[RayService] Avoid Duplicate Serve Service (ray-project#1867)

MortalHappiness · Jan 30, 2024 · edd332b · edd332b
1 parent acafbfe
commit edd332b
Show file tree

Hide file tree

Showing 5 changed files with 45 additions and 60 deletions.
diff --git a/ray-operator/controllers/ray/common/pod.go b/ray-operator/controllers/ray/common/pod.go
@@ -243,7 +243,7 @@ func DefaultWorkerPodTemplate(ctx context.Context, instance rayv1.RayCluster, wo
 	return podTemplate
 }
 
-func initLivenessAndReadinessProbe(rayContainer *corev1.Container, rayNodeType rayv1.RayNodeType, enableServeService bool) {
+func initLivenessAndReadinessProbe(rayContainer *corev1.Container, rayNodeType rayv1.RayNodeType, creator string) {
 	rayAgentRayletHealthCommand := fmt.Sprintf(utils.BaseWgetHealthCommand, utils.DefaultDashboardAgentListenPort, utils.RayAgentRayletHealthPath)
 	rayDashboardGCSHealthCommand := fmt.Sprintf(utils.BaseWgetHealthCommand, utils.DefaultDashboardPort, utils.RayDashboardGCSHealthPath)
 
@@ -281,7 +281,7 @@ func initLivenessAndReadinessProbe(rayContainer *corev1.Container, rayNodeType r
 		// For worker Pods serving traffic, we need to add an additional HTTP proxy health check for the readiness probe.
 		// Note: head Pod checks the HTTP proxy's health at every rayservice controller reconcile instaed of using readiness probe.
 		// See https://github.com/ray-project/kuberay/pull/1808 for reasons.
-		if enableServeService && rayNodeType == rayv1.WorkerNode {
+		if strings.EqualFold(creator, string(utils.RayServiceCRD)) && rayNodeType == rayv1.WorkerNode {
 			rayContainer.ReadinessProbe.FailureThreshold = utils.ServeReadinessProbeFailureThreshold
 			rayServeProxyHealthCommand := fmt.Sprintf(utils.BaseWgetHealthCommand,
 				utils.FindContainerPort(rayContainer, utils.ServingPortName, utils.DefaultServingPort), utils.RayServeProxyHealthPath)
@@ -292,14 +292,14 @@ func initLivenessAndReadinessProbe(rayContainer *corev1.Container, rayNodeType r
 }
 
 // BuildPod a pod config
-func BuildPod(ctx context.Context, podTemplateSpec corev1.PodTemplateSpec, rayNodeType rayv1.RayNodeType, rayStartParams map[string]string, headPort string, enableRayAutoscaler *bool, creator string, fqdnRayIP string, enableServeService bool) (aPod corev1.Pod) {
+func BuildPod(ctx context.Context, podTemplateSpec corev1.PodTemplateSpec, rayNodeType rayv1.RayNodeType, rayStartParams map[string]string, headPort string, enableRayAutoscaler *bool, creator string, fqdnRayIP string) (aPod corev1.Pod) {
+	log := ctrl.LoggerFrom(ctx)
+
 	// For Worker Pod: Traffic readiness is determined by the readiness probe.
 	// Therefore, the RayClusterServingServiceLabelKey label is not utilized and should always be set to true.
 	// For Head Pod: Traffic readiness is determined by the value of the RayClusterServingServiceLabelKey label.
 	// Initially, set the label to false and let the rayservice controller to manage its value.
-	log := ctrl.LoggerFrom(ctx)
-
-	if enableServeService {
+	if strings.EqualFold(creator, string(utils.RayServiceCRD)) {
 		podTemplateSpec.Labels[utils.RayClusterServingServiceLabelKey] = utils.EnableRayClusterServingServiceTrue
 		if rayNodeType == rayv1.HeadNode {
 			podTemplateSpec.Labels[utils.RayClusterServingServiceLabelKey] = utils.EnableRayClusterServingServiceFalse
@@ -380,7 +380,7 @@ func BuildPod(ctx context.Context, podTemplateSpec corev1.PodTemplateSpec, rayNo
 		// Configure the readiness and liveness probes for the Ray container. These probes
 		// play a crucial role in KubeRay health checks. Without them, certain failures,
 		// such as the Raylet process crashing, may go undetected.
-		initLivenessAndReadinessProbe(&pod.Spec.Containers[utils.RayContainerIndex], rayNodeType, enableServeService)
+		initLivenessAndReadinessProbe(&pod.Spec.Containers[utils.RayContainerIndex], rayNodeType, creator)
 	}
 
 	return pod

diff --git a/ray-operator/controllers/ray/common/pod_test.go b/ray-operator/controllers/ray/common/pod_test.go
@@ -335,11 +335,7 @@ func TestBuildPod(t *testing.T) {
 	// Test head pod
 	podName := strings.ToLower(cluster.Name + utils.DashSymbol + string(rayv1.HeadNode) + utils.DashSymbol + utils.FormatInt32(0))
 	podTemplateSpec := DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "", true)
-
-	val, ok := pod.Labels[utils.RayClusterServingServiceLabelKey]
-	assert.True(t, ok, "Expected serve label is not present")
-	assert.Equal(t, utils.EnableRayClusterServingServiceFalse, val, "Wrong serve label value")
+	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "")
 
 	// Check environment variables
 	rayContainer := pod.Spec.Containers[utils.RayContainerIndex]
@@ -392,11 +388,7 @@ func TestBuildPod(t *testing.T) {
 	podName = cluster.Name + utils.DashSymbol + string(rayv1.WorkerNode) + utils.DashSymbol + worker.GroupName + utils.DashSymbol + utils.FormatInt32(0)
 	fqdnRayIP := utils.GenerateFQDNServiceName(ctx, *cluster, cluster.Namespace)
 	podTemplateSpec = DefaultWorkerPodTemplate(ctx, *cluster, worker, podName, fqdnRayIP, "6379")
-	pod = BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP, true)
-
-	val, ok = pod.Labels[utils.RayClusterServingServiceLabelKey]
-	assert.True(t, ok, "Expected serve label is not present")
-	assert.Equal(t, utils.EnableRayClusterServingServiceTrue, val, "Wrong serve label value")
+	pod = BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP)
 
 	// Check environment variables
 	rayContainer = pod.Spec.Containers[utils.RayContainerIndex]
@@ -435,7 +427,7 @@ func TestBuildPod_WithOverwriteCommand(t *testing.T) {
 
 	podName := strings.ToLower(cluster.Name + utils.DashSymbol + string(rayv1.HeadNode) + utils.DashSymbol + utils.FormatInt32(0))
 	podTemplateSpec := DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	headPod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "", false)
+	headPod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "")
 	headContainer := headPod.Spec.Containers[utils.RayContainerIndex]
 	assert.Equal(t, headContainer.Command, []string{"I am head"})
 	assert.Equal(t, headContainer.Args, []string{"I am head again"})
@@ -444,7 +436,7 @@ func TestBuildPod_WithOverwriteCommand(t *testing.T) {
 	podName = cluster.Name + utils.DashSymbol + string(rayv1.WorkerNode) + utils.DashSymbol + worker.GroupName + utils.DashSymbol + utils.FormatInt32(0)
 	fqdnRayIP := utils.GenerateFQDNServiceName(ctx, *cluster, cluster.Namespace)
 	podTemplateSpec = DefaultWorkerPodTemplate(ctx, *cluster, worker, podName, fqdnRayIP, "6379")
-	workerPod := BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP, false)
+	workerPod := BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP)
 	workerContainer := workerPod.Spec.Containers[utils.RayContainerIndex]
 	assert.Equal(t, workerContainer.Command, []string{"I am worker"})
 	assert.Equal(t, workerContainer.Args, []string{"I am worker again"})
@@ -456,7 +448,7 @@ func TestBuildPod_WithAutoscalerEnabled(t *testing.T) {
 	cluster.Spec.EnableInTreeAutoscaling = &trueFlag
 	podName := strings.ToLower(cluster.Name + utils.DashSymbol + string(rayv1.HeadNode) + utils.DashSymbol + utils.FormatInt32(0))
 	podTemplateSpec := DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", &trueFlag, "", "", false)
+	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", &trueFlag, "", "")
 
 	actualResult := pod.Labels[utils.RayClusterLabelKey]
 	expectedResult := cluster.Name
@@ -507,29 +499,29 @@ func TestBuildPod_WithAutoscalerEnabled(t *testing.T) {
 }
 
 func TestBuildPod_WithCreatedByRayService(t *testing.T) {
+	ctx := context.Background()
+
 	cluster := instance.DeepCopy()
 	cluster.Spec.EnableInTreeAutoscaling = &trueFlag
 	podName := strings.ToLower(cluster.Name + utils.DashSymbol + string(rayv1.HeadNode) + utils.DashSymbol + utils.FormatInt32(0))
-	podTemplateSpec := DefaultHeadPodTemplate(context.Background(), *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	pod := BuildPod(context.Background(), podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", &trueFlag, string(utils.RayServiceCRD), "", false)
+	podTemplateSpec := DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
+	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", &trueFlag, string(utils.RayServiceCRD), "")
 
-	hasCorrectDeathEnv := false
-	for _, container := range pod.Spec.Containers {
-		if container.Name != "ray-head" {
-			continue
-		}
-		if container.Env == nil || len(container.Env) == 0 {
-			t.Fatalf("Expected death env `%v`", container)
-		}
-		for _, env := range container.Env {
-			if env.Name == utils.RAY_TIMEOUT_MS_TASK_WAIT_FOR_DEATH_INFO {
-				assert.Equal(t, "0", env.Value)
-				hasCorrectDeathEnv = true
-				break
-			}
-		}
-	}
-	assert.True(t, hasCorrectDeathEnv)
+	val, ok := pod.Labels[utils.RayClusterServingServiceLabelKey]
+	assert.True(t, ok, "Expected serve label is not present")
+	assert.Equal(t, utils.EnableRayClusterServingServiceFalse, val, "Wrong serve label value")
+	utils.EnvVarExists(utils.RAY_TIMEOUT_MS_TASK_WAIT_FOR_DEATH_INFO, pod.Spec.Containers[utils.RayContainerIndex].Env)
+
+	worker := cluster.Spec.WorkerGroupSpecs[0]
+	podName = cluster.Name + utils.DashSymbol + string(rayv1.WorkerNode) + utils.DashSymbol + worker.GroupName + utils.DashSymbol + utils.FormatInt32(0)
+	fqdnRayIP := utils.GenerateFQDNServiceName(ctx, *cluster, cluster.Namespace)
+	podTemplateSpec = DefaultWorkerPodTemplate(ctx, *cluster, worker, podName, fqdnRayIP, "6379")
+	pod = BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, string(utils.RayServiceCRD), fqdnRayIP)
+
+	val, ok = pod.Labels[utils.RayClusterServingServiceLabelKey]
+	assert.True(t, ok, "Expected serve label is not present")
+	assert.Equal(t, utils.EnableRayClusterServingServiceTrue, val, "Wrong serve label value")
+	utils.EnvVarExists(utils.RAY_TIMEOUT_MS_TASK_WAIT_FOR_DEATH_INFO, pod.Spec.Containers[utils.RayContainerIndex].Env)
 }
 
 func TestBuildPod_WithGcsFtEnabled(t *testing.T) {
@@ -543,7 +535,7 @@ func TestBuildPod_WithGcsFtEnabled(t *testing.T) {
 	// Build a head Pod.
 	podName := strings.ToLower(cluster.Name + utils.DashSymbol + string(rayv1.HeadNode) + utils.DashSymbol + utils.FormatInt32(0))
 	podTemplateSpec := DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "", false)
+	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "")
 
 	// Check environment variable "RAY_GCS_RPC_SERVER_RECONNECT_TIMEOUT_S"
 	rayContainer := pod.Spec.Containers[utils.RayContainerIndex]
@@ -561,7 +553,7 @@ func TestBuildPod_WithGcsFtEnabled(t *testing.T) {
 	cluster.Spec.HeadGroupSpec.Template.Spec.Containers[utils.RayContainerIndex].Env = append(cluster.Spec.HeadGroupSpec.Template.Spec.Containers[utils.RayContainerIndex].Env,
 		corev1.EnvVar{Name: utils.RAY_GCS_RPC_SERVER_RECONNECT_TIMEOUT_S, Value: "60"})
 	podTemplateSpec = DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	pod = BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "", false)
+	pod = BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "")
 	rayContainer = pod.Spec.Containers[utils.RayContainerIndex]
 
 	// Check environment variable "RAY_GCS_RPC_SERVER_RECONNECT_TIMEOUT_S"
@@ -578,7 +570,7 @@ func TestBuildPod_WithGcsFtEnabled(t *testing.T) {
 	podName = cluster.Name + utils.DashSymbol + string(rayv1.WorkerNode) + utils.DashSymbol + worker.GroupName + utils.DashSymbol + utils.FormatInt32(0)
 	fqdnRayIP := utils.GenerateFQDNServiceName(ctx, *cluster, cluster.Namespace)
 	podTemplateSpec = DefaultWorkerPodTemplate(ctx, *cluster, worker, podName, fqdnRayIP, "6379")
-	pod = BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP, false)
+	pod = BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP)
 
 	// Check the default value of "RAY_GCS_RPC_SERVER_RECONNECT_TIMEOUT_S"
 	rayContainer = pod.Spec.Containers[utils.RayContainerIndex]
@@ -595,7 +587,7 @@ func TestBuildPod_WithGcsFtEnabled(t *testing.T) {
 		corev1.EnvVar{Name: utils.RAY_GCS_RPC_SERVER_RECONNECT_TIMEOUT_S, Value: "120"})
 	worker = cluster.Spec.WorkerGroupSpecs[0]
 	podTemplateSpec = DefaultWorkerPodTemplate(ctx, *cluster, worker, podName, fqdnRayIP, "6379")
-	pod = BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP, false)
+	pod = BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, "6379", nil, "", fqdnRayIP)
 
 	// Check the default value of "RAY_GCS_RPC_SERVER_RECONNECT_TIMEOUT_S"
 	rayContainer = pod.Spec.Containers[utils.RayContainerIndex]
@@ -666,7 +658,7 @@ func TestBuildPodWithAutoscalerOptions(t *testing.T) {
 		SecurityContext:    &customSecurityContext,
 	}
 	podTemplateSpec := DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", &trueFlag, "", "", false)
+	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", &trueFlag, "", "")
 	expectedContainer := *autoscalerContainer.DeepCopy()
 	expectedContainer.Image = customAutoscalerImage
 	expectedContainer.ImagePullPolicy = customPullPolicy
@@ -840,7 +832,7 @@ func TestCleanupInvalidVolumeMounts(t *testing.T) {
 	// Test head pod
 	podName := strings.ToLower(cluster.Name + utils.DashSymbol + string(rayv1.HeadNode) + utils.DashSymbol + utils.FormatInt32(0))
 	podTemplateSpec := DefaultHeadPodTemplate(ctx, *cluster, cluster.Spec.HeadGroupSpec, podName, "6379")
-	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "", false)
+	pod := BuildPod(ctx, podTemplateSpec, rayv1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, "6379", nil, "", "")
 
 	pod.Spec.Containers[0].VolumeMounts = append(pod.Spec.Containers[0].VolumeMounts, []corev1.VolumeMount{
 		{
@@ -1211,16 +1203,18 @@ func TestInitLivenessAndReadinessProbe(t *testing.T) {
 
 	rayContainer.LivenessProbe = &httpGetProbe
 	rayContainer.ReadinessProbe = &httpGetProbe
-	initLivenessAndReadinessProbe(rayContainer, rayv1.HeadNode, false)
+	initLivenessAndReadinessProbe(rayContainer, rayv1.HeadNode, "")
 	assert.NotNil(t, rayContainer.LivenessProbe.HTTPGet)
 	assert.NotNil(t, rayContainer.ReadinessProbe.HTTPGet)
 	assert.Nil(t, rayContainer.LivenessProbe.Exec)
 	assert.Nil(t, rayContainer.ReadinessProbe.Exec)
 
 	// Test 2: User does not define a custom probe. KubeRay will inject Exec probe.
+	// Here we test the case where the Ray Pod originates from RayServiceCRD,
+	// implying that an additional serve health check will be added to the readiness probe.
 	rayContainer.LivenessProbe = nil
 	rayContainer.ReadinessProbe = nil
-	initLivenessAndReadinessProbe(rayContainer, rayv1.WorkerNode, true)
+	initLivenessAndReadinessProbe(rayContainer, rayv1.WorkerNode, utils.RayOriginatedFromCRDLabelValue(utils.RayServiceCRD))
 	assert.NotNil(t, rayContainer.LivenessProbe.Exec)
 	assert.NotNil(t, rayContainer.ReadinessProbe.Exec)
 	assert.False(t, strings.Contains(strings.Join(rayContainer.LivenessProbe.Exec.Command, " "), utils.RayServeProxyHealthPath))

diff --git a/ray-operator/controllers/ray/raycluster_controller.go b/ray-operator/controllers/ray/raycluster_controller.go
@@ -1040,19 +1040,14 @@ func (r *RayClusterReconciler) buildHeadPod(ctx context.Context, instance rayv1.
 	fqdnRayIP := utils.GenerateFQDNServiceName(ctx, instance, instance.Namespace) // Fully Qualified Domain Name
 	// The Ray head port used by workers to connect to the cluster (GCS server port for Ray >= 1.11.0, Redis port for older Ray.)
 	headPort := common.GetHeadPort(instance.Spec.HeadGroupSpec.RayStartParams)
-	// Check whether serve is enabled and add serve label
-	serveLabel := false
-	if enableServeServiceValue, exist := instance.Annotations[utils.EnableServeServiceKey]; exist && enableServeServiceValue == utils.EnableServeServiceTrue {
-		serveLabel = true
-	}
 	autoscalingEnabled := instance.Spec.EnableInTreeAutoscaling
 	podConf := common.DefaultHeadPodTemplate(ctx, instance, instance.Spec.HeadGroupSpec, podName, headPort)
 	if len(r.headSidecarContainers) > 0 {
 		podConf.Spec.Containers = append(podConf.Spec.Containers, r.headSidecarContainers...)
 	}
 	r.Log.Info("head pod labels", "labels", podConf.Labels)
 	creatorName := getCreator(instance)
-	pod := common.BuildPod(ctx, podConf, rayv1.HeadNode, instance.Spec.HeadGroupSpec.RayStartParams, headPort, autoscalingEnabled, creatorName, fqdnRayIP, serveLabel)
+	pod := common.BuildPod(ctx, podConf, rayv1.HeadNode, instance.Spec.HeadGroupSpec.RayStartParams, headPort, autoscalingEnabled, creatorName, fqdnRayIP)
 	// Set raycluster instance as the owner and controller
 	if err := controllerutil.SetControllerReference(&instance, &pod, r.Scheme); err != nil {
 		r.Log.Error(err, "Failed to set controller reference for raycluster pod")
@@ -1089,12 +1084,7 @@ func (r *RayClusterReconciler) buildWorkerPod(ctx context.Context, instance rayv
 		podTemplateSpec.Spec.Containers = append(podTemplateSpec.Spec.Containers, r.workerSidecarContainers...)
 	}
 	creatorName := getCreator(instance)
-	// Check whether serve is enabled and add serve label
-	serveLabel := false
-	if enableServeServiceValue, exist := instance.Annotations[utils.EnableServeServiceKey]; exist && enableServeServiceValue == utils.EnableServeServiceTrue {
-		serveLabel = true
-	}
-	pod := common.BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, headPort, autoscalingEnabled, creatorName, fqdnRayIP, serveLabel)
+	pod := common.BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, headPort, autoscalingEnabled, creatorName, fqdnRayIP)
 	// Set raycluster instance as the owner and controller
 	if err := controllerutil.SetControllerReference(&instance, &pod, r.Scheme); err != nil {
 		r.Log.Error(err, "Failed to set controller reference for raycluster pod")

diff --git a/ray-operator/controllers/ray/rayservice_controller.go b/ray-operator/controllers/ray/rayservice_controller.go
@@ -683,7 +683,6 @@ func (r *RayServiceReconciler) constructRayClusterForRayService(rayService *rayv
 	for k, v := range rayService.Annotations {
 		rayClusterAnnotations[k] = v
 	}
-	rayClusterAnnotations[utils.EnableServeServiceKey] = utils.EnableServeServiceTrue
 	errContext := "Failed to serialize RayCluster config. " +
 		"Manual config updates will NOT be tracked accurately. " +
 		"Please tear down the cluster and apply a new config."

diff --git a/ray-operator/controllers/ray/utils/constant.go b/ray-operator/controllers/ray/utils/constant.go
@@ -43,6 +43,8 @@ const (
 	// Finalizers for GCS fault tolerance
 	GCSFaultToleranceRedisCleanupFinalizer = "ray.io/gcs-ft-redis-cleanup-finalizer"
 
+	// EnableServeServiceKey is exclusively utilized to indicate if a RayCluster is directly used for serving.
+	// See https://github.com/ray-project/kuberay/pull/1672 for more details.
 	EnableServeServiceKey  = "ray.io/enable-serve-service"
 	EnableServeServiceTrue = "true"