search

usatlas / analysisbase-dask-uc

UC AF specific image building
https://hub.docker.com/r/sslhep/analysis-dask-uc/tags
2 stars 0 forks source link

Update deployment to use Dask Kubernetes Operator #9

Open matthewfeickert opened 6 months ago

matthewfeickert commented 6 months ago

So it feels like the time to finally retire the classic KubeCluster implementation along with the HelmCluster.

I suspect that the update to dask-kubernetes is the reason. Maybe some JupyterLab integration would need to be updated accordingly, or perhaps some other backend thing, but for now I think it would be good to revert the change.

Originally posted by @alexander-held in https://github.com/usatlas/analysisbase-dask/issues/49#issue-2278937281

The Dask Kubernetes Operator has been the standard for deploying Dask on Kubernetes for over a year now. So it feels like the time to finally retire the classic KubeCluster implementation along with the HelmCluster.

https://github.com/dask/dask-kubernetes/releases/tag/2024.5.0

This would be useful to switch to as otherwise Dask will be stuck on the 2024.4.x releases.

@ivukotic is this possible?

matthewfeickert commented 6 months ago

You can see our dask configuration in the derived AB image or in Jupyter directly. Our clusters are deployed using Dask Operators. (We are testing a Dask Gateway with authentication and everything but don’t want to change anything till the DESY meetings. — @ivukotic

(From IRIS-HEP Slack)

@ivukotic Sorry, I'm not experience here, so can you point out what section of https://github.com/usatlas/analysisbase-dask-uc/blob/ccd0a8000c6ee3d82b53f3e022d81c4f4bc8d34f/dask_config.yaml is this specified in (or what section of the Dask Kubernetes Operator docs I should be looking at)? Also, if the deployment is already done with Dask Operators, why did @alexander-held encounter the issue from https://github.com/usatlas/analysisbase-dask/pull/48?

(We are testing a Dask Gateway with authentication and everything but don’t want to change anything till the DESY meetings).

:+1: Nice, thank you!

matthewfeickert commented 2 months ago

@LincolnBryant @fengpinghu can you also take a look at this?

fengpinghu commented 2 months ago

It looks it's indeed using the classic KubeCluster because in the configuration file KubeCluster users worker template and scheduler template to create pods directly. We will look into either dask-operator or dask-gateway integration.

matthewfeickert commented 1 month ago

We will look into either dask-operator or dask-gateway integration.

Great! Thanks very much, @fengpinghu!