[Discussion] Enhance/optimise current implementation of processing newly added Job Frameworks.

[ChristianZaccaria]

This is a follow-up issue to PR #2574: Start JobFramework controller and webhook on new CRDs availability.

What would you like to be added: Enhance/optimise current design and implementation of https://github.com/kubernetes-sigs/kueue/pull/2574

Why is this needed:

• Current implementation works as expected, however, it could be enhanced with:
  • Enhance current design and implementation of PR: Start JobFramework controller and webhook on new CRDs availability. As part of this issue, we would want to further discuss on other more optimal alternatives to the functionality in place. Please find in this thread the main points of discussion.
  • Only start waitForAPI() once the CRD is registered (wait either by controller or watcher on CRD type).
  • Watching CRDs directly with apiextensions. Once we discover a new CRD, then we start waiting for the restmapper to be filled.
  • One go-routine for all frameworks as opposed to one per framework.
  • Include integration tests aligned with enhanced implementation.

Completion requirements:

This enhancement requires the following artifacts:

• [ ] Design doc
• [ ] API change
• [ ] Docs update

The artifacts should be linked in subsequent comments.

[mimowo]

Just a little bit of motivation, this will allow us:

• reduce the amount of logging for not installed CRDs
• shorten the time for the API to be ready - currently due to the exponential backoff the detection of API is likely to take 2min (max period). if we start active polling after the CRD is installed the REST mapper will be detected quickly.

I think an integration test can be added even to the current mechanism, so I think it makes sense to keep https://github.com/kubernetes-sigs/kueue/issues/3015 open