[feat][RayCluster] shorten HeadlessServiceSuffix to have more space for CR names

[rueian]

Why are these changes needed?

Currently, KubeRay adds different suffixes to make different services for a RayCluster. It will also trim these service names if they are too long to fit the 63-character limitation by k8s. However, the silent cut is not user-friendly as users can't easily know the resulting service names beforehand.

As a part of #3076: we are going to add a length validation to RayCluster names in v1.4 to make sure the silent cut does not take place while keeping sufficient room for the names specified by users, this PR shortens the -headless-worker-svc suffix to just -headless, saving 11 characters for users to use (63-len("-serve-svc") = 53 characters).

Related issue number

#3076

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(

[kevin85421]

can you split this into two PRs, one for headless service and one for RayCluster?

[rueian]

can you split this into two PRs, one for headless service and one for RayCluster?

Sure, another PR is here: #3102

[rueian]

Hi @andrewsykim,

We'd like to add a length validation to RayCluster names in the next release to provide non-trimmed service names for a better user experience. But the current headless service suffix for the TPU webhook, -headless-worker-svc, is already too long, we'd like to shorten it to just -headless to leave more space for RayCluster names.

Please let me know if you have any concerns about this. Thanks!

[kevin85421]

#3101 (comment)

cc @ryanaoleary

[ryanaoleary]

#3101 (comment)

cc @ryanaoleary

I'll need to change the logic of the webhook here before this change so that TPU_WORKER_HOSTNAMES are still generated correctly. I'll change it to look up the headless service name based on the RayCluster label.

[kevin85421]

GoogleCloudPlatform/ai-on-gke#1003 is merged.