)(i) for i in range(10)):
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 934, in __call__
    self.retrieve()
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 833, in retrieve
    self._output.extend(job.get(timeout=self.timeout))
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/_parallel_backends.py", line 521, in wrap_future_result
    return future.result(timeout=timeout)
  File "~/miniconda3/envs/myenv/lib/python3.6/concurrent/futures/_base.py", line 432, in result
    return self.__get_result()
  File "~/miniconda3/envs/myenv/lib/python3.6/concurrent/futures/_base.py", line 384, in __get_result
    raise self._exception
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/_base.py", line 625, in _invoke_callbacks
    callback(self)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 309, in __call__
    self.parallel.dispatch_next()
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 731, in dispatch_next
    if not self.dispatch_one_batch(self._original_iterator):
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 759, in dispatch_one_batch
    self._dispatch(tasks)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/parallel.py", line 716, in _dispatch
    job = self._backend.apply_async(batch, callback=cb)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/_parallel_backends.py", line 510, in apply_async
    future = self._workers.submit(SafeFunction(func))
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/reusable_executor.py", line 151, in submit
    fn, *args, **kwargs)
  File "~/miniconda3/envs/myenv/lib/python3.6/site-packages/joblib/externals/loky/process_executor.py", line 1022, in submit
    raise self._flags.broken
joblib.externals.loky.process_executor.BrokenProcessPool: A task has failed to un-serialize. Please ensure that the arguments of the function are all picklable.

Looking at the traceback, it seems like the root cause is an issue importing the matlab package in the child process.

It's probably worth noting that this all runs just fine if instead I had defined x = np.array([[0.0]]) (after importing numpy as np ). And of course the main process has no problem with any matlab imports, so I am not sure why the child process would.

I'm not sure if this error has anything in particular to do with the matlab package, or if it's something to do with global variables and cloudpickle or loky . In my application it would help to stick with loky , so I'd appreciate any insight!

I should also note that I'm using the official Matlab engine for Python: https://www.mathworks.com/help/matlab/matlab-engine-for-python.html . I suppose that might make it hard for others to try out the test cases, so I wish I could reproduce this error with a type other than matlab.double , but I haven't found another yet.

Digging around more, I've noticed that the process of importing the matlab package is more circular than I would expect, and I'm speculating that this could be part of the problem? The issue is that when import matlab is run by loky 's _ForkingPickler , first some file matlab/mlarray.py is imported, which imports some other files, one of which contains import matlab , and this causes matlab/__init__.py to be run, which internally has from mlarray import double, single, uint8, ... which is the line that causes the crash.

Could this circularity be the issue? If so, why can I import this module in the main process but not in the loky backend?

Answer 1

The error is caused by incorrect loading order of global objects in the child processes. It can be seen clearly in the traceback _ForkingPickler.loads(res) -> ... -> import matlab -> from mlarray import ... that matlab is not yet imported when the global variable x is loaded by cloudpickle .

joblib with loky seems to treat modules as normal global objects and send them dynamically to the child processes. joblib doesn't record the order in which those objects/modules were defined. Therefore they are loaded (initialized) in a random order in the child processes.

A simple workaround is to manually pickle the matlab object and load it after importing matlab inside your function.

import matlab
import pickle

px = pickle.dumps(matlab.double([[0.0]]))

def f(i):
    import matlab
    x=pickle.loads(px)
    print(i, x)

Of course you can also use the joblib.dumps and loads to serialize the objects.

Use initializer

Thanks to the suggestion of @Aaron, you can also use an initializer ( for loky ) to import Matlab before loading x .

Currently there's no simple API to specify initializer . So I wrote a simple function:

def with_initializer(self, f_init):
    # Overwrite initializer hook in the Loky ProcessPoolExecutor
    # https://github.com/tomMoral/loky/blob/f4739e123acb711781e46581d5ed31ed8201c7a9/loky/process_executor.py#L850
    hasattr(self._backend, '_workers') or self.__enter__()
    origin_init = self._backend._workers._initializer
    def new_init():
        origin_init()
        f_init()
    self._backend._workers._initializer = new_init if callable(origin_init) else f_init
    return self

It is a little bit hacky but works well with the current version of joblib and loky. Then you can use it like:

import matlab
from joblib import Parallel, delayed

x = matlab.double([[0.0]])

def f(i):
    print(i, x)

def _init_matlab():
    import matlab

with Parallel(4) as p:
    for _ in with_initializer(p, _init_matlab)(delayed(f)(i) for i in range(10)):
        pass

I hope the developers of joblib will add initializer argument to the constructor of Parallel in the future.