I find myself often wanting to run code on a worker, rather than on my local client. This happens in a few settings:

1. My workers have access to a data store that I don't, so I need to call something like dd.read_parquet remotely (cc @martindurant @jcrist )
2. My workers are far away from my client, so client-heavy operations like joblib or Hyperband incur a serious bottleneck from client-scheduler communication (cc @stsievert )
3. My workers have hardware or libraries like GPUs/RAPIDS that I don't have locally (cc @quasiben @kkraus14)

Today I can do this by writing a function and submitting that function as a task

def f():
    import dask_cudf
    df = dask_cudf.read_parquet("s3://...")
    return df.sum().compute()

result = client.submit(f).result()

It might make sense to provide syntax around this to make it more magical (or it might not). We might do something like the following:

with dask.distributed.remote as result:
    import dask_cudf
    df = dask_cudf.read_parquet("s3://...")
    result = df.sum().compute()

I know that @eriknw has done magic like this in the past. We could enlist this help. However, we may not want to do this due to the magical and novel behavior.