[Backport 7.79.x] Track evictions in VPA

pkg/clusteragent/autoscaling/workload/controller_vertical.go

@@ -127,6 +127,13 @@ func (u *verticalController) sync(ctx context.Context, podAutoscaler *datadoghq.
 		podsPerDirectOwner[pod.Owners[0].ID] = podsPerDirectOwner[pod.Owners[0].ID] + 1
 	}
 
+	// Get the live pod UIDs, prune pod operations for pods that are no longer in the live set.
+	livePodUIDs := make(map[string]struct{}, len(pods))
+	for _, pod := range pods {
+		livePodUIDs[pod.EntityID.ID] = struct{}{}
+	}
+	autoscalerInternal.PrunePodOperations(recommendationID, livePodUIDs)
+
 	// Classify each non-terminating pod by resize status so we can set scaled replicas
 	// (completed pods count) accurately. Pass this slice to syncInternal to avoid
 	// a duplicate call to getPodResizeStatus.
@@ -135,6 +142,10 @@ func (u *verticalController) sync(ctx context.Context, podAutoscaler *datadoghq.
 		if pod.DeletionTimestamp != nil {
 			continue
 		}
+		if autoscalerInternal.HasPendingOperation(pod.EntityID.ID, recommendationID) {
+			podsByResizeStatus[PodResizeStatusEvicting] = append(podsByResizeStatus[PodResizeStatusEvicting], classifiedPod{pod: pod})
+			continue
+		}
 		status, ltt := getPodResizeStatus(pod, recommendationID)
 		podsByResizeStatus[status] = append(podsByResizeStatus[status], classifiedPod{pod: pod, lastTransitionTime: ltt})
 	}
@@ -276,6 +287,7 @@ func (u *verticalController) syncInternal(
 		if result == evictor.Evicted {
 			evictedThisSync++
 			autoscalerInternal.InPlaceEvictionSuccessInc()
+			autoscalerInternal.TrackPodOperation(cp.pod.EntityID.ID, recommendationID)
 		}
 		if result == evictor.PDBLockedOrThrottle || result == evictor.Skipped {
 			pdbBlocked = true
@@ -303,8 +315,14 @@ func (u *verticalController) syncInternal(
 			eventType = corev1.EventTypeWarning
 			reason = model.FailedToEvictEventReason
 		}
+		inFlight := len(podsByResizeStatus[PodResizeStatusEvicting])
+		remaining := int(int32(len(toEvict)) - evictedThisSync - failedEvictions)
+		suffix := fmt.Sprintf("%d remaining", remaining)
+		if inFlight > 0 {
+			suffix += fmt.Sprintf(", %d in-flight", inFlight)
+		}
 		u.eventRecorder.Eventf(podAutoscaler, eventType, reason,
-			"In-place resize eviction: %s (%d pods pending)", strings.Join(parts, ", "), len(toEvict))
+			"In-place resize eviction: %s (%s)", strings.Join(parts, ", "), suffix)
 	}
 
 	// Terminating pods are excluded from podsByResizeStatus, so summing all bucket lengths
@@ -314,6 +332,7 @@ func (u *verticalController) syncInternal(
 		totalActive += len(bucket)
 	}
 	if len(podsByResizeStatus[PodResizeStatusCompleted]) == totalActive {
+		autoscalerInternal.ClearPodOperations()
 		if lastAction := autoscalerInternal.VerticalLastAction(); lastAction != nil &&
 			lastAction.Type == datadoghqcommon.DatadogPodAutoscalerResizeTriggeredVerticalActionType {
 			u.eventRecorder.Eventf(podAutoscaler, corev1.EventTypeNormal, model.ResizeSuccessfulEventReason,

pkg/clusteragent/autoscaling/workload/controller_vertical_helpers.go

@@ -92,6 +92,8 @@ const (
 	PodResizeStatusError
 	PodResizeStatusInfeasible
 	PodResizeStatusDeferred
+	// PodResizeStatusEvicting marks pods with an accepted eviction pending termination.
+	PodResizeStatusEvicting
 )
 
 // classifiedPod pairs a pod with the LastTransitionTime of the condition that

pkg/clusteragent/autoscaling/workload/model/pod_autoscaler.go

@@ -202,6 +202,10 @@ type PodAutoscalerInternal struct {
 	// customRecommenderConfiguration holds the configuration for custom recommenders,
 	// Parsed from annotations on the autoscaler
 	customRecommenderConfiguration *RecommenderConfiguration
+
+	// submittedPodOps maps pod UID to the recommendationID of the last accepted patch or eviction,
+	// used to suppress redundant API calls during the informer-cache lag window.
+	submittedPodOps map[string]string
 }
 
 // NewPodAutoscalerInternal creates a new PodAutoscalerInternal from a Kubernetes CR
@@ -901,6 +905,40 @@ func (p *PodAutoscalerInternal) InPlaceResizeCompletedInc() {
 	p.inPlaceResizeCompletedCount++
 }
 
+// TrackPodOperation records that a patch or eviction for podUID was accepted by the API server.
+func (p *PodAutoscalerInternal) TrackPodOperation(podUID, recommendationID string) {
+	if p.submittedPodOps == nil {
+		p.submittedPodOps = make(map[string]string)
+	}
+	p.submittedPodOps[podUID] = recommendationID
+}
+
+// HasPendingOperation reports whether podUID has a tracked operation for the current recommendationID.
+func (p *PodAutoscalerInternal) HasPendingOperation(podUID, recommendationID string) bool {
+	if p.submittedPodOps == nil {
+		return false
+	}
+	return p.submittedPodOps[podUID] == recommendationID
+}
+
+// ClearPodOperations drops all tracked pod operations; called on resize completion.
+func (p *PodAutoscalerInternal) ClearPodOperations() {
+	p.submittedPodOps = nil
+}
+
+// PrunePodOperations removes entries whose stored recommendationID no longer matches
+// the current one, and entries whose pod UID have been removed from the livePodUIDs map.
+func (p *PodAutoscalerInternal) PrunePodOperations(recommendationID string, livePodUIDs map[string]struct{}) {
+	for uid, rid := range p.submittedPodOps {
+		if rid != recommendationID {
+			delete(p.submittedPodOps, uid)
+		}
+		if _, ok := livePodUIDs[uid]; !ok {
+			delete(p.submittedPodOps, uid)
+		}
+	}
+}
+
 // CurrentReplicas returns the current number of PODs for the targetRef
 func (p *PodAutoscalerInternal) CurrentReplicas() *int32 {
 	return p.currentReplicas