add wildcard support for evpa container policy, and update the status…

… even if it is off mode, because we need to record the metrics

pkg/autoscaling/estimator/percentile.go

@@ -92,7 +92,7 @@ func (e *PercentileResourceEstimator) GetResourceEstimation(evpa *autoscalingapi
 		predictErrs = append(predictErrs, err)
 	}
 
-	if len(tsList) > 1 && len(tsList[0].Samples) > 1 {
+	if len(tsList) > 0 && len(tsList[0].Samples) > 0 {
 		cpuValue := int64(tsList[0].Samples[0].Value * 1000)
 		recommendResource[corev1.ResourceCPU] = *resource.NewMilliQuantity(cpuValue, resource.DecimalSI)
 	} else {
@@ -104,7 +104,7 @@ func (e *PercentileResourceEstimator) GetResourceEstimation(evpa *autoscalingapi
 		predictErrs = append(predictErrs, err)
 	}
 
-	if len(tsList) > 1 && len(tsList[0].Samples) > 1 {
+	if len(tsList) > 0 && len(tsList[0].Samples) > 0 {
 		memValue := int64(tsList[0].Samples[0].Value)
 		recommendResource[corev1.ResourceMemory] = *resource.NewQuantity(memValue, resource.BinarySI)
 	} else {

pkg/controller/evpa/container_policy.go

@@ -28,23 +28,28 @@ func (c *EffectiveVPAController) ReconcileContainerPolicies(evpa *autoscalingapi
 	recommendation = evpa.Status.Recommendation
 
 	rankedEstimators := RankEstimators(resourceEstimators)
-	for _, containerPolicy := range evpa.Spec.ResourcePolicy.ContainerPolicies {
-		// container scaling is disabled
-		if (containerPolicy.ScaleUpPolicy.ScaleMode != nil && *containerPolicy.ScaleUpPolicy.ScaleMode == vpatypes.ContainerScalingModeOff) ||
-			(containerPolicy.ScaleDownPolicy.ScaleMode != nil && *containerPolicy.ScaleDownPolicy.ScaleMode == vpatypes.ContainerScalingModeOff) {
-			continue
+	needReconciledContainers := make(map[string]autoscalingapi.ContainerResourcePolicy)
+	for _, container := range podTemplate.Spec.Containers {
+		for _, containerPolicy := range evpa.Spec.ResourcePolicy.ContainerPolicies {
+			if containerPolicy.ContainerName == "*" || containerPolicy.ContainerName == container.Name {
+				out := containerPolicy.DeepCopy()
+				out.ContainerName = container.Name
+				needReconciledContainers[container.Name] = *out
+			}
 		}
-
+	}
+	for container, containerPolicy := range needReconciledContainers {
 		// get current resource by pod template
-		// todo: support "*"
-		resourceRequirement, found := utils.GetResourceByPodTemplate(podTemplate, containerPolicy.ContainerName)
+		resourceRequirement, found := utils.GetResourceByPodTemplate(podTemplate, container)
 		if !found {
 			klog.Warningf("ContainerName %s not found", containerPolicy.ContainerName)
 			continue
 		}
 
 		// loop estimator and get final estimated resource for container
 		recommendResourceContainer, currentStatus := GetEstimatedResourceForContainer(evpa, containerPolicy, resourceRequirement, rankedEstimators, currentEstimatorStatus)
+		// record the recommended resource each time to do estimating. so we can get more observability
+		recordResourceRecommendation(evpa, containerPolicy, recommendResourceContainer)
 		currentEstimatorStatus = currentStatus
 		if IsResourceListEmpty(recommendResourceContainer) {
 			klog.V(4).Infof("Container %s recommend resource is empty, skip scaling. ", containerPolicy.ContainerName)

pkg/controller/evpa/effective_vpa_controller.go

@@ -92,7 +92,7 @@ func (c *EffectiveVPAController) Reconcile(ctx context.Context, req ctrl.Request
 			c.UpdateStatus(ctx, evpa, newStatus)
 			return ctrl.Result{}, err
 		}
-		c.Recorder.Event(evpa, v1.EventTypeNormal, "RemoveFinalizers", err.Error())
+		c.Recorder.Event(evpa, v1.EventTypeNormal, "RemoveFinalizers", "")
 	} else if !utils.ContainsString(evpa.Finalizers, known.AutoscalingFinalizer) {
 		evpa.Finalizers = append(evpa.Finalizers, known.AutoscalingFinalizer)
 		err = c.Client.Update(ctx, evpa)
@@ -155,15 +155,38 @@ func (c *EffectiveVPAController) SetupWithManager(mgr ctrl.Manager) error {
 		Complete(c)
 }
 
-func recordMetric(evpa *autoscalingapi.EffectiveVerticalPodAutoscaler, status *autoscalingapi.EffectiveVerticalPodAutoscalerStatus, podTemplate *v1.PodTemplateSpec) {
-	labels := map[string]string{
-		"resourceName": fmt.Sprintf("%s/%s", evpa.Namespace, evpa.Spec.TargetRef.Name),
+func recordResourceRecommendation(evpa *autoscalingapi.EffectiveVerticalPodAutoscaler, containerPolicy autoscalingapi.ContainerResourcePolicy, resourceList v1.ResourceList) {
+	for resourceName, resource := range resourceList {
+		labels := map[string]string{
+			"apiversion": evpa.Spec.TargetRef.APIVersion,
+			"owner_kind": evpa.Spec.TargetRef.Kind,
+			"namespace":  evpa.Namespace,
+			"owner_name": evpa.Spec.TargetRef.Name,
+			"container":  containerPolicy.ContainerName,
+			"resource":   resourceName.String(),
+		}
+		switch resourceName {
+		case v1.ResourceCPU:
+			metrics.EVPAResourceRecommendation.With(labels).Set(float64(resource.MilliValue()) / 1000.)
+		case v1.ResourceMemory:
+			metrics.EVPAResourceRecommendation.With(labels).Set(float64(resource.Value()))
+		}
 	}
+}
+
+func recordMetric(evpa *autoscalingapi.EffectiveVerticalPodAutoscaler, status *autoscalingapi.EffectiveVerticalPodAutoscalerStatus, podTemplate *v1.PodTemplateSpec) {
 
 	if status.Recommendation == nil {
 		return
 	}
 	for _, container := range status.Recommendation.ContainerRecommendations {
+		labels := map[string]string{
+			"apiversion": evpa.Spec.TargetRef.APIVersion,
+			"owner_kind": evpa.Spec.TargetRef.Kind,
+			"namespace":  evpa.Namespace,
+			"owner_name": evpa.Spec.TargetRef.Name,
+			"container":  container.ContainerName,
+		}
 		resourceRequirement, found := utils.GetResourceByPodTemplate(podTemplate, container.ContainerName)
 		if !found {
 			klog.Warningf("ContainerName %s not found", container.ContainerName)
@@ -172,30 +195,32 @@ func recordMetric(evpa *autoscalingapi.EffectiveVerticalPodAutoscaler, status *a
 
 		recommendCpu := container.Target[v1.ResourceCPU]
 		currentCpu := resourceRequirement.Requests[v1.ResourceCPU]
+		labels["resource"] = v1.ResourceCPU.String()
 		if currentCpu.Cmp(recommendCpu) > 0 {
 			// scale down
 			currCopy := currentCpu.DeepCopy()
 			currCopy.Sub(recommendCpu)
-			metrics.EVPACpuScaleDownMilliCores.With(labels).Set(float64(currCopy.MilliValue()))
+			metrics.EVPACpuScaleDownMilliCores.With(labels).Set(float64(currCopy.MilliValue()) / 1000.)
 		} else if currentCpu.Cmp(recommendCpu) < 0 {
 			// scale up
 			recommendCopy := recommendCpu.DeepCopy()
 			recommendCopy.Sub(currentCpu)
-			metrics.EVPACpuScaleUpMilliCores.With(labels).Set(float64(recommendCopy.MilliValue()))
+			metrics.EVPACpuScaleUpMilliCores.With(labels).Set(float64(recommendCopy.MilliValue()) / 1000.)
 		}
 
 		recommendMem := container.Target[v1.ResourceMemory]
 		currentMem := resourceRequirement.Requests[v1.ResourceMemory]
+		labels["resource"] = v1.ResourceMemory.String()
 		if currentMem.Cmp(recommendMem) > 0 {
 			// scale down
 			currCopy := currentMem.DeepCopy()
 			currCopy.Sub(recommendMem)
-			metrics.EVPAMemoryScaleDownMB.With(labels).Set(float64(currCopy.Value() / 1024 / 1024))
+			metrics.EVPAMemoryScaleDownMB.With(labels).Set(float64(currCopy.Value()))
 		} else if currentMem.Cmp(recommendMem) < 0 {
 			// scale up
 			recommendCopy := recommendMem.DeepCopy()
 			recommendCopy.Sub(currentMem)
-			metrics.EVPAMemoryScaleUpMB.With(labels).Set(float64(recommendCopy.Value() / 1024 / 1024))
+			metrics.EVPAMemoryScaleUpMB.With(labels).Set(float64(recommendCopy.Value()))
 		}
 	}
 }

pkg/metrics/autoscaling.go

@@ -49,51 +49,52 @@ var (
 		prometheus.GaugeOpts{
 			Namespace: "crane",
 			Subsystem: "autoscaling",
-			Name:      "effective_vpa_cpu_scale_up_millicores",
+			Name:      "effective_vpa_cpu_scale_up",
 			Help:      "The cpu scale up for Effective VPA",
 		},
-		[]string{
-			"resourceName",
-		},
+		[]string{"apiversion", "owner_kind", "namespace", "owner_name", "container", "resource"},
 	)
 	EVPACpuScaleDownMilliCores = prometheus.NewGaugeVec(
 		prometheus.GaugeOpts{
 			Namespace: "crane",
 			Subsystem: "autoscaling",
-			Name:      "effective_vpa_cpu_scale_down_millicores",
+			Name:      "effective_vpa_cpu_scale_down",
 			Help:      "The cpu scale down for Effective VPA",
 		},
-		[]string{
-			"resourceName",
-		},
+		[]string{"apiversion", "owner_kind", "namespace", "owner_name", "container", "resource"},
 	)
 	EVPAMemoryScaleUpMB = prometheus.NewGaugeVec(
 		prometheus.GaugeOpts{
 			Namespace: "crane",
 			Subsystem: "autoscaling",
-			Name:      "effective_vpa_memory_scale_up_mb",
+			Name:      "effective_vpa_memory_scale_up",
 			Help:      "The memory scale up for Effective VPA",
 		},
-		[]string{
-			"resourceName",
-		},
+		[]string{"apiversion", "owner_kind", "namespace", "owner_name", "container", "resource"},
 	)
 	EVPAMemoryScaleDownMB = prometheus.NewGaugeVec(
 		prometheus.GaugeOpts{
 			Namespace: "crane",
 			Subsystem: "autoscaling",
-			Name:      "effective_vpa_memory_scale_down_mb",
+			Name:      "effective_vpa_memory_scale_down",
 			Help:      "The memory scale down for Effective VPA",
 		},
-		[]string{
-			"resourceName",
+		[]string{"apiversion", "owner_kind", "namespace", "owner_name", "container", "resource"},
+	)
+	EVPAResourceRecommendation = prometheus.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace: "crane",
+			Subsystem: "autoscaling",
+			Name:      "effective_vpa_resource_recommendation",
+			Help:      "The resource recommendation for Effective VPA",
 		},
+		[]string{"apiversion", "owner_kind", "namespace", "owner_name", "container", "resource"},
 	)
 )
 
 func init() {
 	// Register custom metrics with the global prometheus registry
-	metrics.Registry.MustRegister(HPAReplicas, EHPAReplicas, OOMCount, HPAScaleCount, EVPACpuScaleUpMilliCores, EVPACpuScaleDownMilliCores, EVPAMemoryScaleDownMB, EVPAMemoryScaleUpMB)
+	metrics.Registry.MustRegister(HPAReplicas, EHPAReplicas, OOMCount, HPAScaleCount, EVPACpuScaleUpMilliCores, EVPACpuScaleDownMilliCores, EVPAMemoryScaleDownMB, EVPAMemoryScaleUpMB, EVPAResourceRecommendation)
 
 }
 

pkg/prediction/percentile/prediction.go

@@ -282,10 +282,12 @@ func (p *percentilePrediction) Run(stopCh <-chan struct{}) {
 			QueryExpr := qc.MetricNamer.BuildUniqueKey()
 
 			if _, ok := p.queryRoutines.Load(QueryExpr); ok {
+				klog.V(6).InfoS("Prediction percentile routine %v already registered.", "queryExpr", QueryExpr, "caller", qc.Caller)
 				continue
 			}
 
 			if _, ok := p.stopChMap.Load(QueryExpr); ok {
+				klog.V(6).InfoS("Prediction percentile routine %v already stopped.", "queryExpr", QueryExpr, "caller", qc.Caller)
 				continue
 			}
 
@@ -466,6 +468,11 @@ func (p *percentilePrediction) addSamples(namer metricnaming.MetricNamer) {
 		if signal == nil {
 			return
 		}
+		// maybe we can use other aggregated way to deal with the container instances of the same container in different pods of the same workload,
+		// aggregated by reducing all the samples to just one p99 or avg value and so on.
+		// saw that when the workload is daemonset, container in different node has very different resource usage. this is unexpected in production, maybe the daemonset in different nodes has different loads.
+		// NOTE: now it there are N instance of the workload, there are N samples in latest, then the aggregationWindowLength is N times growth to accumulate data fastly.
+		// it is not a time dimension, but we use N samples of different container instances of the workload to represent the N intervals samples
 		for _, ts := range latestTimeSeriesList {
 			if len(ts.Samples) < 1 {
 				klog.V(4).InfoS("Sample not found.", "key", key)