[Autoscaler][V2] Check IM instance_status before terminating nodes

[ryanaoleary]

Why are these changes needed?

Currently when the autoscaler attempts to scale down Ray nodes due to max number of worker nodes per type reached, it's possible for the autoscaler to error and fail to reconcile the nodes with Invalid status transition from ALLOCATED to RAY_STOP_REQUESTED. This occurs when cloud IM instances (for KubeRay this would be Pods) have not yet been able to start Ray, but are still attempted to be scaled down due to max number of nodes per type constraints. This PR adds a check for im_instance_status when terminating nodes, transitioning directly from ALLOCATED to TERMINATING in the case where Ray has not yet started on the node but the autoscaler requires it to be scaled down. The linked issue contains a reproduction script and more context for why this change is necessary.

Related issue number

Closes #50868

Checks

• I've signed off every commit(by using the -s flag, i.e., git commit -s) in this PR.
• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
  • I've added any new APIs to the API Reference. For example, if I added a
    method in Tune, I've added it in doc/source/tune/api/ under the
    corresponding .rst file.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[ryanaoleary]

Closes #50868

[ryanaoleary]

cc: @kevin85421 @rueian

[kevin85421]

would you mind providing more context in the PR description and which case this PR fixes (a simple repro program is helpful)? I actually don't understand what does this PR wants to solve.

[ryanaoleary]

would you mind providing more context in the PR description and which case this PR fixes (a simple repro program is helpful)? I actually don't understand what does this PR wants to solve.

I have a reproduction script for the issue in #50868. I'll update the PR description with more context and a link to that issue.

[kevin85421]

Sorry for the multiple review iterations on this PR that exceeded our expectations. I believe the main issue is that the existing API (e.g., termination request/scheduling node) does not clearly indicate which fields are optional and when they should be provided. This is something we need to improve the existing codebase.

[kevin85421]

Would you mind adding some tests? I know you have a PR for end-to-end tests in the KubeRay repo, but would you mind adding a non-KubeRay test as well?

[ryanaoleary]

Would you mind adding some tests? I know you have a PR for end-to-end tests in the KubeRay repo, but would you mind adding a non-KubeRay test as well?

Sounds good, I added a unit test in 7736214 that validates that allocated instances can be deleted by the scheduler when max_worker_nodes limits are changed for a NodeTypeConfig.

[kevin85421]

cc @jjyao @edoakes would you mind merging or reviewing this PR? Thanks!