Serialization of un-picklable objects

This example highlights the options for tempering with joblib serialization process.

# Code source: Thomas Moreau
# License: BSD 3 clause

import sys
import time
import traceback
from joblib.externals.loky import set_loky_pickler
from joblib import parallel_config
from joblib import Parallel, delayed
from joblib import wrap_non_picklable_objects

First, define functions which cannot be pickled with the standard pickle protocol. They cannot be serialized with pickle because they are defined in the __main__ module. They can however be serialized with cloudpickle. With the default behavior, loky is to use cloudpickle to serialize the objects that are sent to the workers.

def func_async(i, *args):
    return 2 * i


print(Parallel(n_jobs=2)(delayed(func_async)(21) for _ in range(1))[0])

42

For most use-cases, using cloudpickle is efficient enough. However, this solution can be very slow to serialize large python objects, such as dict or list, compared to the standard pickle serialization.

def func_async(i, *args):
    return 2 * i


# We have to pass an extra argument with a large list (or another large python
# object).
large_list = list(range(1000000))

t_start = time.time()
Parallel(n_jobs=2)(delayed(func_async)(21, large_list) for _ in range(1))
print("With loky backend and cloudpickle serialization: {:.3f}s"
      .format(time.time() - t_start))

With loky backend and cloudpickle serialization: 0.072s

If you are on a UNIX system, it is possible to fallback to the old multiprocessing backend, which can pickle interactively defined functions with the default pickle module, which is faster for such large objects.

import multiprocessing as mp
if mp.get_start_method() != "spawn":
    def func_async(i, *args):
        return 2 * i

    with parallel_config('multiprocessing'):
        t_start = time.time()
        Parallel(n_jobs=2)(
            delayed(func_async)(21, large_list) for _ in range(1))
        print("With multiprocessing backend and pickle serialization: {:.3f}s"
              .format(time.time() - t_start))

With multiprocessing backend and pickle serialization: 0.156s

However, using fork to start new processes can cause violation of the POSIX specification and can have bad interaction with compiled extensions that use openmp. Also, it is not possible to start processes with fork on windows where only spawn is available. The loky backend has been developed to mitigate these issues.

To have fast pickling with loky, it is possible to rely on pickle to serialize all communications between the main process and the workers with the loky backend. This can be done by setting the environment variable LOKY_PICKLER=pickle before the script is launched. Here we use an internal programmatic switch loky.set_loky_pickler for demonstration purposes but it has the same effect as setting LOKY_PICKLER. Note that this switch should not be used as it has some side effects with the workers.

# Now set the `loky_pickler` to use the pickle serialization from stdlib. Here,
# we do not pass the desired function ``func_async`` as it is not picklable
# but it is replaced by ``id`` for demonstration purposes.

set_loky_pickler('pickle')
t_start = time.time()
Parallel(n_jobs=2)(delayed(id)(large_list) for _ in range(1))
print("With pickle serialization: {:.3f}s".format(time.time() - t_start))

With pickle serialization: 0.071s

However, the function and objects defined in __main__ are not serializable anymore using pickle and it is not possible to call func_async using this pickler.

def func_async(i, *args):
    return 2 * i


try:
    Parallel(n_jobs=2)(delayed(func_async)(21, large_list) for _ in range(1))
except Exception:
    traceback.print_exc(file=sys.stdout)

joblib.externals.loky.process_executor._RemoteTraceback:
"""
Traceback (most recent call last):
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/externals/loky/process_executor.py", line 426, in _process_worker
    call_item = call_queue.get(block=True, timeout=timeout)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/docs/.asdf/installs/python/3.11.9/lib/python3.11/multiprocessing/queues.py", line 122, in get
    return _ForkingPickler.loads(res)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: Can't get attribute 'func_async' on <module 'joblib.externals.loky.backend.popen_loky_posix' from '/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/externals/loky/backend/popen_loky_posix.py'>
"""

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

Traceback (most recent call last):
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/checkouts/latest/examples/serialization_and_wrappers.py", line 114, in <module>
    Parallel(n_jobs=2)(delayed(func_async)(21, large_list) for _ in range(1))
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/parallel.py", line 2007, in __call__
    return output if self.return_generator else list(output)
                                                ^^^^^^^^^^^^
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/parallel.py", line 1650, in _get_outputs
    yield from self._retrieve()
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/parallel.py", line 1754, in _retrieve
    self._raise_error_fast()
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/parallel.py", line 1789, in _raise_error_fast
    error_job.get_result(self.timeout)
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/parallel.py", line 745, in get_result
    return self._return_or_raise()
           ^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/docs/checkouts/readthedocs.org/user_builds/joblib/envs/latest/lib/python3.11/site-packages/joblib/parallel.py", line 763, in _return_or_raise
    raise self._result
joblib.externals.loky.process_executor.BrokenProcessPool: A task has failed to un-serialize. Please ensure that the arguments of the function are all picklable.

To have both fast pickling, safe process creation and serialization of interactive functions, joblib provides a wrapper function wrap_non_picklable_objects() to wrap the non-picklable function and indicate to the serialization process that this specific function should be serialized using cloudpickle. This changes the serialization behavior only for this function and keeps using pickle for all other objects. The drawback of this solution is that it modifies the object. This should not cause many issues with functions but can have side effects with object instances.

@delayed
@wrap_non_picklable_objects
def func_async_wrapped(i, *args):
    return 2 * i


t_start = time.time()
Parallel(n_jobs=2)(func_async_wrapped(21, large_list) for _ in range(1))
print("With pickle from stdlib and wrapper: {:.3f}s"
      .format(time.time() - t_start))

With pickle from stdlib and wrapper: 0.337s

The same wrapper can also be used for non-picklable classes. Note that the side effects of wrap_non_picklable_objects on objects can break magic methods such as __add__ and can mess up the isinstance and issubclass functions. Some improvements will be considered if use-cases are reported.

# Reset the loky_pickler to avoid border effects with other examples in
# sphinx-gallery.
set_loky_pickler()