fix: add cooldown to prevent resetting autoheal exp backoff preemptively (cherry-pick #23057)

cmd/argocd-application-controller/commands/argocd_application_controller.go

@@ -68,6 +68,7 @@ func NewCommand() *cobra.Command {
 		selfHealBackoffTimeoutSeconds    int
 		selfHealBackoffFactor            int
 		selfHealBackoffCapSeconds        int
+		selfHealBackoffCooldownSeconds   int
 		syncTimeout                      int
 		statusProcessors                 int
 		operationProcessors              int
@@ -201,6 +202,7 @@ func NewCommand() *cobra.Command {
 				time.Duration(appResyncJitter)*time.Second,
 				time.Duration(selfHealTimeoutSeconds)*time.Second,
 				selfHealBackoff,
+				time.Duration(selfHealBackoffCooldownSeconds)*time.Second,
 				time.Duration(syncTimeout)*time.Second,
 				time.Duration(repoErrorGracePeriod)*time.Second,
 				metricsPort,
@@ -272,6 +274,7 @@ func NewCommand() *cobra.Command {
 	command.Flags().IntVar(&selfHealBackoffTimeoutSeconds, "self-heal-backoff-timeout-seconds", env.ParseNumFromEnv("ARGOCD_APPLICATION_CONTROLLER_SELF_HEAL_BACKOFF_TIMEOUT_SECONDS", 2, 0, math.MaxInt32), "Specifies initial timeout of exponential backoff between self heal attempts")
 	command.Flags().IntVar(&selfHealBackoffFactor, "self-heal-backoff-factor", env.ParseNumFromEnv("ARGOCD_APPLICATION_CONTROLLER_SELF_HEAL_BACKOFF_FACTOR", 3, 0, math.MaxInt32), "Specifies factor of exponential timeout between application self heal attempts")
 	command.Flags().IntVar(&selfHealBackoffCapSeconds, "self-heal-backoff-cap-seconds", env.ParseNumFromEnv("ARGOCD_APPLICATION_CONTROLLER_SELF_HEAL_BACKOFF_CAP_SECONDS", 300, 0, math.MaxInt32), "Specifies max timeout of exponential backoff between application self heal attempts")
+	command.Flags().IntVar(&selfHealBackoffCooldownSeconds, "self-heal-backoff-cooldown-seconds", env.ParseNumFromEnv("ARGOCD_APPLICATION_CONTROLLER_SELF_HEAL_BACKOFF_COOLDOWN_SECONDS", 330, 0, math.MaxInt32), "Specifies period of time the app needs to stay synced before the self heal backoff can reset")
 	command.Flags().IntVar(&syncTimeout, "sync-timeout", env.ParseNumFromEnv("ARGOCD_APPLICATION_CONTROLLER_SYNC_TIMEOUT", 0, 0, math.MaxInt32), "Specifies the timeout after which a sync would be terminated. 0 means no timeout (default 0).")
 	command.Flags().Int64Var(&kubectlParallelismLimit, "kubectl-parallelism-limit", env.ParseInt64FromEnv("ARGOCD_APPLICATION_CONTROLLER_KUBECTL_PARALLELISM_LIMIT", 20, 0, math.MaxInt64), "Number of allowed concurrent kubectl fork/execs. Any value less than 1 means no limit.")
 	command.Flags().BoolVar(&repoServerPlaintext, "repo-server-plaintext", env.ParseBoolFromEnv("ARGOCD_APPLICATION_CONTROLLER_REPO_SERVER_PLAINTEXT", false), "Disable TLS on connections to repo server")

controller/appcontroller.go

@@ -41,6 +41,7 @@ import (
 	"k8s.io/client-go/kubernetes"
 	"k8s.io/client-go/tools/cache"
 	"k8s.io/client-go/util/workqueue"
+	"k8s.io/utils/ptr"
 
 	commitclient "github.com/argoproj/argo-cd/v3/commitserver/apiclient"
 	"github.com/argoproj/argo-cd/v3/common"
@@ -133,6 +134,7 @@ type ApplicationController struct {
 	statusRefreshJitter           time.Duration
 	selfHealTimeout               time.Duration
 	selfHealBackOff               *wait.Backoff
+	selfHealBackoffCooldown       time.Duration
 	syncTimeout                   time.Duration
 	db                            db.ArgoDB
 	settingsMgr                   *settings_util.SettingsManager
@@ -168,6 +170,7 @@ func NewApplicationController(
 	appResyncJitter time.Duration,
 	selfHealTimeout time.Duration,
 	selfHealBackoff *wait.Backoff,
+	selfHealBackoffCooldown time.Duration,
 	syncTimeout time.Duration,
 	repoErrorGracePeriod time.Duration,
 	metricsPort int,
@@ -214,6 +217,7 @@ func NewApplicationController(
 		settingsMgr:                       settingsMgr,
 		selfHealTimeout:                   selfHealTimeout,
 		selfHealBackOff:                   selfHealBackoff,
+		selfHealBackoffCooldown:           selfHealBackoffCooldown,
 		syncTimeout:                       syncTimeout,
 		clusterSharding:                   clusterSharding,
 		projByNameCache:                   sync.Map{},
@@ -2245,17 +2249,22 @@ func (ctrl *ApplicationController) shouldSelfHeal(app *appv1.Application, alread
 		return true, time.Duration(0)
 	}
 
-	// Reset counter if the prior sync was successful OR if the revision has changed
-	if !alreadyAttempted || app.Status.Sync.Status == appv1.SyncStatusCodeSynced {
+	var timeSinceOperation *time.Duration
+	if app.Status.OperationState.FinishedAt != nil {
+		timeSinceOperation = ptr.To(time.Since(app.Status.OperationState.FinishedAt.Time))
+	}
+
+	// Reset counter if the prior sync was successful and the cooldown period is over OR if the revision has changed
+	if !alreadyAttempted || (timeSinceOperation != nil && *timeSinceOperation >= ctrl.selfHealBackoffCooldown && app.Status.Sync.Status == appv1.SyncStatusCodeSynced) {
 		app.Status.OperationState.Operation.Sync.SelfHealAttemptsCount = 0
 	}
 
 	var retryAfter time.Duration
 	if ctrl.selfHealBackOff == nil {
-		if app.Status.OperationState.FinishedAt == nil {
+		if timeSinceOperation == nil {
 			retryAfter = ctrl.selfHealTimeout
 		} else {
-			retryAfter = ctrl.selfHealTimeout - time.Since(app.Status.OperationState.FinishedAt.Time)
+			retryAfter = ctrl.selfHealTimeout - *timeSinceOperation
 		}
 	} else {
 		backOff := *ctrl.selfHealBackOff
@@ -2265,10 +2274,11 @@ func (ctrl *ApplicationController) shouldSelfHeal(app *appv1.Application, alread
 		for i := 0; i < steps; i++ {
 			delay = backOff.Step()
 		}
-		if app.Status.OperationState.FinishedAt == nil {
+
+		if timeSinceOperation == nil {
 			retryAfter = delay
 		} else {
-			retryAfter = delay - time.Since(app.Status.OperationState.FinishedAt.Time)
+			retryAfter = delay - *timeSinceOperation
 		}
 	}
 	return retryAfter <= 0, retryAfter

controller/appcontroller_test.go

@@ -172,6 +172,7 @@ func newFakeControllerWithResync(data *fakeData, appResyncPeriod time.Duration,
 		time.Second,
 		time.Minute,
 		nil,
+		time.Minute,
 		0,
 		time.Second*10,
 		common.DefaultPortArgoCDMetrics,
@@ -2607,10 +2608,18 @@ func TestSelfHealExponentialBackoff(t *testing.T) {
 		alreadyAttempted: false,
 		expectedAttempts: 0,
 		syncStatus:       v1alpha1.SyncStatusCodeOutOfSync,
-	}, {
+	}, { // backoff will not reset as finished tme isn't >= cooldown
 		attempts:         6,
-		finishedAt:       nil,
-		expectedDuration: 0,
+		finishedAt:       ptr.To(metav1.Now()),
+		expectedDuration: 120 * time.Second,
+		shouldSelfHeal:   false,
+		alreadyAttempted: true,
+		expectedAttempts: 6,
+		syncStatus:       v1alpha1.SyncStatusCodeSynced,
+	}, { // backoff will reset as finished time is >= cooldown
+		attempts:         40,
+		finishedAt:       &metav1.Time{Time: time.Now().Add(-(1 * time.Minute))},
+		expectedDuration: -60 * time.Second,
 		shouldSelfHeal:   true,
 		alreadyAttempted: true,
 		expectedAttempts: 0,

manifests/base/application-controller-deployment/argocd-application-controller-deployment.yaml

@@ -121,6 +121,12 @@ spec:
               name: argocd-cmd-params-cm
               key: controller.self.heal.backoff.cap.seconds
               optional: true
+        - name: ARGOCD_APPLICATION_CONTROLLER_SELF_HEAL_BACKOFF_COOLDOWN_SECONDS
+          valueFrom:
+            configMapKeyRef:
+              name: argocd-cmd-params-cm
+              key: controller.self.heal.backoff.cooldown.seconds
+              optional: true
         - name: ARGOCD_APPLICATION_CONTROLLER_SYNC_TIMEOUT
           valueFrom:
             configMapKeyRef:

manifests/base/application-controller/argocd-application-controller-statefulset.yaml

@@ -124,6 +124,12 @@ spec:
               name: argocd-cmd-params-cm
               key: controller.self.heal.backoff.cap.seconds
               optional: true
+        - name: ARGOCD_APPLICATION_CONTROLLER_SELF_HEAL_BACKOFF_COOLDOWN_SECONDS
+          valueFrom:
+            configMapKeyRef:
+              name: argocd-cmd-params-cm
+              key: controller.self.heal.backoff.cooldown.seconds
+              optional: true
         - name: ARGOCD_APPLICATION_CONTROLLER_SYNC_TIMEOUT
           valueFrom:
             configMapKeyRef: