How to properly use dask's upload_file() to pass local code to workers

Question

I have functions in a local_code.py file that I would like to pass to workers through dask. I've seen answers to questions on here saying that this can be done using the upload_file() function, but I can't seem to get it working because I'm still getting a ModuleNotFoundError.

The relevant part of the code is as follows.

from dask.distributed import Client
from dask_jobqueue import SLURMCluster

from local_code import *
helper_file = '/absolute/path/to/local_code.py'

def main():
    with SLURMCluster(**slurm_params) as cluster:

        cluster.scale(n_workers)

        with Client(cluster) as client:
            client.upload_file(helper_file)
            mapping = client.map(myfunc, data)
            client.gather(mapping)

if __name__ == '__main__':
    main()

Note, myfunc is imported from local_code, and there's no error importing it to map. The function myfunc also depends on other functions that are defined in local_code.

With this code, I'm still getting this error

distributed.protocol.pickle - INFO - Failed to deserialize b'\x80\x04\x95+\x00\x00\x00\x00\x00\x00\x00\x8c\x11local_code\x94\x8c\x$
Traceback (most recent call last):
  File "/home/gallagher.r/.local/lib/python3.7/site-packages/distributed/protocol/pickle.py", line 61, in loads
    return pickle.loads(x)
ModuleNotFoundError: No module named 'local_code'

Using upload_file() seems so straightforward that I'm not sure what I'm doing wrong. I must have it in the wrong place or not be understanding correctly what is passed to it.

I'd appreciate any help with this. Please let me know if you need any other information or if there's anything else I can supply from the error file.

score 7 · Accepted Answer · answered Jul 21 '19 at 04:51

7

The upload_file method only uploads the file to the currently available workers. If a worker arrives after you call upload_file then that worker won't have the provided file.

If your situation the easiest thing to do is probably to wait until all of the workers arrive before you call upload file

cluster.scale(n)
with Client(cluster) as client:
    client.wait_for_workers(n)
    client.upload_file(...)

answered Jul 21 '19 at 04:51

MRocklin

55,641
23
163
235

3

Has anyone figured out how to call upload_file as part of worker registration? – Jeffrey Hulten Dec 04 '19 at 21:34
1

client.upload_file doesn't seem to work with FargateCluster in AWS. It uploads the file to worker and also get the message "Reload module " in worker logs. But, still I get error "No module found" in scheduler logs. – jsanjayce May 26 '23 at 12:50

score 3 · Answer 2 · answered May 06 '20 at 20:56

Another option when you have workers going in/out is to use the Client.register_worker_callbacks to hook into whenever a new worker is registered/added. The one caveat is you will need to serialize your file(s) in the callback partial:

fname = ...
with open(fname, 'rb') as f:
  data = f.read()

client.register_worker_callbacks(
  setup=functools.partial(
    _worker_upload, data=data, fname=fname,
  )
)

def _worker_upload(dask_worker, *, data, fname):
  dask_worker.loop.add_callback(
    callback=dask_worker.upload_file,
    comm=None,  # not used
    filename=fname,
    data=data,
    load=True)

This will also upload the file the first time the callback is registered so you can avoid calling client.upload_file entirely.

How to properly use dask's upload_file() to pass local code to workers

2 Answers2

Linked