Merge pull request kubernetes#130189 from ania-borowiec/129967_get_rid_of_caching_cluster_events_in_binding

Call queue.Done() before PreBind phase, removing the pod in binding from inFlightPods to save memory

Author: k8s-ci-robot

pkg/scheduler/backend/queue/scheduling_queue.go

@@ -134,6 +134,7 @@ type SchedulingQueue interface {
 	PodsInActiveQ() []*v1.Pod
 	// PodsInBackoffQ returns all the Pods in the backoffQ.
 	PodsInBackoffQ() []*v1.Pod
+	UnschedulablePods() []*v1.Pod
 }
 
 // NewSchedulingQueue initializes a priority queue as a new scheduling queue.
@@ -1205,6 +1206,15 @@ func (p *PriorityQueue) PodsInBackoffQ() []*v1.Pod {
 	return p.backoffQ.list()
 }
 
+// UnschedulablePods returns all the pods in unschedulable state.
+func (p *PriorityQueue) UnschedulablePods() []*v1.Pod {
+	var result []*v1.Pod
+	for _, pInfo := range p.unschedulablePods.podInfoMap {
+		result = append(result, pInfo.Pod)
+	}
+	return result
+}
+
 var pendingPodsSummary = "activeQ:%v; backoffQ:%v; unschedulablePods:%v"
 
 // GetPod searches for a pod in the activeQ, backoffQ, and unschedulablePods.
@@ -1241,9 +1251,9 @@ func (p *PriorityQueue) GetPod(name, namespace string) (pInfo *framework.QueuedP
 func (p *PriorityQueue) PendingPods() ([]*v1.Pod, string) {
 	p.lock.RLock()
 	defer p.lock.RUnlock()
-	result := p.activeQ.list()
+	result := p.PodsInActiveQ()
 	activeQLen := len(result)
-	backoffQPods := p.backoffQ.list()
+	backoffQPods := p.PodsInBackoffQ()
 	backoffQLen := len(backoffQPods)
 	result = append(result, backoffQPods...)
 	for _, pInfo := range p.unschedulablePods.podInfoMap {

pkg/scheduler/schedule_one.go

@@ -292,30 +292,19 @@ func (sched *Scheduler) bindingCycle(
 		return status
 	}
 
-	// Run "prebind" plugins.
-	if status := fwk.RunPreBindPlugins(ctx, state, assumedPod, scheduleResult.SuggestedHost); !status.IsSuccess() {
-		if status.IsRejected() {
-			fitErr := &framework.FitError{
-				NumAllNodes: 1,
-				Pod:         assumedPodInfo.Pod,
-				Diagnosis: framework.Diagnosis{
-					NodeToStatus:         framework.NewDefaultNodeToStatus(),
-					UnschedulablePlugins: sets.New(status.Plugin()),
-				},
-			}
-			fitErr.Diagnosis.NodeToStatus.Set(scheduleResult.SuggestedHost, status)
-			return framework.NewStatus(status.Code()).WithError(fitErr)
-		}
-		return status
-	}
-
 	// Any failures after this point cannot lead to the Pod being considered unschedulable.
-	// We define the Pod as "unschedulable" only when Pods are rejected at specific extension points, and PreBind is the last one in the scheduling/binding cycle.
+	// We define the Pod as "unschedulable" only when Pods are rejected at specific extension points, and Permit is the last one in the scheduling/binding cycle.
+	// If a Pod fails on PreBind or Bind, it should be moved to BackoffQ for retry.
 	//
 	// We can call Done() here because
-	// we can free the cluster events stored in the scheduling queue sonner, which is worth for busy clusters memory consumption wise.
+	// we can free the cluster events stored in the scheduling queue sooner, which is worth for busy clusters memory consumption wise.
 	sched.SchedulingQueue.Done(assumedPod.UID)
 
+	// Run "prebind" plugins.
+	if status := fwk.RunPreBindPlugins(ctx, state, assumedPod, scheduleResult.SuggestedHost); !status.IsSuccess() {
+		return status
+	}
+
 	// Run "bind" plugins.
 	if status := sched.bind(ctx, fwk, assumedPod, scheduleResult.SuggestedHost, state); !status.IsSuccess() {
 		return status