What to do when threads_per_worker is 0

[jsignell]

tl;dr Maybe threads_per_worker should throw an error if set to 0.

While working on #3984 I ran into some odd behaviors when threads_per_worker is set to 0 in LocalCluster. I am not sure what the intended behavior should be or if this PR makes things better.

Some examples:

On master the following results in a divide by 0 error. This PR sets n_workers to 0 which I think is a good solution:

from distributed import LocalCluster, Client
cluster = LocalCluster(threads_per_worker=0)
client = Client(cluster)

client
# <Client: 'tcp://127.0.0.1:46121' processes=0 threads=0, memory=0 B> 

For the next case, I am not sure that the behavior should be. Here the nthreads is being passed int the worker_kwargs as 0. So we don't really have control over what happens after that. In this case it sets nthreads for each worker to CPU_COUNT

from distributed import LocalCluster, Client

cluster = LocalCluster(n_workers=2, threads_per_worker=0)
client = Client(cluster)

client
# <Client: 'tcp://127.0.0.1:46509' processes=2 threads=24, memory=33.37 GB>

[jsignell]

@gforsyth who I was looking at this with.

[mrocklin]

I think that we should raise if threads_per_worker < 1

When testing in the distributed repository async tests are preferred, if they're easy to use. You'll also want to use the cleanup fixture.

[jsignell]

I think that we should raise if threads_per_worker < 1

That makes sense to me and should result in unambiguous behavior.

[jsignell]

@mrocklin I made it a warning instead of an error since it is easy enough to set it to None.

[mrocklin]

Thanks @jsignell ! Merged.