[RFC][autoscaler] Add autoscaler container overrides and config optio…

…ns for scale behavior. (ray-project#278)

This PR adds an autoscalerOptions field to the RayCluster CRD.

ray-operator/apis/ray/v1alpha1/raycluster_types.go

@@ -20,6 +20,8 @@ type RayClusterSpec struct {
 	RayVersion string `json:"rayVersion,omitempty"`
 	// EnableInTreeAutoscaling indicates whether operator should create in tree autoscaling configs
 	EnableInTreeAutoscaling *bool `json:"enableInTreeAutoscaling,omitempty"`
+	// AutoscalerOptions specifies optional configuration for the Ray autoscaler.
+	AutoscalerOptions *AutoscalerOptions `json:"autoscalerOptions,omitempty"`
 }
 
 // HeadGroupSpec are the spec for the head pod
@@ -62,6 +64,25 @@ type ScaleStrategy struct {
 	WorkersToDelete []string `json:"workersToDelete,omitempty"`
 }
 
+// AutoscalerOptions specifies optional configuration for the Ray autoscaler.
+type AutoscalerOptions struct {
+	// Resources specifies resource requests and limits for the autoscaler container.
+	// Default values: 256m CPU request, 512m CPU limit, 256Mi memory request, 512Mi memory limit.
+	Resources *v1.ResourceRequirements `json:"resources,omitempty"`
+	// Image optionally overrides the autoscaler's container image. This override is for provided for autoscaler testing and development.
+	Image *string `json:"image,omitempty"`
+	// IdleTimeoutSeconds is the number of seconds to wait before scaling down a worker pod which is not using Ray resources.
+	// Defaults to 300 (five minutes).
+	IdleTimeoutSeconds *int32 `json:"idleTimeoutSeconds,omitempty"`
+	// UpscalineMode is "Default" or "Aggressive."
+	// Default: Upscaling is rate-limited; the number of pending worker pods is at most the size of the Ray cluster.
+	// Aggressive: Upscaling is not rate-limited.
+	UpscalingMode *UpscalingMode `json:"upscalingMode,omitempty"`
+}
+
+// +kubebuilder:validation:Enum=Default;Aggressive
+type UpscalingMode string
+
 // The overall state of the Ray cluster.
 type ClusterState string
 

ray-operator/config/crd/bases/ray.io_rayclusters.yaml

@@ -34,6 +34,51 @@ spec:
           spec:
             description: Specification of the desired behavior of the RayCluster.
             properties:
+              autoscalerOptions:
+                description: AutoscalerOptions specifies optional configuration for
+                  the Ray autoscaler.
+                properties:
+                  idleTimeoutSeconds:
+                    description: IdleTimeoutSeconds is the number of seconds to wait
+                      before scaling down a worker pod which is not us
+                    format: int32
+                    type: integer
+                  image:
+                    description: Image optionally overrides the autoscaler's container
+                      image.
+                    type: string
+                  resources:
+                    description: Resources specifies resource requests and limits
+                      for the autoscaler container.
+                    properties:
+                      limits:
+                        additionalProperties:
+                          anyOf:
+                          - type: integer
+                          - type: string
+                          pattern: ^(\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))))?$
+                          x-kubernetes-int-or-string: true
+                        description: 'Limits describes the maximum amount of compute
+                          resources allowed. More info: https://kubernetes.'
+                        type: object
+                      requests:
+                        additionalProperties:
+                          anyOf:
+                          - type: integer
+                          - type: string
+                          pattern: ^(\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))))?$
+                          x-kubernetes-int-or-string: true
+                        description: Requests describes the minimum amount of compute
+                          resources required.
+                        type: object
+                    type: object
+                  upscalingMode:
+                    description: UpscalineMode is "Default" or "Aggressive.
+                    enum:
+                    - Default
+                    - Aggressive
+                    type: string
+                type: object
               enableInTreeAutoscaling:
                 description: EnableInTreeAutoscaling indicates whether operator should
                   create in tree autoscaling configs

ray-operator/controllers/ray/common/pod.go

@@ -49,8 +49,10 @@ func DefaultHeadPodTemplate(instance rayiov1alpha1.RayCluster, headSpec rayiov1a
 		podTemplate.Spec.ServiceAccountName = utils.GetHeadGroupServiceAccountName(&instance)
 
 		// inject autoscaler container into head pod
-		container := BuildAutoscalerContainer()
-		podTemplate.Spec.Containers = append(podTemplate.Spec.Containers, container)
+		autoscalerContainer := BuildAutoscalerContainer()
+		// Merge the user overrides from autoscalerOptions into the autoscaler container config.
+		mergeAutoscalerOverrides(&autoscalerContainer, instance.Spec.AutoscalerOptions)
+		podTemplate.Spec.Containers = append(podTemplate.Spec.Containers, autoscalerContainer)
 		// set custom service account which can be authorized to talk with apiserver
 		podTemplate.Spec.ServiceAccountName = instance.Name
 	}
@@ -214,6 +216,18 @@ func BuildAutoscalerContainer() v1.Container {
 	return container
 }
 
+// Merge the user overrides from autoscalerOptions into the autoscaler container config.
+func mergeAutoscalerOverrides(autoscalerContainer *v1.Container, autoscalerOptions *rayiov1alpha1.AutoscalerOptions) {
+	if autoscalerOptions != nil {
+		if autoscalerOptions.Resources != nil {
+			autoscalerContainer.Resources = *autoscalerOptions.Resources
+		}
+		if autoscalerOptions.Image != nil {
+			autoscalerContainer.Image = *autoscalerOptions.Image
+		}
+	}
+}
+
 func isRayStartWithBlock(rayStartParams map[string]string) bool {
 	if blockValue, exist := rayStartParams["block"]; exist {
 		return strings.ToLower(blockValue) == "true"

ray-operator/controllers/ray/common/pod_test.go

@@ -334,7 +334,45 @@ func TestBuildPod_WithAutoscalerEnabled(t *testing.T) {
 	if !reflect.DeepEqual(expectedContainer, actualContainer) {
 		t.Fatalf("Expected `%v` but got `%v`", expectedContainer, actualContainer)
 	}
+}
+
+// Check that autoscaler container overrides work as expected.
+func TestBuildPodWithAutoscalerOptions(t *testing.T) {
+	cluster := instance.DeepCopy()
+	cluster.Spec.EnableInTreeAutoscaling = &trueFlag
+	podName := strings.ToLower(cluster.Name + DashSymbol + string(rayiov1alpha1.HeadNode) + DashSymbol + utils.FormatInt32(0))
+	svcName := utils.GenerateServiceName(cluster.Name)
+
+	customAutoscalerImage := "custom-autoscaler-xxx"
+	customTimeout := int32(100)
+	customUpscaling := "Aggressive"
+	customResources := v1.ResourceRequirements{
+		Requests: v1.ResourceList{
+			v1.ResourceCPU:    resource.MustParse("1"),
+			v1.ResourceMemory: testMemoryLimit,
+		},
+		Limits: v1.ResourceList{
+			v1.ResourceCPU:    resource.MustParse("1"),
+			v1.ResourceMemory: testMemoryLimit,
+		},
+	}
 
+	cluster.Spec.AutoscalerOptions = &rayiov1alpha1.AutoscalerOptions{
+		UpscalingMode:      (*rayiov1alpha1.UpscalingMode)(&customUpscaling),
+		IdleTimeoutSeconds: &customTimeout,
+		Image:              &customAutoscalerImage,
+		Resources:          &customResources,
+	}
+	podTemplateSpec := DefaultHeadPodTemplate(*cluster, cluster.Spec.HeadGroupSpec, podName, svcName)
+	pod := BuildPod(podTemplateSpec, rayiov1alpha1.HeadNode, cluster.Spec.HeadGroupSpec.RayStartParams, svcName, &trueFlag)
+	expectedContainer := *autoscalerContainer.DeepCopy()
+	expectedContainer.Image = customAutoscalerImage
+	expectedContainer.Resources = customResources
+	index := getAutoscalerContainerIndex(pod)
+	actualContainer := pod.Spec.Containers[index]
+	if !reflect.DeepEqual(expectedContainer, actualContainer) {
+		t.Fatalf("Expected `%v` but got `%v`", expectedContainer, actualContainer)
+	}
 }
 
 func TestDefaultHeadPodTemplate_WithAutoscalingEnabled(t *testing.T) {