Matthew Rocklin

I've been thinking more about locking. Assuming that we're only locking in-process (presumably `os.environ` is the protected resource) then if we can arrange things to put a lock around process...

My preference would be a `mode={"at least", "at most", "exactly"}` keyword.

I think that we need to think a little bit about how this would be used. When submitting a script how does that script get access to the scheduler? Naively...

I wouldn't consider them to be public to users necessarily, but certainly they're public to other dask libraries and other friends-of-dask like Xarray. From my perspective those libraries are within...

Rebasing shouldn't be necessary. a Merge commit would be welcome as well (and may be easier) On Tue, Jun 29, 2021 at 4:06 AM Florian Jetter ***@***.***> wrote: > @charlesbluca...

Thanks for the issue @cjnolet . I'm sorry to hear that you've had a frustrating time. If you can provide a minimal reproducible example I'd be happy to run through...

Thanks @cjnolet . In the future if you're able to provide a copy-pasteable example that anyone can run that would help. For example, if you don't need `LocalCUDACluster` then please...

Ah, that's perfect. Thank you @cjnolet . I'm able to reproduce. I'll try to take a look at this this weekend.

Here is a simpler failure. I've removed the tree reduction. ```python from dask.distributed import Client, LocalCluster async def test(n_parts, client): a = await client.scatter(range(n_parts), hash=True) future = client.submit(sum, a) await...

The error that you have uncovered here happens when 1. We create some futures by scattering data (not normal computations) 2. We delete those futures 3. We recreate those exact...