Merge pull request #14163 from rook/mergify/bp/release-1.13/pr-14160

mgr: fix UpdateActiveMgrLabel to retry label update on failure (backport #14160)

cmd/rook/ceph/mgr.go

@@ -96,9 +96,9 @@ func runMgrSidecar(cmd *cobra.Command, args []string) error {
 	clusterInfo.CephVersion = *version
 
 	m := mgr.New(context, &clusterInfo, clusterSpec, "")
-	prevActiveMgr := "unknown"
+	activeMgr := "unknown"
 	for {
-		prevActiveMgr, err = m.UpdateActiveMgrLabel(daemonName, prevActiveMgr)
+		activeMgr, err = m.UpdateActiveMgrLabel(daemonName, activeMgr)
 		if err != nil {
 			logger.Errorf("failed to reconcile services. %v", err)
 		} else {

pkg/operator/ceph/cluster/mgr/mgr.go

@@ -272,6 +272,10 @@ func (c *Cluster) UpdateActiveMgrLabel(daemonNameToUpdate string, prevActiveMgr
 		return "", err // force mrg_role update in the next call
 	}
 
+	// Normally, there should only be one mgr pod with the specific name daemonNameToUpdate. However,
+	// during transitions, there might be additional mgr pods shutting down. To handle this, the code
+	// updates the label mgrRoleLabelName on all mgr pods. If this update fails, the system rolls back
+	// the currently active manager (currActiveMgr). This way the next call will retry the update.
 	for i, pod := range pods.Items {
 
 		labels := pod.GetLabels()
@@ -289,7 +293,10 @@ func (c *Cluster) UpdateActiveMgrLabel(daemonNameToUpdate string, prevActiveMgr
 			pod.SetLabels(labels)
 			_, err = c.context.Clientset.CoreV1().Pods(c.clusterInfo.Namespace).Update(c.clusterInfo.Context, &pods.Items[i], metav1.UpdateOptions{})
 			if err != nil {
-				logger.Infof("cannot update the active mgr pod %q. err=%v", pods.Items[i].Name, err)
+				// In case of failure we report as 'active manager' the previous value, this way
+				// we force refreshing mgrRoleLabelName label next time this function is called
+				currActiveMgr = prevActiveMgr
+				logger.Warningf("cannot update the active mgr pod %q. err=%v", pods.Items[i].Name, err)
 			}
 		}
 	}