lmeyerov

@aminsshahab I'm checking w/ our AWS collaborators for guidance -- assuming not complicated, maybe end-of-week / next? I _think_ the private subnet thing is: 1. If creds specified, remove INSTANCE...

Great -- and how would web browsers access your graph-app-kit instance... say via an AWS-friendly VPN that allows access to the private IP, or maybe a public ELB that has...

- @briantrice @trxcllnt @quinnhj The rest of us are, historically, terrible at naming, and this one in particular has been an open issue ;-)

This sounds like `node-opencl` is failing. Can you try doing a vanilla install of node-opencl, e.g., `npm install node-opencl ; npm test`?

Oops, this should be in `dask-cuda`, sorry!

I'm not sure that's true of async clusters. Either way, reran with the edit, same thing: ```python import asyncio, cudf, dask_cudf from dask_cuda import LocalCUDACluster from dask.distributed import Client async...

@kkraus14 It's a known bug in `LocalCluster` - I had dug up this one a few days ago, where @mrocklin reported it didn't matter as much for dask's impl of...

Good news: when using a remote cluster, we can having both sync + async clients hit the same cluster. So this is really scoped down to `localcudacluster` not supporting mixed...

Clever! Interestingly, while it still times out, I wonder if it's partly due to running in docker. I added a printf to the address (`print('making sync client for cluster', cluster_async.scheduler.address)`),...

I don't think so -- we work around this now via a remote dask service, which does support simultaneous async+sync clients