Expose classification metrics from LowNodeUtilization plugin

Enhance the plugin interface to support metrics export in a
backward-compatible way. All plugins now have access to the
export setting via handle.ShouldExportMetrics(), without
requiring code changes to existing plugins.

Expose two new metrics from the LowNodeUtilization plugin:
- low_node_utilization_thresholds
- low_node_utilization_classification

These metrics allow users to properly monitor plugin behavior,
which was previously only visible in descheduler logs.

Signed-off-by: Simone Tiraboschi <[email protected]>

Author: tiraboschi

metrics/metrics.go

@@ -92,6 +92,25 @@ var (
 			Buckets:        []float64{0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2.5, 5, 10, 25, 50, 100},
 		}, []string{"strategy", "profile"})
 
+	LowNodeUtilizationThresholds = metrics.NewGaugeVec(
+		&metrics.GaugeOpts{
+			Subsystem:      DeschedulerSubsystem,
+			Name:           "low_node_utilization_thresholds",
+			Help:           "Thresholds used by the LowNodeUtilization to classify nodes by node, by class, by resource",
+			StabilityLevel: metrics.ALPHA,
+		},
+		[]string{"node", "class", "resource"},
+	)
+	LowNodeUtilizationClassification = metrics.NewGaugeVec(
+		&metrics.GaugeOpts{
+			Subsystem:      DeschedulerSubsystem,
+			Name:           "low_node_utilization_classification",
+			Help:           "Number of nodes by class",
+			StabilityLevel: metrics.ALPHA,
+		},
+		[]string{"class"},
+	)
+
 	metricsList = []metrics.Registerable{
 		PodsEvicted,
 		PodsEvictedTotal,
@@ -100,6 +119,8 @@ var (
 		DeschedulerStrategyDuration,
 		LoopDuration,
 		StrategyDuration,
+		LowNodeUtilizationThresholds,
+		LowNodeUtilizationClassification,
 	}
 )
 

pkg/descheduler/descheduler.go

@@ -425,6 +425,7 @@ func (d *descheduler) runProfiles(ctx context.Context, client clientset.Interfac
 			frameworkprofile.WithGetPodsAssignedToNodeFnc(d.getPodsAssignedToNode),
 			frameworkprofile.WithMetricsCollector(d.metricsCollector),
 			frameworkprofile.WithPrometheusClient(d.prometheusClient),
+			frameworkprofile.WithShouldExportMetrics(!d.rs.DisableMetrics),
 		)
 		if err != nil {
 			klog.ErrorS(err, "unable to create a profile", "profile", profile.Name)

pkg/framework/fake/fake.go

@@ -23,6 +23,7 @@ type HandleImpl struct {
 	PodEvictorImpl                *evictions.PodEvictor
 	MetricsCollectorImpl          *metricscollector.MetricsCollector
 	PrometheusClientImpl          promapi.Client
+	ShouldExportMetricsImpl       bool
 }
 
 var _ frameworktypes.Handle = &HandleImpl{}
@@ -62,3 +63,7 @@ func (hi *HandleImpl) PreEvictionFilter(pod *v1.Pod) bool {
 func (hi *HandleImpl) Evict(ctx context.Context, pod *v1.Pod, opts evictions.EvictOptions) error {
 	return hi.PodEvictorImpl.EvictPod(ctx, pod, opts)
 }
+
+func (hi *HandleImpl) ShouldExportMetrics() bool {
+	return hi.ShouldExportMetricsImpl
+}

pkg/framework/plugins/nodeutilization/lownodeutilization.go

@@ -19,11 +19,13 @@ package nodeutilization
 import (
 	"context"
 	"fmt"
+	"strconv"
 
 	v1 "k8s.io/api/core/v1"
 	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/klog/v2"
 
+	"sigs.k8s.io/descheduler/metrics"
 	"sigs.k8s.io/descheduler/pkg/api"
 	"sigs.k8s.io/descheduler/pkg/descheduler/evictions"
 	nodeutil "sigs.k8s.io/descheduler/pkg/descheduler/node"
@@ -179,6 +181,16 @@ func (l *LowNodeUtilization) Balance(ctx context.Context, nodes []*v1.Node) *fra
 		)
 	}
 
+	if l.handle.ShouldExportMetrics() {
+		for key, rt := range thresholds {
+			for i, t := range rt {
+				for resource, value := range t {
+					metrics.LowNodeUtilizationThresholds.WithLabelValues(key, strconv.Itoa(i), resource.String()).Set(float64(value))
+				}
+			}
+		}
+	}
+
 	// classify nodes in under and over utilized. we will later try to move
 	// pods from the overutilized nodes to the underutilized ones.
 	nodeGroups := classifier.Classify(
@@ -255,6 +267,11 @@ func (l *LowNodeUtilization) Balance(ctx context.Context, nodes []*v1.Node) *fra
 	logger.V(1).Info("Criteria for a node above target utilization", l.overCriteria...)
 	logger.V(1).Info("Number of overutilized nodes", "totalNumber", len(highNodes))
 
+	if l.handle.ShouldExportMetrics() {
+		metrics.LowNodeUtilizationClassification.WithLabelValues("underutilized").Set(float64(len(lowNodes)))
+		metrics.LowNodeUtilizationClassification.WithLabelValues("overutilized").Set(float64(len(highNodes)))
+	}
+
 	if len(lowNodes) == 0 {
 		logger.V(1).Info(
 			"No node is underutilized, nothing to do here, you might tune your thresholds further",

pkg/framework/profile/profile.go

@@ -74,6 +74,7 @@ type handleImpl struct {
 	getPodsAssignedToNodeFunc podutil.GetPodsAssignedToNodeFunc
 	sharedInformerFactory     informers.SharedInformerFactory
 	evictor                   *evictorImpl
+	shouldExportMetrics       bool
 }
 
 var _ frameworktypes.Handle = &handleImpl{}
@@ -106,6 +107,11 @@ func (hi *handleImpl) Evictor() frameworktypes.Evictor {
 	return hi.evictor
 }
 
+// ShouldExportMetrics returns whether plugins should export metrics
+func (hi *handleImpl) ShouldExportMetrics() bool {
+	return hi.shouldExportMetrics
+}
+
 type filterPlugin interface {
 	frameworktypes.Plugin
 	Filter(pod *v1.Pod) bool
@@ -142,6 +148,7 @@ type handleImplOpts struct {
 	getPodsAssignedToNodeFunc podutil.GetPodsAssignedToNodeFunc
 	podEvictor                *evictions.PodEvictor
 	metricsCollector          *metricscollector.MetricsCollector
+	shouldExportMetrics       bool
 }
 
 // WithClientSet sets clientSet for the scheduling frameworkImpl.
@@ -182,6 +189,12 @@ func WithMetricsCollector(metricsCollector *metricscollector.MetricsCollector) O
 	}
 }
 
+func WithShouldExportMetrics(shouldExportMetrics bool) Option {
+	return func(o *handleImplOpts) {
+		o.shouldExportMetrics = shouldExportMetrics
+	}
+}
+
 func getPluginConfig(pluginName string, pluginConfigs []api.PluginConfig) (*api.PluginConfig, int) {
 	for idx, pluginConfig := range pluginConfigs {
 		if pluginConfig.Name == pluginName {
@@ -280,8 +293,9 @@ func NewProfile(ctx context.Context, config api.DeschedulerProfile, reg pluginre
 			profileName: config.Name,
 			podEvictor:  hOpts.podEvictor,
 		},
-		metricsCollector: hOpts.metricsCollector,
-		prometheusClient: hOpts.prometheusClient,
+		metricsCollector:    hOpts.metricsCollector,
+		prometheusClient:    hOpts.prometheusClient,
+		shouldExportMetrics: hOpts.shouldExportMetrics,
 	}
 
 	pluginNames := append(config.Plugins.Deschedule.Enabled, config.Plugins.Balance.Enabled...)

pkg/framework/types/types.go

@@ -41,6 +41,8 @@ type Handle interface {
 	GetPodsAssignedToNodeFunc() podutil.GetPodsAssignedToNodeFunc
 	SharedInformerFactory() informers.SharedInformerFactory
 	MetricsCollector() *metricscollector.MetricsCollector
+	// ShouldExportMetrics returns whether plugins should export metrics
+	ShouldExportMetrics() bool
 }
 
 // Evictor defines an interface for filtering and evicting pods