adding feature to uncordon node if full drain is unsucessful

Author: jonahjon

pkg/descheduler/node/node.go

@@ -26,6 +26,7 @@ import (
 	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/labels"
+	"k8s.io/apimachinery/pkg/types"
 	clientset "k8s.io/client-go/kubernetes"
 	listersv1 "k8s.io/client-go/listers/core/v1"
 	"k8s.io/client-go/util/workqueue"
@@ -400,3 +401,26 @@ func podMatchesInterPodAntiAffinity(nodeIndexer podutil.GetPodsAssignedToNodeFun
 
 	return false, nil
 }
+
+// UncordonNode removes the Unschedulable flag from a node to allow new pods to be scheduled on it.
+// This is useful when eviction has failed and we want to allow the node to receive new workloads again.
+func UncordonNode(ctx context.Context, client clientset.Interface, node *v1.Node) error {
+	if !node.Spec.Unschedulable {
+		// Node is already uncordoned, nothing to do
+		return nil
+	}
+
+	logger := klog.FromContext(ctx)
+	logger.V(2).InfoS("Uncordoning node", "node", klog.KObj(node))
+
+	// Create a JSON patch to set Unschedulable to false
+	patch := []byte(`[{"op": "replace", "path": "/spec/unschedulable", "value": false}]`)
+	_, err := client.CoreV1().Nodes().Patch(ctx, node.Name, types.JSONPatchType, patch, metav1.PatchOptions{})
+	if err != nil {
+		logger.Error(err, "Failed to uncordon node", "node", klog.KObj(node))
+		return fmt.Errorf("failed to uncordon node %s: %w", node.Name, err)
+	}
+
+	logger.V(1).InfoS("Successfully uncordoned node", "node", klog.KObj(node))
+	return nil
+}

pkg/framework/plugins/nodeutilization/highnodeutilization.go

@@ -275,6 +275,7 @@ func (h *HighNodeUtilization) Balance(ctx context.Context, nodes []*v1.Node) *fr
 		continueEvictionCond,
 		h.usageClient,
 		nil,
+		h.handle,
 	)
 
 	return nil

pkg/framework/plugins/nodeutilization/lownodeutilization.go

@@ -316,6 +316,7 @@ func (l *LowNodeUtilization) Balance(ctx context.Context, nodes []*v1.Node) *fra
 		continueEvictionCond,
 		l.usageClient,
 		nodeLimit,
+		l.handle,
 	)
 
 	return nil

pkg/framework/plugins/nodeutilization/nodeutilization.go

@@ -175,6 +175,7 @@ func evictPodsFromSourceNodes(
 	continueEviction continueEvictionCond,
 	usageClient usageClient,
 	maxNoOfPodsToEvictPerNode *uint,
+	handle frameworktypes.Handle,
 ) {
 	logger := klog.FromContext(ctx)
 	available, err := assessAvailableResourceInNodes(destinationNodes, resourceNames)
@@ -240,6 +241,13 @@ func evictPodsFromSourceNodes(
 			case *evictions.EvictionTotalLimitError:
 				return
 			default:
+				// Eviction failed, uncordon the node to allow new pods to be scheduled
+				if node.node.Spec.Unschedulable {
+					logger.V(1).Info("Eviction failed, uncordoning node", "node", klog.KObj(node.node))
+					if uncordonErr := nodeutil.UncordonNode(ctx, handle.ClientSet(), node.node); uncordonErr != nil {
+						logger.Error(uncordonErr, "Failed to uncordon node after eviction failure", "node", klog.KObj(node.node))
+					}
+				}
 			}
 		}
 	}