[feature/idea] Integrate kubernetes's models into kfp-kubernetes

cjidboon94 commented 11 months ago

Feature Area

/area backend /area sdk

What feature would you like to see?

kfp-kubernetes is a library that aims to extend kfp v2 with kubernetes functionality. Currently the kfp-kubernetes package is still missing some key features imo:

affinities/tolerations https://github.com/kubeflow/pipelines/issues/9682
extended resource requests/limits
init_containers + sidecars https://github.com/kubeflow/pipelines/issues/9679 and https://github.com/kubeflow/pipelines/issues/9700
pod annotation/labels https://github.com/kubeflow/pipelines/issues/9714
Something like ResourceOp from v1: https://github.com/kubeflow/pipelines/issues/9703
Set pull policies: https://github.com/kubeflow/pipelines/issues/9980
Add secrets used for pulling images

Some of the implemented features of kfp-kubernetes are currently limited. For example, only secrets can be loaded as env vars/volumes while this could also be from a Config Map, a resource's field or an object's field: (via V1EnvVarSource/V1Volume respectively ) . Integration with the kubernetes package (as was the case with KFP v1) could facillitate this. The protos that kfp-kubernetes requires could be generated based on the OpenAPI spec. The methods in go/python for manipulating the objects would still need to be implemented. I believe integrating the models from the kubernetes python client package is a natural way for the kfp-kubernetes to grow in functionality and ability.

What is the use case or pain point?

It would enable different kind of Env var sources, tolerations and affinities would be easier to set/reused from the KFP v1.
Init_containers and sidecars can be again set (via V1Container).

Is there a workaround currently?

N/A

Love this idea? Give it a 👍.

schrodervictor commented 10 months ago

Is this already being worked or is it up for grabs? What is contribution guideline and workflow in case I volunteer to implement the tolerations model? We really need it for our project.

pythonking6 commented 8 months ago

it is shocking that v2 pipelines would be released without this feature!

roytman commented 6 months ago

The above list of missing functionality can be extended by:

only secrets can be mounted as a volume, while this could also be from a Config Map
we cannot set an image pull policy
a secret to pull images

cjidboon94 commented 5 months ago

The above list of missing functionality can be extended by:

only secrets can be mounted as a volume, while this could also be from a Config Map

we cannot set an image pull policy

a secret to pull images

Added them to the list!

github-actions[bot] commented 3 months ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

cjidboon94 commented 3 months ago

Commenting as this issue is still relevant

rimolive commented 2 months ago

/lifecycle frozen

I'm trying to get the last issue PR and merge asap. The original contributor cannot keep contributing with it.

cjidboon94 commented 2 months ago

@rimolive Hey, I'm working on the node affinities backend + sdk implementation https://github.com/kubeflow/pipelines/pull/10671 this change is necessary and hoping to have the rest of the node affinity work done at end of this weekend and have a PR out

Pod (anti)affinity is I think less urgent

rimolive commented 2 months ago

Thanks for contributing with this, @cjidboon94 !

Did you see #9913? It is different from your implementation. Just wanted to check if we're not missing anything.

kubeflow / pipelines