delayed does not work with local_client

Using delayed to launch tasks from workers does not always work. As reproducible example, I have modified the fibonacci example to use delayed (fib_2 function) instead of submit/gather (fib function).:

from distributed import Client, local_client

def fib(n):
    # this is using the submit idiom, which is working
    if n < 2:
        return n
    else:
        with local_client() as lc:
            a = lc.submit(fib, n - 1)
            b = lc.submit(fib, n - 2)
            a, b = lc.gather([a, b])
            return a + b

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

def fib_2(n):
    # this is using delayed, which is not working
    from dask import delayed
    if n < 2:
        return n
    else:
        with local_client() as lc:
            a = delayed(fib_2)(n - 1)
            b = delayed(fib_2)(n - 2)
            s = delayed(add)(a, b)
            r = s.compute()
            return r

def fib_delayed():
    # get a delayed fib calculation, to make sure the local_client is only spawned from a worker
    from dask import delayed
    d = delayed(fib_2)(10)
    return d

if __name__ == '__main__':
    c = Client()
    fib_d = fib_delayed()
    print(fib_d.dask) # should only have 1 key, the rest is submitted from the worker
    print(fib_d.compute())

running this gives me lots of “Compute failed” errors and a an endless traceback which I’ll omit here.

[vincentschut]

So this does work when I replace “s.compute()” with “lc.compute(s); r=s.result()”. I always assumed delayed.compute() to be a shortcut of “f=client.compute(d); f.result()” but apparently that is not true within the context of local_client(). Thus I think we can narrow this down to “delayed.compute()” does not work with local_client".

[GenevieveBuckley]

It looks like things have changed since this issue was opened, I don’t get this problem happening today.

When I run the example with a more recent version of dask/distributed (2021.09.0) I don’t see any errors. (It does tell me that local_client has been moved/renamed to worker_client, but that’s not a problem).