Implemented UnschedulablePodsCount metric

pkg/controllers/provisioning/provisioner.go

@@ -345,6 +345,9 @@ func (p *Provisioner) Schedule(ctx context.Context) (scheduler.Results, error) {
 		return scheduler.Results{}, fmt.Errorf("creating scheduler, %w", err)
 	}
 	results := s.Solve(ctx, pods).TruncateInstanceTypes(scheduler.MaxInstanceTypes)
+	scheduler.UnschedulablePodsCount.With(
+		prometheus.Labels{scheduler.ControllerLabel: injection.GetControllerName(ctx)},
+	).Set(float64(len(results.PodErrors)))
 	if len(results.NewNodeClaims) > 0 {
 		log.FromContext(ctx).WithValues("Pods", pretty.Slice(lo.Map(pods, func(p *corev1.Pod, _ int) string { return klog.KRef(p.Namespace, p.Name).String() }), 5), "duration", time.Since(start)).Info("found provisionable pod(s)")
 	}

pkg/controllers/provisioning/scheduling/metrics.go

@@ -24,7 +24,7 @@ import (
 )
 
 func init() {
-	crmetrics.Registry.MustRegister(SchedulingDurationSeconds, QueueDepth, IgnoredPodCount)
+	crmetrics.Registry.MustRegister(SchedulingDurationSeconds, QueueDepth, IgnoredPodCount, UnschedulablePodsCount)
 }
 
 const (
@@ -65,4 +65,15 @@ var (
 			Help:      "Number of pods ignored during scheduling by Karpenter",
 		},
 	)
+	UnschedulablePodsCount = prometheus.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace: metrics.Namespace,
+			Subsystem: schedulerSubsystem,
+			Name:      "unschedulable_pods_count",
+			Help:      "The number of unschedulable Pods.",
+		},
+		[]string{
+			ControllerLabel,
+		},
+	)
 )

pkg/controllers/provisioning/scheduling/scheduler.go

@@ -203,7 +203,9 @@ func (s *Scheduler) Solve(ctx context.Context, pods []*corev1.Pod) Results {
 	// had 5xA pods and 5xB pods were they have a zonal topology spread, but A can only go in one zone and B in another.
 	// We need to schedule them alternating, A, B, A, B, .... and this solution also solves that as well.
 	errors := map[*corev1.Pod]error{}
-	QueueDepth.DeletePartialMatch(prometheus.Labels{ControllerLabel: injection.GetControllerName(ctx)}) // Reset the metric for the controller, so we don't keep old ids around
+	// Reset the metric for the controller, so we don't keep old ids around
+	UnschedulablePodsCount.DeletePartialMatch(prometheus.Labels{ControllerLabel: injection.GetControllerName(ctx)})
+	QueueDepth.DeletePartialMatch(prometheus.Labels{ControllerLabel: injection.GetControllerName(ctx)})
 	q := NewQueue(pods...)
 	for {
 		QueueDepth.With(

pkg/controllers/provisioning/scheduling/suite_test.go

@@ -112,6 +112,7 @@ var _ = AfterEach(func() {
 	cluster.Reset()
 	scheduling.QueueDepth.Reset()
 	scheduling.SchedulingDurationSeconds.Reset()
+	scheduling.UnschedulablePodsCount.Reset()
 })
 
 var _ = Context("Scheduling", func() {
@@ -3676,6 +3677,43 @@ var _ = Context("Scheduling", func() {
 			s.Solve(injection.WithControllerName(ctx, "provisioner"), pods)
 			wg.Wait()
 		})
+		It("should surface the UnschedulablePodsCount metric while executing the scheduling loop", func() {
+			nodePool := test.NodePool(v1.NodePool{
+				Spec: v1.NodePoolSpec{
+					Template: v1.NodeClaimTemplate{
+						Spec: v1.NodeClaimTemplateSpec{
+							Requirements: []v1.NodeSelectorRequirementWithMinValues{
+								{
+									NodeSelectorRequirement: corev1.NodeSelectorRequirement{
+										Key:      corev1.LabelInstanceTypeStable,
+										Operator: corev1.NodeSelectorOpIn,
+										Values: []string{
+											"default-instance-type",
+										},
+									},
+								},
+							},
+						},
+					},
+				},
+			})
+			ExpectApplied(ctx, env.Client, nodePool)
+			//Creates 15 pods, 5 schedulable and 10 unschedulable
+			podsUnschedulable := test.UnschedulablePods(test.PodOptions{NodeSelector: map[string]string{corev1.LabelInstanceTypeStable: "unknown"}}, 10)
+			podsSchedulable := test.UnschedulablePods(test.PodOptions{NodeSelector: map[string]string{corev1.LabelInstanceTypeStable: "default-instance-type"}}, 5)
+			pods := append(podsUnschedulable, podsSchedulable...)
+			ExpectApplied(ctx, env.Client, nodePool)
+			//Adds UID to pods for queue in solve. Solve pushes any unschedulable pod back onto the queue and
+			//then maps the current length of the queue to the pod using the UID
+			for _, i := range pods {
+				ExpectApplied(ctx, env.Client, i)
+			}
+			_, err := prov.Schedule(injection.WithControllerName(ctx, "provisioner"))
+			m, ok := FindMetricWithLabelValues("karpenter_scheduler_unschedulable_pods_count", map[string]string{"controller": "provisioner"})
+			Expect(ok).To(BeTrue())
+			Expect(lo.FromPtr(m.Gauge.Value)).To(BeNumerically("==", 10))
+			Expect(err).To(BeNil())
+		})
 		It("should surface the schedulingDuration metric after executing a scheduling loop", func() {
 			nodePool := test.NodePool()
 			ExpectApplied(ctx, env.Client, nodePool)