dask worker: daemonic processes are not allowed to have children

[costrouc]

I have a package that I have written that used multiprocessing and some complex communication with multiprocessing.Pipe. I run into an issue when trying to using multiprocessing within dask workers. Is there any workaround that I could do? Is there a way to run a dask worker not as a daemon?

Minimum working example.

dask-scheduler

dask-worker <scheduler-ip:scheduler-port>

def test_process():
    import multiprocessing

    def adder(a, b):
        return a + b

    p = multiprocessing.Process(target=adder, args=(1, 3))
    p.start()
    p.join()

if __name__ == "__main__":
    from dask.distributed import Client
    client = Client('<scheduler-ip:scheduler-port>')

    f = client.submit(test_process)
    f.result()

Error message

distributed.worker - WARNING -  Compute Failed
Function:  test_process
args:      ()
kwargs:    {}
Exception: AssertionError('daemonic processes are not allowed to have children',)

Traceback (most recent call last):
  File "test.py", line 16, in <module>
    f.result()
  File "/home/costrouc/.virtualenvs/dask-6qByegfR/lib/python3.6/site-packages/distributed/client.py", line 192, in result
    six.reraise(*result)
  File "/home/costrouc/.virtualenvs/dask-6qByegfR/lib/python3.6/site-packages/six.py", line 692, in reraise
    raise value.with_traceback(tb)
  File "test.py", line 8, in test_process
    p.start()
  File "/usr/lib/python3.6/multiprocessing/process.py", line 103, in start
    'daemonic processes are not allowed to have children'
AssertionError: daemonic processes are not allowed to have children

[mrocklin]

Your minimal example is odd. You create a dask-worker and dask-scheduler but you also create a separate local cluster in the Client() call.

However, I suspect that what you're after is the --no-nanny flag to dask-worker or the processes=False keyword to Client, depending on how you actually create your cluster.

[costrouc]

Thank you for correcting my misunderstanding. I have updated my comment above to show the correct Client('<scheduler-ip:port>') since I was interested in the distributed case.

Yes --no-nanny works for the Distributed cluster and processes=False works for the Local cluster. Issue is resolved.

[zhanghang1989]

I am new to dask. Is that possible to set --no-nanny when using dask-ssh?