Albert DeFusco Albert DeFusco - 11 months ago 60
Python Question

Can I use functions imported from .py files in Dask/Distributed?

I have a question about serialization and imports.

  • should functions have their own imports? like I've seen done with PySpark

  • Is the following just plain wrong? Does
    need to be a conda/pip package?
    was written to a shared filesystem.

In [1]: from distributed import Executor

In [2]: e = Executor('')

In [3]: e
Out[3]: <Executor: scheduler="" processes=2 cores=2>

In [4]: import socket

In [5]:
Out[5]: {'': 'n1015', '': 'n1016'}

In [6]: %%file
...: def hostname():
...: return 'the hostname'

In [7]: import mod

In [8]: mod.hostname()
Out[8]: 'the hostname'

In [9]:
distributed.utils - ERROR - No module named 'mod'

Answer Source

Quick Answer

Upload your file to all of your workers. You can do this using whatever mechanism you used to set up dask.distributed, or you can use the upload_file method


Alternatively, if your function is made in IPython, rather than being part of a module, it will be sent along without a problem.

Long Answer

This all has to do with how functions get serialized in Python. Functions from modules are serialized by their module name and function name

In [1]: from math import sin

In [2]: import pickle

In [3]: pickle.dumps(sin)
Out[3]: b'\x80\x03cmath\nsin\nq\x00.'

So if the client machine wants to refer to the math.sin function it sends along this bytestring (which you'll notice has 'math' and 'sin' in it buried among other bytes) to the worker machine. The worker looks at this bytestring and says "OK great, the function I want is in such and such a module, let me go and find that in my local file system. If the module isn't present then it'll raise an error, much like what you received above.

For dynamically created functions (functions that you make in IPython) it uses a completely different approach, bundling up all of the code. This approach generally works fine.

Generally speaking Dask assumes that the workers and the client all have the same software environment. Typically this is mostly handled by whoever sets up your cluster, using some other tool like Docker. Methods like upload_file are there to fill in the gaps when you have files or scripts that get updated more frequently.