How the the computation code is transferreed to dask worker from the client?

How does the client side source code get transferred to dask worker(s) ?

from dask.distributed  import Client


def hello3(a, b):
    print('I am hello 3')
    return a + 5 + b * 3


def hello2(a, b):
    print('I am hello2')
    return hello3(a * 10, b + 5)


def xadd(a, b):
    return hello2(a + 2, b * 5)

if __name__ == '__main__':
    client = Client('127.0.0.1:8786')
    x = client.submit(xadd, 1, 2)
    print(x.result())

In the above code snippet, how will submit transfer code for xadd, hello2 and hello3 functions to the workers ?

Dask uses pickle for sending most things from the client the scheduler, which then sends stuff on to workers. In the case of code defined in __main__, you will (very likely) end up falling back to cloudpickle, which is a similar protocol that's able to cope with a wider variety of situations.

> Among other things, cloudpickle supports pickling for lambda functions along with functions and classes defined interactively in the __main__ module (for instance in a script, a shell or a Jupyter notebook).

( https://github.com/cloudpipe/cloudpickle#readme )

I'm not certain that the following is a correct description, but might give some useful leads:

Upon submission of the function and arguments via client.submit, dask will serialize the function and arguments (as well as any additional kwargs). The description of the serialization is provided in the docs.

The specific function used to extract the function code is the _get_computation_code method.