Fix node tainting when using UseProviderId with SQS monitors (#1203)

Author: Sithembiso-Mashinini

pkg/monitor/sqsevent/asg-lifecycle-event.go

@@ -126,7 +126,18 @@ func (m SQSMonitor) asgTerminationToInterruptionEvent(event *EventBridgeEvent, m
 			go m.SendHeartbeats(nthConfig.HeartbeatInterval, nthConfig.HeartbeatUntil, lifecycleDetail, stopHeartbeatCh, cancelHeartbeatCh)
 		}
 
-		err := n.TaintASGLifecycleTermination(interruptionEvent.NodeName, interruptionEvent.EventID)
+		// Use provider ID to resolve the actual Kubernetes node name if UseProviderId is configured
+		nodeName := interruptionEvent.NodeName
+		if nthConfig.UseProviderId && interruptionEvent.ProviderID != "" {
+			resolvedNodeName, err := n.GetNodeNameFromProviderID(interruptionEvent.ProviderID)
+			if err != nil {
+				log.Warn().Err(err).Str("provider_id", interruptionEvent.ProviderID).Msg("Failed to resolve node name from provider ID, falling back to NodeName from event")
+			} else {
+				nodeName = resolvedNodeName
+			}
+		}
+
+		err := n.TaintASGLifecycleTermination(nodeName, interruptionEvent.EventID)
 		if err != nil {
 			log.Err(err).Msgf("unable to taint node with taint %s:%s", node.ASGLifecycleTerminationTaint, interruptionEvent.EventID)
 		}

pkg/monitor/sqsevent/rebalance-recommendation-event.go

@@ -78,7 +78,19 @@ func (m SQSMonitor) rebalanceRecommendationToInterruptionEvent(event *EventBridg
 		return nil
 	}
 	interruptionEvent.PreDrainTask = func(interruptionEvent monitor.InterruptionEvent, n node.Node) error {
-		err := n.TaintRebalanceRecommendation(interruptionEvent.NodeName, interruptionEvent.EventID)
+		// Use provider ID to resolve the actual Kubernetes node name if UseProviderId is configured
+		nthConfig := n.GetNthConfig()
+		nodeName := interruptionEvent.NodeName
+		if nthConfig.UseProviderId && interruptionEvent.ProviderID != "" {
+			resolvedNodeName, err := n.GetNodeNameFromProviderID(interruptionEvent.ProviderID)
+			if err != nil {
+				log.Warn().Err(err).Str("provider_id", interruptionEvent.ProviderID).Msg("Failed to resolve node name from provider ID, falling back to NodeName from event")
+			} else {
+				nodeName = resolvedNodeName
+			}
+		}
+
+		err := n.TaintRebalanceRecommendation(nodeName, interruptionEvent.EventID)
 		if err != nil {
 			log.Err(err).Msgf("Unable to taint node with taint %s:%s", node.RebalanceRecommendationTaint, interruptionEvent.EventID)
 		}

pkg/monitor/sqsevent/scheduled-change-event.go

@@ -113,7 +113,19 @@ func (m SQSMonitor) scheduledEventToInterruptionEvents(event *EventBridgeEvent,
 			return nil
 		}
 		interruptionEvent.PreDrainTask = func(interruptionEvent monitor.InterruptionEvent, n node.Node) error {
-			if err := n.TaintScheduledMaintenance(interruptionEvent.NodeName, interruptionEvent.EventID); err != nil {
+			// Use provider ID to resolve the actual Kubernetes node name if UseProviderId is configured
+			nthConfig := n.GetNthConfig()
+			nodeName := interruptionEvent.NodeName
+			if nthConfig.UseProviderId && interruptionEvent.ProviderID != "" {
+				resolvedNodeName, err := n.GetNodeNameFromProviderID(interruptionEvent.ProviderID)
+				if err != nil {
+					log.Warn().Err(err).Str("provider_id", interruptionEvent.ProviderID).Msg("Failed to resolve node name from provider ID, falling back to NodeName from event")
+				} else {
+					nodeName = resolvedNodeName
+				}
+			}
+
+			if err := n.TaintScheduledMaintenance(nodeName, interruptionEvent.EventID); err != nil {
 				log.Err(err).Msgf("Unable to taint node with taint %s:%s", node.ScheduledMaintenanceTaint, interruptionEvent.EventID)
 			}
 			return nil

pkg/monitor/sqsevent/spot-itn-event.go

@@ -80,7 +80,19 @@ func (m SQSMonitor) spotITNTerminationToInterruptionEvent(event *EventBridgeEven
 		return nil
 	}
 	interruptionEvent.PreDrainTask = func(interruptionEvent monitor.InterruptionEvent, n node.Node) error {
-		err := n.TaintSpotItn(interruptionEvent.NodeName, interruptionEvent.EventID)
+		// Use provider ID to resolve the actual Kubernetes node name if UseProviderId is configured
+		nthConfig := n.GetNthConfig()
+		nodeName := interruptionEvent.NodeName
+		if nthConfig.UseProviderId && interruptionEvent.ProviderID != "" {
+			resolvedNodeName, err := n.GetNodeNameFromProviderID(interruptionEvent.ProviderID)
+			if err != nil {
+				log.Warn().Err(err).Str("provider_id", interruptionEvent.ProviderID).Msg("Failed to resolve node name from provider ID, falling back to NodeName from event")
+			} else {
+				nodeName = resolvedNodeName
+			}
+		}
+
+		err := n.TaintSpotItn(nodeName, interruptionEvent.EventID)
 		if err != nil {
 			log.Err(err).Msgf("Unable to taint node with taint %s:%s", node.SpotInterruptionTaint, interruptionEvent.EventID)
 		}

pkg/node/node.go

@@ -662,6 +662,7 @@ func (n Node) fetchKubernetesNode(nodeName string) (*corev1.Node, error) {
 			},
 		},
 	}
+
 	listOptions := metav1.ListOptions{LabelSelector: metav1.FormatLabelSelector(&labelSelector)}
 	matchingNodes, err := n.drainHelper.Client.CoreV1().Nodes().List(context.TODO(), listOptions)
 	if err != nil || len(matchingNodes.Items) == 0 {