fix: don't mark pods with 'karpenter.sh/do-not-disrupt=true' as resch…

…edulable

pkg/controllers/disruption/drift_test.go

@@ -553,8 +553,7 @@ var _ = Describe("Drift", func() {
 			Expect(ExpectNodes(ctx, env.Client)).To(HaveLen(1))
 			ExpectExists(ctx, env.Client, nodeClaim)
 		})
-		It("should drift nodes that have pods with the karpenter.sh/do-not-disrupt annotation when the NodePool's TerminationGracePeriod is not nil", func() {
-			nodeClaim.Spec.TerminationGracePeriod = &metav1.Duration{Duration: time.Second * 300}
+		It("should ignore nodes that have pods with the karpenter.sh/do-not-disrupt annotation when the NodePool's TerminationGracePeriod is not nil", func() {
 			pod := test.Pod(test.PodOptions{
 				ObjectMeta: metav1.ObjectMeta{
 					Annotations: map[string]string{
@@ -570,8 +569,9 @@ var _ = Describe("Drift", func() {
 
 			ExpectSingletonReconciled(ctx, disruptionController)
 
-			// Expect to create a replacement but not delete the old nodeclaim
-			Expect(ExpectNodeClaims(ctx, env.Client)).To(HaveLen(2)) // new nodeclaim is created for drift
+			// Pods with `karpenter.sh/do-not-disrupt` can't be evicted and hence can't be rescheduled on a new node.
+			// Expect no new nodeclaims to be created.
+			Expect(ExpectNodeClaims(ctx, env.Client)).To(HaveLen(1))
 			Expect(ExpectNodes(ctx, env.Client)).To(HaveLen(1))
 			ExpectExists(ctx, env.Client, nodeClaim)
 		})

pkg/utils/pod/scheduling.go

@@ -39,13 +39,15 @@ func IsActive(pod *corev1.Pod) bool {
 // - Is an active pod (isn't terminal or actively terminating) OR Is owned by a StatefulSet and Is Terminating
 // - Isn't owned by a DaemonSet
 // - Isn't a mirror pod (https://kubernetes.io/docs/tasks/configure-pod-container/static-pod/)
+// - Does not have the "karpenter.sh/do-not-disrupt=true" annotation (https://karpenter.sh/docs/concepts/disruption/#pod-level-controls)
 func IsReschedulable(pod *corev1.Pod) bool {
 	// StatefulSet pods can be handled differently here because we know that StatefulSet pods MUST
 	// get deleted before new pods are re-created. This means that we can model terminating pods for StatefulSets
 	// differently for higher availability by considering terminating pods for scheduling
 	return (IsActive(pod) || (IsOwnedByStatefulSet(pod) && IsTerminating(pod))) &&
 		!IsOwnedByDaemonSet(pod) &&
-		!IsOwnedByNode(pod)
+		!IsOwnedByNode(pod) &&
+		!HasDoNotDisrupt(pod)
 }
 
 // IsEvictable checks if a pod is evictable by Karpenter by ensuring that the pod: