set a custom port for the Dask cluster about kluster HOT 10 OPEN

@epifanio Oh, I see now. Get_client will automatically resolve that port for you, if it detects the local scheduler. But I can certainly add that option to the LocalCluster initialization for you, and expose it in the dialog. That should be a simple change. I can do a new pull request tonight for that change.

from kluster.

@epifanio ok cool, I'll be interested to hear how that goes. I've changed jobs recently, which is why you haven't seen much activity here lately. But I'm still invested, so keep in touch with the results!

from kluster.

Hi @epifanio,

Thanks for your interest and all the commits! I do this already in the dask client

which runs dask_find_or_start_client with the IP:PORT string.

The docs aren't clear (I should probably update), but you can pass anything that dask get_client accepts.

You can pass around a client in the API as follows:

from HSTB.kluster.dask_helpers import dask_find_or_start_client
from HSTB.kluster.fqpr_intelligence import intel_process

cl = dask_find_or_start_client(address='192.168.0.1:6454', number_of_workers=4)
intel_process(r'multibeam_file_path', outfold='destination_path')

I haven't used the client this way yet, so if you run into issues, let me know.

from kluster.

ok, didn't see that settings - this means I can expose a port on the docker file and then use that one in the UI - that should work, thanks!

from kluster.

@ericgyounkin those highlighted settings are made to attach the processing to an existing dusk cluster and not to spin a local one - is that correct?
I am trying to spin a new cluster on the container, but using a user-provided port instead of a random one - this way I can expose that port and have access to the dask cluster (which runs in docker) from the browser on the host machine.

like, having an entry: PORT: ____ in the Local Cluster settings

from kluster.

@epifanio

Highlighted settings are for a remote cluster, yes. You use the LocalCluster to setup a Dask distributed instance on your computer.

It needs to align with the get_client API, so I believe in that instance you would use something like '127.0.0.1:PORT' as the address. Basically, you need to provide some kind of address I believe for the port to be accessable.

from kluster.

@ericgyounkin , thanks for your reply

What I was looking for is a way to access the "local cluster" which starts on the containerized app - at the moment I can spin the local cluster and I can see in the kluster logs the following:

2022-09-09 12:05:48,245 - INFO - Starting local cluster client...
2022-09-09 12:05:49,726 - INFO - <Client: 'tcp://127.0.0.1:43579' processes=4 threads=8, memory=15.58 GiB>

which means the dask cluster is up and running, but is not accessible from outside the container because the PORT used is chosen randomnly - we need the PORT to be configurable also for the local cluster, so we can expose it in the docker configuration :)

I am looking on how to pass the PORT as argument to the dask api - when the cluster is initialized so that get_client() will return the localhost:$PORT

from kluster.

@epifanio Unfortunately there seems to be a bug in the version of Dask that we currently use in Kluster. If we set the scheduler port in LocalCluster

cl = LocalCluster(scheduler_port='53647')

We get an error based on how it parses the default address. I checked and the bug is resolved in the latest dask.distributed (committed June 2022)

https://github.com/dask/distributed/blob/main/distributed/comm/addressing.py#L201

But I don't think we'll be updating the library used in Kluster in the short term. So I'm not sure how to support this.

from kluster.

I see, that is unfortunate but I understand -
As a back-up solution, I may try to add an additional container in the docker-compose just for dask, spin it at runtime and then try to attach the GUI to it.

from kluster.

Hi, @ericgyounkin I am back on kluster :) I tested a docker-compose environment which seems to work as expected but I need t run some testing ' will try to post it here when ready.

from kluster.