RayJobs with clusterSelectors should not be managed/validated by kueue.

They will be submitted against RayClusters which are already managed by kueue so kueueing them would double up on the quota consumed.

Author: laurafitzgerald

pkg/controller/jobs/rayjob/rayjob_controller.go

@@ -78,6 +78,7 @@ type RayJob rayv1.RayJob
 
 var _ jobframework.GenericJob = (*RayJob)(nil)
 var _ jobframework.JobWithManagedBy = (*RayJob)(nil)
+var _ jobframework.JobWithSkip = (*RayJob)(nil)
 
 func (j *RayJob) Object() client.Object {
 	return (*rayv1.RayJob)(j)
@@ -100,6 +101,12 @@ func (j *RayJob) Suspend() {
 	j.Spec.Suspend = true
 }
 
+func (j *RayJob) Skip(ctx context.Context) bool {
+	// Skip reconciliation for RayJobs that use clusterSelector to reference existing clusters.
+	// These jobs are not managed by Kueue.
+	return len(j.Spec.ClusterSelector) > 0
+}
+
 func (j *RayJob) GVK() schema.GroupVersionKind {
 	return gvk
 }

pkg/controller/jobs/rayjob/rayjob_webhook.go

@@ -107,6 +107,12 @@ func (w *RayJobWebhook) validateCreate(ctx context.Context, job *rayv1.RayJob) (
 	var allErrors field.ErrorList
 	kueueJob := (*RayJob)(job)
 
+	// RayJobs with clusterSelector use an existing cluster and should not be managed by Kueue.
+	// Skip all Kueue validation for these jobs.
+	if len(job.Spec.ClusterSelector) > 0 {
+		return allErrors, nil
+	}
+
 	if w.manageJobsWithoutQueueName || jobframework.QueueName(kueueJob) != "" {
 		spec := &job.Spec
 		specPath := field.NewPath("spec")
@@ -116,11 +122,6 @@ func (w *RayJobWebhook) validateCreate(ctx context.Context, job *rayv1.RayJob) (
 			allErrors = append(allErrors, field.Invalid(specPath.Child("shutdownAfterJobFinishes"), spec.ShutdownAfterJobFinishes, "a kueue managed job should delete the cluster after finishing"))
 		}
 
-		// Should not want existing cluster. Kueue (workload) should be able to control the admission of the actual work, not only the trigger.
-		if len(spec.ClusterSelector) > 0 {
-			allErrors = append(allErrors, field.Invalid(specPath.Child("clusterSelector"), spec.ClusterSelector, "a kueue managed job should not use an existing cluster"))
-		}
-
 		clusterSpec := spec.RayClusterSpec
 		clusterSpecPath := specPath.Child("rayClusterSpec")