Merge pull request #8099 from x13n/estimate-metric

Introduce a metric measuring heterogeneity of binpacking

cluster-autoscaler/estimator/binpacking_estimator.go

@@ -18,9 +18,11 @@ package estimator
 
 import (
 	"fmt"
+	"strconv"
 
 	apiv1 "k8s.io/api/core/v1"
 	"k8s.io/autoscaler/cluster-autoscaler/cloudprovider"
+	"k8s.io/autoscaler/cluster-autoscaler/metrics"
 	core_utils "k8s.io/autoscaler/cluster-autoscaler/simulator"
 	"k8s.io/autoscaler/cluster-autoscaler/simulator/clustersnapshot"
 	"k8s.io/autoscaler/cluster-autoscaler/simulator/framework"
@@ -93,6 +95,7 @@ func (e *BinpackingNodeEstimator) Estimate(
 	nodeTemplate *framework.NodeInfo,
 	nodeGroup cloudprovider.NodeGroup,
 ) (int, []*apiv1.Pod) {
+	observeBinpackingHeterogeneity(podsEquivalenceGroups, nodeTemplate)
 
 	e.limiter.StartEstimation(podsEquivalenceGroups, nodeGroup, e.context)
 	defer e.limiter.EndEstimation()
@@ -231,3 +234,31 @@ func (e *BinpackingNodeEstimator) addNewNodeToSnapshot(
 	estimationState.newNodeNames[estimationState.lastNodeName] = true
 	return nil
 }
+
+func observeBinpackingHeterogeneity(podsEquivalenceGroups []PodEquivalenceGroup, nodeTemplate *framework.NodeInfo) {
+	node := nodeTemplate.Node()
+	var instanceType, cpuCount string
+	if node != nil {
+		if node.Labels != nil {
+			instanceType = node.Labels[apiv1.LabelInstanceTypeStable]
+		}
+		cpuCount = node.Status.Capacity.Cpu().String()
+	}
+	namespaces := make(map[string]bool)
+	for _, peg := range podsEquivalenceGroups {
+		e := peg.Exemplar()
+		if e != nil {
+			namespaces[e.Namespace] = true
+		}
+	}
+	// Quantize # of namespaces to limit metric cardinality.
+	nsCountBucket := ""
+	if len(namespaces) <= 5 {
+		nsCountBucket = strconv.Itoa(len(namespaces))
+	} else if len(namespaces) <= 10 {
+		nsCountBucket = "6-10"
+	} else {
+		nsCountBucket = "11+"
+	}
+	metrics.ObserveBinpackingHeterogeneity(instanceType, cpuCount, nsCountBucket, len(podsEquivalenceGroups))
+}

cluster-autoscaler/metrics/metrics.go

@@ -418,6 +418,15 @@ var (
 			Help:      "Number of migs where instance count according to InstanceGroupManagers.List() differs from the results of Instances.List(). This can happen when some instances are abandoned or a user edits instance 'created-by' metadata.",
 		},
 	)
+
+	binpackingHeterogeneity = k8smetrics.NewHistogramVec(
+		&k8smetrics.HistogramOpts{
+			Namespace: caNamespace,
+			Name:      "binpacking_heterogeneity",
+			Help:      "Number of groups of equivalent pods being processed as a part of the same binpacking simulation.",
+			Buckets:   []float64{1, 2, 4, 6, 10},
+		}, []string{"instance_type", "cpu_count", "namespace_count"},
+	)
 )
 
 // RegisterAll registers all metrics.
@@ -453,6 +462,7 @@ func RegisterAll(emitPerNodeGroupMetrics bool) {
 	legacyregistry.MustRegister(pendingNodeDeletions)
 	legacyregistry.MustRegister(nodeTaintsCount)
 	legacyregistry.MustRegister(inconsistentInstancesMigsCount)
+	legacyregistry.MustRegister(binpackingHeterogeneity)
 
 	if emitPerNodeGroupMetrics {
 		legacyregistry.MustRegister(nodesGroupMinNodes)
@@ -736,3 +746,9 @@ func ObserveNodeTaintsCount(taintType string, count float64) {
 func UpdateInconsistentInstancesMigsCount(migCount int) {
 	inconsistentInstancesMigsCount.Set(float64(migCount))
 }
+
+// ObserveBinpackingHeterogeneity records the number of pod equivalence groups
+// considered in a single binpacking estimation.
+func ObserveBinpackingHeterogeneity(instanceType, cpuCount, namespaceCount string, pegCount int) {
+	binpackingHeterogeneity.WithLabelValues(instanceType, cpuCount, namespaceCount).Observe(float64(pegCount))
+}