Remarks

Author: mimowo

crd/kubeflow.org_mpijobs.yaml

@@ -7886,15 +7886,15 @@ spec:
                 type: string
               suspend:
                 default: false
-                description: "suspend specifies whether the Job controller should
-                  create Pods or not. If a Job is created with suspend set to true,
-                  no Pods are created by the Job controller. If a Job is suspended
-                  after creation (i.e. the flag goes from false to true), the Job
-                  controller will delete all active Pods associated with this Job.
-                  Users must design their workload to gracefully handle this. Suspending
-                  a Job will reset the StartTime field of the Job, effectively resetting
-                  the ActiveDeadlineSeconds timer too. Defaults to false. \n Defaults
-                  to false."
+                description: "suspend specifies whether the MPIJob controller should
+                  create Pods or not. If a MPIJob is created with suspend set to true,
+                  no Pods are created by the MPIJob controller. If a MPIJob is suspended
+                  after creation (i.e. the flag goes from false to true), the MPIJob
+                  controller will delete all active Pods associated with this MPIJob.
+                  Also, it will suspend the Launcher Job. Users must design their
+                  workload to gracefully handle this. Suspending a Job will reset
+                  the StartTime field of the MPIJob, effectively resetting the activeDeadlineSeconds
+                  timer too. Defaults to false. \n Defaults to false."
                 type: boolean
             required:
             - mpiReplicaSpecs

pkg/apis/kubeflow/v2beta1/swagger.json

@@ -272,7 +272,7 @@
           "type": "string"
         },
         "suspend": {
-          "description": "suspend specifies whether the Job controller should create Pods or not. If a Job is created with suspend set to true, no Pods are created by the Job controller. If a Job is suspended after creation (i.e. the flag goes from false to true), the Job controller will delete all active Pods associated with this Job. Users must design their workload to gracefully handle this. Suspending a Job will reset the StartTime field of the Job, effectively resetting the ActiveDeadlineSeconds timer too. Defaults to false.\n\nDefaults to false.",
+          "description": "suspend specifies whether the MPIJob controller should create Pods or not. If a MPIJob is created with suspend set to true, no Pods are created by the MPIJob controller. If a MPIJob is suspended after creation (i.e. the flag goes from false to true), the MPIJob controller will delete all active Pods associated with this MPIJob. Also, it will suspend the Launcher Job. Users must design their workload to gracefully handle this. Suspending a Job will reset the StartTime field of the MPIJob, effectively resetting the activeDeadlineSeconds timer too. Defaults to false.\n\nDefaults to false.",
           "type": "boolean"
         }
       }

pkg/apis/kubeflow/v2beta1/types.go

@@ -63,13 +63,14 @@ type MPIJobSpec struct {
 	// +kubebuilder:default:=OpenMPI
 	MPIImplementation MPIImplementation `json:"mpiImplementation,omitempty"`
 
-	// suspend specifies whether the Job controller should create Pods or not. If
-	// a Job is created with suspend set to true, no Pods are created by the Job
-	// controller. If a Job is suspended after creation (i.e. the flag goes from
-	// false to true), the Job controller will delete all active Pods associated
-	// with this Job. Users must design their workload to gracefully handle this.
-	// Suspending a Job will reset the StartTime field of the Job, effectively
-	// resetting the ActiveDeadlineSeconds timer too. Defaults to false.
+	// suspend specifies whether the MPIJob controller should create Pods or not.
+	// If a MPIJob is created with suspend set to true, no Pods are created by
+	// the MPIJob controller. If a MPIJob is suspended after creation (i.e. the
+	// flag goes from false to true), the MPIJob controller will delete all
+	// active Pods associated with this MPIJob. Also, it will suspend the
+	// Launcher Job. Users must design their workload to gracefully handle this.
+	// Suspending a Job will reset the StartTime field of the MPIJob, effectively
+	// resetting the activeDeadlineSeconds timer too. Defaults to false.
 	//
 	// Defaults to false.
 	// +kubebuilder:default:=false

pkg/controller/mpi_job_controller.go

@@ -505,7 +505,7 @@ func (c *MPIJobController) syncHandler(key string) error {
 	// cleanup and stop retrying the MPIJob.
 	if isFinished(mpiJob.Status) && mpiJob.Status.CompletionTime != nil {
 		if isCleanUpPods(mpiJob.Spec.RunPolicy.CleanPodPolicy) {
-			return cleanUpPods(mpiJob, c)
+			return cleanUpWorkerPods(mpiJob, c)
 		}
 		return nil
 	}
@@ -578,17 +578,9 @@ func (c *MPIJobController) syncHandler(key string) error {
 	}
 
 	if launcher != nil {
-		launcherSuspendUpdate := false
-		if isMPIJobSuspended(mpiJob) && !isJobSuspended(launcher) {
-			// suspend the launcher first if the MPI job is suspended
-			launcherSuspendUpdate = true
-			launcher.Spec.Suspend = pointer.Bool(true)
-		} else if !isMPIJobSuspended(mpiJob) && isJobSuspended(launcher) {
-			launcherSuspendUpdate = true
-			// unsuspend the launcher first if the MPI job is unsuspended
-			launcher.Spec.Suspend = pointer.Bool(false)
-		}
-		if launcherSuspendUpdate {
+		if isMPIJobSuspended(mpiJob) != isJobSuspended(launcher) {
+			// align the suspension state of launcher with the MPIJob
+			launcher.Spec.Suspend = pointer.Bool(isMPIJobSuspended(mpiJob))
 			if _, err := c.kubeClient.BatchV1().Jobs(namespace).Update(context.TODO(), launcher, metav1.UpdateOptions{}); err != nil {
 				return err
 			}
@@ -597,14 +589,14 @@ func (c *MPIJobController) syncHandler(key string) error {
 
 	// cleanup the running worker pods if the MPI job is suspended
 	if isMPIJobSuspended(mpiJob) {
-		if err := cleanUpPods(mpiJob, c); err != nil {
+		if err := cleanUpWorkerPods(mpiJob, c); err != nil {
 			return err
 		}
 	}
 	return nil
 }
 
-func cleanUpPods(mpiJob *kubeflow.MPIJob, c *MPIJobController) error {
+func cleanUpWorkerPods(mpiJob *kubeflow.MPIJob, c *MPIJobController) error {
 	// set worker StatefulSet Replicas to 0.
 	if err := c.deleteWorkerPods(mpiJob); err != nil {
 		return err