Files
kuberay/ray-operator/controllers/ray/raycluster_controller.go

1509 lines
63 KiB
Go

package ray
import (
"context"
errstd "errors"
"fmt"
"os"
"reflect"
"runtime"
"strconv"
"strings"
"time"
"k8s.io/apimachinery/pkg/api/meta"
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
"k8s.io/utils/ptr"
"github.com/ray-project/kuberay/ray-operator/controllers/ray/batchscheduler"
"github.com/ray-project/kuberay/ray-operator/controllers/ray/common"
"github.com/ray-project/kuberay/ray-operator/controllers/ray/utils"
"github.com/ray-project/kuberay/ray-operator/pkg/features"
batchv1 "k8s.io/api/batch/v1"
rbacv1 "k8s.io/api/rbac/v1"
"k8s.io/client-go/tools/record"
rayv1 "github.com/ray-project/kuberay/ray-operator/apis/ray/v1"
"github.com/go-logr/logr"
routev1 "github.com/openshift/api/route/v1"
"k8s.io/client-go/discovery"
"k8s.io/client-go/rest"
corev1 "k8s.io/api/core/v1"
networkingv1 "k8s.io/api/networking/v1"
"k8s.io/apimachinery/pkg/api/errors"
"k8s.io/apimachinery/pkg/api/resource"
k8sruntime "k8s.io/apimachinery/pkg/runtime"
ctrl "sigs.k8s.io/controller-runtime"
"sigs.k8s.io/controller-runtime/pkg/builder"
"sigs.k8s.io/controller-runtime/pkg/client"
controller "sigs.k8s.io/controller-runtime/pkg/controller"
"sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
"sigs.k8s.io/controller-runtime/pkg/manager"
"sigs.k8s.io/controller-runtime/pkg/predicate"
"sigs.k8s.io/controller-runtime/pkg/reconcile"
)
type reconcileFunc func(context.Context, *rayv1.RayCluster) error
var (
DefaultRequeueDuration = 2 * time.Second
EnableBatchScheduler bool
// Definition of a index field for pod name
podUIDIndexField = "metadata.uid"
)
// getDiscoveryClient returns a discovery client for the current reconciler
func getDiscoveryClient(config *rest.Config) (*discovery.DiscoveryClient, error) {
return discovery.NewDiscoveryClientForConfig(config)
}
// Check where we are running. We are trying to distinguish here whether
// this is vanilla kubernetes cluster or Openshift
func getClusterType(ctx context.Context) bool {
logger := ctrl.LoggerFrom(ctx)
if os.Getenv("USE_INGRESS_ON_OPENSHIFT") == "true" {
// Environment is set to treat OpenShift cluster as Vanilla Kubernetes
return false
}
// The discovery package is used to discover APIs supported by a Kubernetes API server.
config, err := ctrl.GetConfig()
if err == nil && config != nil {
dclient, err := getDiscoveryClient(config)
if err == nil && dclient != nil {
apiGroupList, err := dclient.ServerGroups()
if err != nil {
logger.Info("Error while querying ServerGroups, assuming we're on Vanilla Kubernetes")
return false
}
for i := 0; i < len(apiGroupList.Groups); i++ {
if strings.HasSuffix(apiGroupList.Groups[i].Name, ".openshift.io") {
logger.Info("We detected being on OpenShift!")
return true
}
}
return false
}
logger.Info("Cannot retrieve a DiscoveryClient, assuming we're on Vanilla Kubernetes")
return false
}
logger.Info("Cannot retrieve config, assuming we're on Vanilla Kubernetes")
return false
}
// NewReconciler returns a new reconcile.Reconciler
func NewReconciler(ctx context.Context, mgr manager.Manager, options RayClusterReconcilerOptions) *RayClusterReconciler {
if err := mgr.GetFieldIndexer().IndexField(ctx, &corev1.Pod{}, podUIDIndexField, func(rawObj client.Object) []string {
pod := rawObj.(*corev1.Pod)
return []string{string(pod.UID)}
}); err != nil {
panic(err)
}
isOpenShift := getClusterType(ctx)
return &RayClusterReconciler{
Client: mgr.GetClient(),
Scheme: mgr.GetScheme(),
Recorder: mgr.GetEventRecorderFor("raycluster-controller"),
BatchSchedulerMgr: batchscheduler.NewSchedulerManager(mgr.GetConfig()),
IsOpenShift: isOpenShift,
headSidecarContainers: options.HeadSidecarContainers,
workerSidecarContainers: options.WorkerSidecarContainers,
}
}
var _ reconcile.Reconciler = &RayClusterReconciler{}
// RayClusterReconciler reconciles a RayCluster object
type RayClusterReconciler struct {
client.Client
Scheme *k8sruntime.Scheme
Recorder record.EventRecorder
BatchSchedulerMgr *batchscheduler.SchedulerManager
headSidecarContainers []corev1.Container
workerSidecarContainers []corev1.Container
IsOpenShift bool
}
type RayClusterReconcilerOptions struct {
HeadSidecarContainers []corev1.Container
WorkerSidecarContainers []corev1.Container
}
// Reconcile reads that state of the cluster for a RayCluster object and makes changes based on it
// and what is in the RayCluster.Spec
// Automatically generate RBAC rules to allow the Controller to read and write workloads
// +kubebuilder:rbac:groups=ray.io,resources=rayclusters,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=ray.io,resources=rayclusters/status,verbs=get;update;patch
// +kubebuilder:rbac:groups=ray.io,resources=rayclusters/finalizers,verbs=update
// +kubebuilder:rbac:groups=core,resources=events,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=core,resources=pods,verbs=get;list;watch;create;update;patch;delete;deletecollection
// +kubebuilder:rbac:groups=core,resources=pods/status,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=core,resources=services,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=core,resources=services/status,verbs=get;update;patch
// +kubebuilder:rbac:groups=coordination.k8s.io,resources=leases,verbs=get;list;create;update
// +kubebuilder:rbac:groups=networking.k8s.io,resources=ingressclasses,verbs=get;list;watch
// +kubebuilder:rbac:groups=networking.k8s.io,resources=ingresses,verbs=get;list;watch;create;update;delete;patch
// +kubebuilder:rbac:groups=route.openshift.io,resources=routes,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=extensions,resources=ingresses,verbs=get;list;watch;create;update;delete;patch
// +kubebuilder:rbac:groups=core,resources=serviceaccounts,verbs=get;list;watch;create;delete
// +kubebuilder:rbac:groups="rbac.authorization.k8s.io",resources=roles,verbs=get;list;watch;create;delete;update
// +kubebuilder:rbac:groups="rbac.authorization.k8s.io",resources=rolebindings,verbs=get;list;watch;create;delete
// [WARNING]: There MUST be a newline after kubebuilder markers.
// Reconcile used to bridge the desired state with the current state
func (r *RayClusterReconciler) Reconcile(ctx context.Context, request ctrl.Request) (ctrl.Result, error) {
logger := ctrl.LoggerFrom(ctx)
var err error
// Try to fetch the RayCluster instance
instance := &rayv1.RayCluster{}
if err = r.Get(ctx, request.NamespacedName, instance); err == nil {
return r.rayClusterReconcile(ctx, request, instance)
}
// No match found
if errors.IsNotFound(err) {
logger.Info("Read request instance not found error!")
} else {
logger.Error(err, "Read request instance error!")
}
// Error reading the object - requeue the request.
return ctrl.Result{}, client.IgnoreNotFound(err)
}
func (r *RayClusterReconciler) deleteAllPods(ctx context.Context, filters common.AssociationOptions) (pods corev1.PodList, err error) {
logger := ctrl.LoggerFrom(ctx)
if err = r.List(ctx, &pods, filters.ToListOptions()...); err != nil {
return pods, err
}
active := 0
for _, pod := range pods.Items {
if pod.DeletionTimestamp.IsZero() {
active++
}
}
if active > 0 {
logger.Info("Deleting all Pods with labels", "filters", filters, "Number of active Pods", active)
return pods, r.DeleteAllOf(ctx, &corev1.Pod{}, filters.ToDeleteOptions()...)
}
return pods, nil
}
func (r *RayClusterReconciler) rayClusterReconcile(ctx context.Context, request ctrl.Request, instance *rayv1.RayCluster) (ctrl.Result, error) {
var reconcileErr error
logger := ctrl.LoggerFrom(ctx)
// Please do NOT modify `originalRayClusterInstance` in the following code.
originalRayClusterInstance := instance.DeepCopy()
// The `enableGCSFTRedisCleanup` is a feature flag introduced in KubeRay v1.0.0. It determines whether
// the Redis cleanup job should be activated. Users can disable the feature by setting the environment
// variable `ENABLE_GCS_FT_REDIS_CLEANUP` to `false`, and undertake the Redis storage namespace cleanup
// manually after the RayCluster CR deletion.
enableGCSFTRedisCleanup := strings.ToLower(os.Getenv(utils.ENABLE_GCS_FT_REDIS_CLEANUP)) != "false"
if enableGCSFTRedisCleanup && common.IsGCSFaultToleranceEnabled(*instance) {
if instance.DeletionTimestamp.IsZero() {
if !controllerutil.ContainsFinalizer(instance, utils.GCSFaultToleranceRedisCleanupFinalizer) {
logger.Info(
"GCS fault tolerance has been enabled. Implementing a finalizer to ensure that Redis is properly cleaned up once the RayCluster custom resource (CR) is deleted.",
"finalizer", utils.GCSFaultToleranceRedisCleanupFinalizer)
controllerutil.AddFinalizer(instance, utils.GCSFaultToleranceRedisCleanupFinalizer)
if err := r.Update(ctx, instance); err != nil {
err = fmt.Errorf("Failed to add the finalizer %s to the RayCluster: %w", utils.GCSFaultToleranceRedisCleanupFinalizer, err)
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, err
}
// Only start the RayCluster reconciliation after the finalizer is added.
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, nil
}
} else {
logger.Info(
fmt.Sprintf("The RayCluster with GCS enabled, %s, is being deleted. Start to handle the Redis cleanup finalizer %s.",
instance.Name, utils.GCSFaultToleranceRedisCleanupFinalizer),
"DeletionTimestamp", instance.ObjectMeta.DeletionTimestamp)
// Delete the head Pod if it exists.
headPods, err := r.deleteAllPods(ctx, common.RayClusterHeadPodsAssociationOptions(instance))
if err != nil {
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, err
}
// Delete all worker Pods if they exist.
if _, err = r.deleteAllPods(ctx, common.RayClusterWorkerPodsAssociationOptions(instance)); err != nil {
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, err
}
if len(headPods.Items) > 0 {
logger.Info(fmt.Sprintf(
"Wait for the head Pod %s to be terminated before initiating the Redis cleanup process. "+
"The storage namespace %s in Redis cannot be fully deleted if the GCS process on the head Pod is still writing to it.",
headPods.Items[0].Name, headPods.Items[0].Annotations[utils.RayExternalStorageNSAnnotationKey]))
// Requeue after 10 seconds because it takes much longer than DefaultRequeueDuration (2 seconds) for the head Pod to be terminated.
return ctrl.Result{RequeueAfter: 10 * time.Second}, nil
}
// We can start the Redis cleanup process now because the head Pod has been terminated.
filterLabels := client.MatchingLabels{utils.RayClusterLabelKey: instance.Name, utils.RayNodeTypeLabelKey: string(rayv1.RedisCleanupNode)}
redisCleanupJobs := batchv1.JobList{}
if err := r.List(ctx, &redisCleanupJobs, client.InNamespace(instance.Namespace), filterLabels); err != nil {
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, err
}
if len(redisCleanupJobs.Items) != 0 {
// Check whether the Redis cleanup Job has been completed.
redisCleanupJob := redisCleanupJobs.Items[0]
logger.Info("Redis cleanup Job status", "Job name", redisCleanupJob.Name,
"Active", redisCleanupJob.Status.Active, "Succeeded", redisCleanupJob.Status.Succeeded, "Failed", redisCleanupJob.Status.Failed)
if condition, finished := utils.IsJobFinished(&redisCleanupJob); finished {
controllerutil.RemoveFinalizer(instance, utils.GCSFaultToleranceRedisCleanupFinalizer)
if err := r.Update(ctx, instance); err != nil {
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, err
}
switch condition {
case batchv1.JobComplete:
logger.Info(fmt.Sprintf(
"The Redis cleanup Job %s has been completed. "+
"The storage namespace %s in Redis has been fully deleted.",
redisCleanupJob.Name, redisCleanupJob.Annotations[utils.RayExternalStorageNSAnnotationKey]))
case batchv1.JobFailed:
logger.Info(fmt.Sprintf(
"The Redis cleanup Job %s has failed, requeue the RayCluster CR after 5 minute. "+
"You should manually delete the storage namespace %s in Redis and remove the RayCluster's finalizer. "+
"Please check https://docs.ray.io/en/master/cluster/kubernetes/user-guides/kuberay-gcs-ft.html for more details.",
redisCleanupJob.Name, redisCleanupJob.Annotations[utils.RayExternalStorageNSAnnotationKey]))
}
return ctrl.Result{}, nil
}
// the redisCleanupJob is still running
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, nil
}
redisCleanupJob := r.buildRedisCleanupJob(ctx, *instance)
if err := r.Create(ctx, &redisCleanupJob); err != nil {
if errors.IsAlreadyExists(err) {
logger.Info(fmt.Sprintf("Redis cleanup Job already exists. Requeue the RayCluster CR %s.", instance.Name))
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, nil
}
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, err
}
logger.Info("Successfully created Redis cleanup Job", "Job name", redisCleanupJob.Name)
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, nil
}
}
if instance.DeletionTimestamp != nil && !instance.DeletionTimestamp.IsZero() {
logger.Info("RayCluster is being deleted, just ignore", "cluster name", request.Name)
return ctrl.Result{}, nil
}
reconcileFuncs := []reconcileFunc{
r.reconcileAutoscalerServiceAccount,
r.reconcileAutoscalerRole,
r.reconcileAutoscalerRoleBinding,
r.reconcileIngress,
r.reconcileHeadService,
r.reconcileHeadlessService,
r.reconcileServeService,
r.reconcilePods,
}
for _, fn := range reconcileFuncs {
if reconcileErr = fn(ctx, instance); reconcileErr != nil {
funcName := runtime.FuncForPC(reflect.ValueOf(fn).Pointer()).Name()
logger.Error(reconcileErr, "Error reconcile resources", "function name", funcName)
break
}
}
// Calculate the new status for the RayCluster. Note that the function will deep copy `instance` instead of mutating it.
newInstance, calculateErr := r.calculateStatus(ctx, instance, reconcileErr)
var updateErr error
if calculateErr != nil {
logger.Info("Got error when calculating new status", "cluster name", request.Name, "error", calculateErr)
} else {
updateErr = r.updateRayClusterStatus(ctx, originalRayClusterInstance, newInstance)
}
// Return error based on order.
var err error
if reconcileErr != nil {
err = reconcileErr
} else if calculateErr != nil {
err = calculateErr
} else {
err = updateErr
}
if err != nil {
return ctrl.Result{RequeueAfter: DefaultRequeueDuration}, err
}
// Unconditionally requeue after the number of seconds specified in the
// environment variable RAYCLUSTER_DEFAULT_REQUEUE_SECONDS_ENV. If the
// environment variable is not set, requeue after the default value.
requeueAfterSeconds, err := strconv.Atoi(os.Getenv(utils.RAYCLUSTER_DEFAULT_REQUEUE_SECONDS_ENV))
if err != nil {
logger.Info(fmt.Sprintf("Environment variable %s is not set, using default value of %d seconds", utils.RAYCLUSTER_DEFAULT_REQUEUE_SECONDS_ENV, utils.RAYCLUSTER_DEFAULT_REQUEUE_SECONDS), "cluster name", request.Name)
requeueAfterSeconds = utils.RAYCLUSTER_DEFAULT_REQUEUE_SECONDS
}
logger.Info("Unconditional requeue after", "cluster name", request.Name, "seconds", requeueAfterSeconds)
return ctrl.Result{RequeueAfter: time.Duration(requeueAfterSeconds) * time.Second}, nil
}
// Checks whether the old and new RayClusterStatus are inconsistent by comparing different fields. If the only
// differences between the old and new status are the `LastUpdateTime` and `ObservedGeneration` fields, the
// status update will not be triggered.
//
// TODO (kevin85421): The field `ObservedGeneration` is not being well-maintained at the moment. In the future,
// this field should be used to determine whether to update this CR or not.
func (r *RayClusterReconciler) inconsistentRayClusterStatus(ctx context.Context, oldStatus rayv1.RayClusterStatus, newStatus rayv1.RayClusterStatus) bool {
logger := ctrl.LoggerFrom(ctx)
if oldStatus.State != newStatus.State || oldStatus.Reason != newStatus.Reason {
logger.Info("inconsistentRayClusterStatus", "detect inconsistency", fmt.Sprintf(
"old State: %s, new State: %s, old Reason: %s, new Reason: %s",
oldStatus.State, newStatus.State, oldStatus.Reason, newStatus.Reason))
return true
}
if oldStatus.ReadyWorkerReplicas != newStatus.ReadyWorkerReplicas ||
oldStatus.AvailableWorkerReplicas != newStatus.AvailableWorkerReplicas ||
oldStatus.DesiredWorkerReplicas != newStatus.DesiredWorkerReplicas ||
oldStatus.MinWorkerReplicas != newStatus.MinWorkerReplicas ||
oldStatus.MaxWorkerReplicas != newStatus.MaxWorkerReplicas {
logger.Info("inconsistentRayClusterStatus", "detect inconsistency", fmt.Sprintf(
"old ReadyWorkerReplicas: %d, new ReadyWorkerReplicas: %d, "+
"old AvailableWorkerReplicas: %d, new AvailableWorkerReplicas: %d, "+
"old DesiredWorkerReplicas: %d, new DesiredWorkerReplicas: %d, "+
"old MinWorkerReplicas: %d, new MinWorkerReplicas: %d, "+
"old MaxWorkerReplicas: %d, new MaxWorkerReplicas: %d",
oldStatus.ReadyWorkerReplicas, newStatus.ReadyWorkerReplicas,
oldStatus.AvailableWorkerReplicas, newStatus.AvailableWorkerReplicas,
oldStatus.DesiredWorkerReplicas, newStatus.DesiredWorkerReplicas,
oldStatus.MinWorkerReplicas, newStatus.MinWorkerReplicas,
oldStatus.MaxWorkerReplicas, newStatus.MaxWorkerReplicas))
return true
}
if !reflect.DeepEqual(oldStatus.Endpoints, newStatus.Endpoints) || !reflect.DeepEqual(oldStatus.Head, newStatus.Head) {
logger.Info("inconsistentRayClusterStatus", "detect inconsistency", fmt.Sprintf(
"old Endpoints: %v, new Endpoints: %v, old Head: %v, new Head: %v",
oldStatus.Endpoints, newStatus.Endpoints, oldStatus.Head, newStatus.Head))
return true
}
if !reflect.DeepEqual(oldStatus.Conditions, newStatus.Conditions) {
logger.Info("inconsistentRayClusterStatus", "old conditions", oldStatus.Conditions, "new conditions", newStatus.Conditions)
return true
}
return false
}
func (r *RayClusterReconciler) reconcileIngress(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
logger.Info("Reconciling Ingress")
if instance.Spec.HeadGroupSpec.EnableIngress == nil || !*instance.Spec.HeadGroupSpec.EnableIngress {
return nil
}
if r.IsOpenShift {
// This is open shift - create route
return r.reconcileRouteOpenShift(ctx, instance)
}
// plain vanilla kubernetes - create ingress
return r.reconcileIngressKubernetes(ctx, instance)
}
func (r *RayClusterReconciler) reconcileRouteOpenShift(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
headRoutes := routev1.RouteList{}
filterLabels := client.MatchingLabels{utils.RayClusterLabelKey: instance.Name}
if err := r.List(ctx, &headRoutes, client.InNamespace(instance.Namespace), filterLabels); err != nil {
return err
}
if headRoutes.Items != nil && len(headRoutes.Items) == 1 {
logger.Info("reconcileIngresses", "head service route found", headRoutes.Items[0].Name)
return nil
}
if headRoutes.Items == nil || len(headRoutes.Items) == 0 {
route, err := common.BuildRouteForHeadService(*instance)
if err != nil {
return err
}
if err := ctrl.SetControllerReference(instance, route, r.Scheme); err != nil {
return err
}
err = r.createHeadRoute(ctx, route, instance)
if err != nil {
return err
}
}
return nil
}
func (r *RayClusterReconciler) reconcileIngressKubernetes(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
headIngresses := networkingv1.IngressList{}
filterLabels := client.MatchingLabels{utils.RayClusterLabelKey: instance.Name}
if err := r.List(ctx, &headIngresses, client.InNamespace(instance.Namespace), filterLabels); err != nil {
return err
}
if headIngresses.Items != nil && len(headIngresses.Items) == 1 {
logger.Info("reconcileIngresses", "head service ingress found", headIngresses.Items[0].Name)
return nil
}
if headIngresses.Items == nil || len(headIngresses.Items) == 0 {
ingress, err := common.BuildIngressForHeadService(ctx, *instance)
if err != nil {
return err
}
if err := ctrl.SetControllerReference(instance, ingress, r.Scheme); err != nil {
return err
}
err = r.createHeadIngress(ctx, ingress, instance)
if err != nil {
return err
}
}
return nil
}
// Return nil only when the head service successfully created or already exists.
func (r *RayClusterReconciler) reconcileHeadService(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
services := corev1.ServiceList{}
filterLabels := client.MatchingLabels{utils.RayClusterLabelKey: instance.Name, utils.RayNodeTypeLabelKey: string(rayv1.HeadNode)}
if err := r.List(ctx, &services, client.InNamespace(instance.Namespace), filterLabels); err != nil {
return err
}
// Check if there's existing head service in the cluster.
if len(services.Items) != 0 {
if len(services.Items) == 1 {
logger.Info("reconcileHeadService", "1 head service found", services.Items[0].Name)
return nil
}
// This should never happen. This protects against the case that users manually create service with the same label.
if len(services.Items) > 1 {
logger.Info("reconcileHeadService", "Duplicate head service found", services.Items)
return fmt.Errorf("%d head service found %v", len(services.Items), services.Items)
}
} else {
// Create head service if there's no existing one in the cluster.
labels := make(map[string]string)
if val, ok := instance.Spec.HeadGroupSpec.Template.ObjectMeta.Labels[utils.KubernetesApplicationNameLabelKey]; ok {
labels[utils.KubernetesApplicationNameLabelKey] = val
}
annotations := make(map[string]string)
// TODO (kevin85421): KubeRay has already exposed the entire head service (#1040) to users.
// We may consider deprecating this field when we bump the CRD version.
for k, v := range instance.Spec.HeadServiceAnnotations {
annotations[k] = v
}
headSvc, err := common.BuildServiceForHeadPod(ctx, *instance, labels, annotations)
// TODO (kevin85421): Provide a detailed and actionable error message. For example, which port is missing?
if len(headSvc.Spec.Ports) == 0 {
logger.Info("Ray head service does not have any ports set up. Service specification: %v", headSvc.Spec)
return fmt.Errorf("Ray head service does not have any ports set up. Service specification: %v", headSvc.Spec)
}
if err != nil {
return err
}
if err := r.createService(ctx, headSvc, instance); err != nil {
return err
}
}
return nil
}
// Return nil only when the serve service successfully created or already exists.
func (r *RayClusterReconciler) reconcileServeService(ctx context.Context, instance *rayv1.RayCluster) error {
// Only reconcile the K8s service for Ray Serve when the "ray.io/enable-serve-service" annotation is set to true.
if enableServeServiceValue, exist := instance.Annotations[utils.EnableServeServiceKey]; !exist || enableServeServiceValue != utils.EnableServeServiceTrue {
return nil
}
// Retrieve the Service from the Kubernetes cluster with the name and namespace.
svc := &corev1.Service{}
err := r.Get(ctx, common.RayClusterServeServiceNamespacedName(instance), svc)
if err == nil {
// service exists, do nothing
return nil
} else if errors.IsNotFound(err) {
// Service does not exist, create it
svc, err = common.BuildServeServiceForRayCluster(ctx, *instance)
if err != nil {
return err
}
// Set the ownwer reference
if err := ctrl.SetControllerReference(instance, svc, r.Scheme); err != nil {
return err
}
// create service
return r.Create(ctx, svc)
}
return err
}
// Return nil only when the headless service for multi-host worker groups is successfully created or already exists.
func (r *RayClusterReconciler) reconcileHeadlessService(ctx context.Context, instance *rayv1.RayCluster) error {
// Check if there are worker groups with NumOfHosts > 1 in the cluster
isMultiHost := false
for _, workerGroup := range instance.Spec.WorkerGroupSpecs {
if workerGroup.NumOfHosts > 1 {
isMultiHost = true
break
}
}
if isMultiHost {
services := corev1.ServiceList{}
options := common.RayClusterHeadlessServiceListOptions(instance)
if err := r.List(ctx, &services, options...); err != nil {
return err
}
// Check if there's an existing headless service in the cluster.
if len(services.Items) != 0 {
// service exists, do nothing
return nil
}
// Create headless tpu worker service if there's no existing one in the cluster.
headlessSvc, err := common.BuildHeadlessServiceForRayCluster(*instance)
if err != nil {
return err
}
if err := r.createService(ctx, headlessSvc, instance); err != nil {
return err
}
}
return nil
}
func (r *RayClusterReconciler) reconcilePods(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
// if RayCluster is suspended, delete all pods and skip reconcile
if instance.Spec.Suspend != nil && *instance.Spec.Suspend {
if _, err := r.deleteAllPods(ctx, common.RayClusterAllPodsAssociationOptions(instance)); err != nil {
return errstd.Join(utils.ErrFailedDeleteAllPods, err)
}
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Deleted",
"Deleted Pods for RayCluster %s/%s due to suspension",
instance.Namespace, instance.Name)
return nil
}
// check if all the pods exist
headPods := corev1.PodList{}
if err := r.List(ctx, &headPods, common.RayClusterHeadPodsAssociationOptions(instance).ToListOptions()...); err != nil {
return err
}
if EnableBatchScheduler {
if scheduler, err := r.BatchSchedulerMgr.GetSchedulerForCluster(instance); err == nil {
if err := scheduler.DoBatchSchedulingOnSubmission(ctx, instance); err != nil {
return err
}
} else {
return err
}
}
// Reconcile head Pod
if len(headPods.Items) == 1 {
headPod := headPods.Items[0]
logger.Info("reconcilePods", "Found 1 head Pod", headPod.Name, "Pod status", headPod.Status.Phase,
"Pod status reason", headPod.Status.Reason,
"Pod restart policy", headPod.Spec.RestartPolicy,
"Ray container terminated status", getRayContainerStateTerminated(headPod))
shouldDelete, reason := shouldDeletePod(headPod, rayv1.HeadNode)
logger.Info("reconcilePods", "head Pod", headPod.Name, "shouldDelete", shouldDelete, "reason", reason)
if shouldDelete {
if err := r.Delete(ctx, &headPod); err != nil {
return errstd.Join(utils.ErrFailedDeleteHeadPod, err)
}
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Deleted",
"Deleted head Pod %s; Pod status: %s; Pod restart policy: %s; Ray container terminated status: %v",
headPod.Name, headPod.Status.Phase, headPod.Spec.RestartPolicy, getRayContainerStateTerminated(headPod))
return fmt.Errorf(reason)
}
} else if len(headPods.Items) == 0 {
// Create head Pod if it does not exist.
logger.Info("reconcilePods", "Found 0 head Pods; creating a head Pod for the RayCluster.", instance.Name)
common.CreatedClustersCounterInc(instance.Namespace)
if err := r.createHeadPod(ctx, *instance); err != nil {
common.FailedClustersCounterInc(instance.Namespace)
return errstd.Join(utils.ErrFailedCreateHeadPod, err)
}
common.SuccessfulClustersCounterInc(instance.Namespace)
} else if len(headPods.Items) > 1 {
logger.Info("reconcilePods", fmt.Sprintf("Found %d head Pods; deleting extra head Pods.", len(headPods.Items)), instance.Name)
// TODO (kevin85421): In-place update may not be a good idea.
itemLength := len(headPods.Items)
for index := 0; index < itemLength; index++ {
if headPods.Items[index].Status.Phase == corev1.PodRunning || headPods.Items[index].Status.Phase == corev1.PodPending {
// Remove the healthy pod at index i from the list of pods to delete
headPods.Items[index] = headPods.Items[len(headPods.Items)-1] // replace last element with the healthy head.
headPods.Items = headPods.Items[:len(headPods.Items)-1] // Truncate slice.
itemLength--
}
}
// delete all the extra head pod pods
for _, extraHeadPodToDelete := range headPods.Items {
if err := r.Delete(ctx, &extraHeadPodToDelete); err != nil {
return errstd.Join(utils.ErrFailedDeleteHeadPod, err)
}
}
}
// Reconcile worker pods now
for _, worker := range instance.Spec.WorkerGroupSpecs {
// workerReplicas will store the target number of pods for this worker group.
var workerReplicas int32 = utils.GetWorkerGroupDesiredReplicas(ctx, worker)
logger.Info("reconcilePods", "desired workerReplicas (always adhering to minReplicas/maxReplica)", workerReplicas, "worker group", worker.GroupName, "maxReplicas", worker.MaxReplicas, "minReplicas", worker.MinReplicas, "replicas", worker.Replicas)
workerPods := corev1.PodList{}
if err := r.List(ctx, &workerPods, common.RayClusterGroupPodsAssociationOptions(instance, worker.GroupName).ToListOptions()...); err != nil {
return err
}
// Delete unhealthy worker Pods.
deletedWorkers := make(map[string]struct{})
deleted := struct{}{}
numDeletedUnhealthyWorkerPods := 0
for _, workerPod := range workerPods.Items {
shouldDelete, reason := shouldDeletePod(workerPod, rayv1.WorkerNode)
logger.Info("reconcilePods", "worker Pod", workerPod.Name, "shouldDelete", shouldDelete, "reason", reason)
if shouldDelete {
numDeletedUnhealthyWorkerPods++
deletedWorkers[workerPod.Name] = deleted
if err := r.Delete(ctx, &workerPod); err != nil {
return errstd.Join(utils.ErrFailedDeleteWorkerPod, err)
}
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Deleted",
"Deleted worker Pod %s; Pod status: %s; Pod restart policy: %s; Ray container terminated status: %v",
workerPod.Name, workerPod.Status.Phase, workerPod.Spec.RestartPolicy, getRayContainerStateTerminated(workerPod))
}
}
// If we delete unhealthy Pods, we will not create new Pods in this reconciliation.
if numDeletedUnhealthyWorkerPods > 0 {
return fmt.Errorf("Delete %d unhealthy worker Pods", numDeletedUnhealthyWorkerPods)
}
// Always remove the specified WorkersToDelete - regardless of the value of Replicas.
// Essentially WorkersToDelete has to be deleted to meet the expectations of the Autoscaler.
logger.Info("reconcilePods", "removing the pods in the scaleStrategy of", worker.GroupName)
for _, podsToDelete := range worker.ScaleStrategy.WorkersToDelete {
pod := corev1.Pod{}
pod.Name = podsToDelete
pod.Namespace = utils.GetNamespace(instance.ObjectMeta)
logger.Info("Deleting pod", "namespace", pod.Namespace, "name", pod.Name)
if err := r.Delete(ctx, &pod); err != nil {
if !errors.IsNotFound(err) {
logger.Info("reconcilePods", "Fail to delete Pod", pod.Name, "error", err)
return errstd.Join(utils.ErrFailedDeleteWorkerPod, err)
}
logger.Info("reconcilePods", "The worker Pod has already been deleted", pod.Name)
} else {
deletedWorkers[pod.Name] = deleted
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Deleted", "Deleted pod %s", pod.Name)
}
}
worker.ScaleStrategy.WorkersToDelete = []string{}
runningPods := corev1.PodList{}
for _, pod := range workerPods.Items {
if _, ok := deletedWorkers[pod.Name]; !ok {
runningPods.Items = append(runningPods.Items, pod)
}
}
// A replica can contain multiple hosts, so we need to calculate this based on the number of hosts per replica.
// If the user doesn't install the CRD with `NumOfHosts`, the zero value of `NumOfHosts`, which is 0, will be used.
// Hence, all workers will be deleted. Here, we set `NumOfHosts` to max(1, `NumOfHosts`) to avoid this situation.
if worker.NumOfHosts <= 0 {
worker.NumOfHosts = 1
}
numExpectedPods := workerReplicas * worker.NumOfHosts
diff := numExpectedPods - int32(len(runningPods.Items))
logger.Info("reconcilePods", "workerReplicas", workerReplicas, "NumOfHosts", worker.NumOfHosts, "runningPods", len(runningPods.Items), "diff", diff)
if diff > 0 {
// pods need to be added
logger.Info("reconcilePods", "Number workers to add", diff, "Worker group", worker.GroupName)
// create all workers of this group
var i int32
for i = 0; i < diff; i++ {
logger.Info("reconcilePods", "creating worker for group", worker.GroupName, fmt.Sprintf("index %d", i), fmt.Sprintf("in total %d", diff))
if err := r.createWorkerPod(ctx, *instance, *worker.DeepCopy()); err != nil {
return errstd.Join(utils.ErrFailedCreateWorkerPod, err)
}
}
} else if diff == 0 {
logger.Info("reconcilePods", "all workers already exist for group", worker.GroupName)
continue
} else {
// diff < 0 indicates the need to delete some Pods to match the desired number of replicas. However,
// randomly deleting Pods is certainly not ideal. So, if autoscaling is enabled for the cluster, we
// will disable random Pod deletion, making Autoscaler the sole decision-maker for Pod deletions.
enableInTreeAutoscaling := (instance.Spec.EnableInTreeAutoscaling != nil) && (*instance.Spec.EnableInTreeAutoscaling)
// TODO (kevin85421): `enableRandomPodDelete` is a feature flag for KubeRay v0.6.0. If users want to use
// the old behavior, they can set the environment variable `ENABLE_RANDOM_POD_DELETE` to `true`. When the
// default behavior is stable enough, we can remove this feature flag.
enableRandomPodDelete := false
if enableInTreeAutoscaling {
if s := os.Getenv(utils.ENABLE_RANDOM_POD_DELETE); strings.ToLower(s) == "true" {
enableRandomPodDelete = true
}
}
// Case 1: If Autoscaler is disabled, we will always enable random Pod deletion no matter the value of the feature flag.
// Case 2: If Autoscaler is enabled, we will respect the value of the feature flag. If the feature flag environment variable
// is not set, we will disable random Pod deletion by default.
if !enableInTreeAutoscaling || enableRandomPodDelete {
// diff < 0 means that we need to delete some Pods to meet the desired number of replicas.
randomlyRemovedWorkers := -diff
logger.Info("reconcilePods", "Number workers to delete randomly", randomlyRemovedWorkers, "Worker group", worker.GroupName)
for i := 0; i < int(randomlyRemovedWorkers); i++ {
randomPodToDelete := runningPods.Items[i]
logger.Info("Randomly deleting Pod", "progress", fmt.Sprintf("%d / %d", i+1, randomlyRemovedWorkers), "with name", randomPodToDelete.Name)
if err := r.Delete(ctx, &randomPodToDelete); err != nil {
if !errors.IsNotFound(err) {
return errstd.Join(utils.ErrFailedDeleteWorkerPod, err)
}
logger.Info("reconcilePods", "The worker Pod has already been deleted", randomPodToDelete.Name)
}
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Deleted", "Deleted Pod %s", randomPodToDelete.Name)
}
} else {
logger.Info(fmt.Sprintf("Random Pod deletion is disabled for cluster %s. The only decision-maker for Pod deletions is Autoscaler.", instance.Name))
}
}
}
return nil
}
// shouldDeletePod returns whether the Pod should be deleted and the reason
//
// @param pod: The Pod to be checked.
// @param nodeType: The type of the node that the Pod belongs to (head or worker).
//
// @return: shouldDelete (bool), reason (string)
// (1) shouldDelete: Whether the Pod should be deleted.
// (2) reason: The reason why the Pod should or should not be deleted.
func shouldDeletePod(pod corev1.Pod, nodeType rayv1.RayNodeType) (bool, string) {
// Based on the logic of the change of the status of the K8S pod, the following judgment is made.
// https://github.com/kubernetes/kubernetes/blob/3361895612dac57044d5dacc029d2ace1865479c/pkg/kubelet/kubelet_pods.go#L1556
// If the Pod's status is `Failed` or `Succeeded`, the Pod will not restart and we can safely delete it.
if pod.Status.Phase == corev1.PodFailed || pod.Status.Phase == corev1.PodSucceeded {
reason := fmt.Sprintf(
"The %s Pod %s status is %s which is a terminal state. "+
"KubeRay will delete the Pod and create new Pods in the next reconciliation if necessary.",
nodeType, pod.Name, pod.Status.Phase)
return true, reason
}
rayContainerTerminated := getRayContainerStateTerminated(pod)
if pod.Status.Phase == corev1.PodRunning && rayContainerTerminated != nil {
if pod.Spec.RestartPolicy == corev1.RestartPolicyNever {
reason := fmt.Sprintf(
"The Pod status of the %s Pod %s is %s, and the Ray container terminated status is %v. "+
"The container is unable to restart due to its restart policy %s, so KubeRay will delete it.",
nodeType, pod.Name, pod.Status.Phase, rayContainerTerminated, pod.Spec.RestartPolicy)
return true, reason
}
// If restart policy is set to `Always` or `OnFailure`, KubeRay will not delete the Pod.
reason := fmt.Sprintf(
"The Pod status of the %s Pod %s is %s, and the Ray container terminated status is %v. "+
"However, KubeRay will not delete the Pod because its restartPolicy is set to %s and it should be able to restart automatically.",
nodeType, pod.Name, pod.Status.Phase, rayContainerTerminated, pod.Spec.RestartPolicy)
return false, reason
}
// TODO (kevin85421): Consider deleting a Pod if its Ray container restarts excessively, as this might
// suggest an unhealthy Kubernetes node. Deleting and then recreating the Pod might allow it to be
// scheduled on a different node.
//
// (1) Head Pod:
// It's aggressive to delete a head Pod that is not in a terminated state (i.e., `Failed` or `Succeeded`).
// We should only delete a head Pod when GCS fault tolerance is enabled, and drain the head Pod before
// deleting it.
//
// (2) Worker Pod:
// Compared to deleting a head Pod, removing a worker Pod is less aggressive and aligns more closely with
// the behavior of the Ray Autoscaler. Nevertheless, we should still carefully drain the node before deleting
// the worker Pod. Enabling GCS fault tolerance might not be necessary when deleting worker Pods. Note that
// the Ray Autoscaler will not delete any worker Pods that have never been registered with the Ray cluster.
// Therefore, we may need to address the Ray Autoscaler's blind spots.
reason := fmt.Sprintf(
"KubeRay does not need to delete the %s Pod %s. The Pod status is %s, and the Ray container terminated status is %v.",
nodeType, pod.Name, pod.Status.Phase, rayContainerTerminated)
return false, reason
}
// `ContainerStatuses` does not guarantee the order of the containers. Therefore, we need to find the Ray
// container's status by name. See the following links for more details:
// (1) https://discuss.kubernetes.io/t/pod-spec-containers-and-pod-status-containerstatuses-can-have-a-different-order-why/25273
// (2) https://github.com/kubernetes/kubernetes/blob/03762cbcb52b2a4394e4d795f9d3517a78a5e1a2/pkg/api/v1/pod/util.go#L261-L268
func getRayContainerStateTerminated(pod corev1.Pod) *corev1.ContainerStateTerminated {
rayContainerName := pod.Spec.Containers[utils.RayContainerIndex].Name
for _, containerStatus := range pod.Status.ContainerStatuses {
if containerStatus.Name == rayContainerName {
return containerStatus.State.Terminated
}
}
// If the Ray container isn't found, we'll assume it hasn't terminated. This scenario
// typically arises during testing (`raycluster_controller_test.go`) as `envtest` lacks
// a Pod controller, preventing automatic Pod status updates.
return nil
}
func (r *RayClusterReconciler) createHeadIngress(ctx context.Context, ingress *networkingv1.Ingress, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
// making sure the name is valid
ingress.Name = utils.CheckName(ingress.Name)
if err := controllerutil.SetControllerReference(instance, ingress, r.Scheme); err != nil {
return err
}
if err := r.Create(ctx, ingress); err != nil {
if errors.IsAlreadyExists(err) {
logger.Info("Ingress already exists, no need to create")
return nil
}
return err
}
logger.Info("Ingress created successfully", "ingress name", ingress.Name)
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Created", "Created ingress %s", ingress.Name)
return nil
}
func (r *RayClusterReconciler) createHeadRoute(ctx context.Context, route *routev1.Route, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
// making sure the name is valid
route.Name = utils.CheckRouteName(ctx, route.Name, route.Namespace)
if err := r.Create(ctx, route); err != nil {
if errors.IsAlreadyExists(err) {
logger.Info("Route already exists, no need to create")
return nil
}
return err
}
logger.Info("Route created successfully", "route name", route.Name)
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Created", "Created route %s", route.Name)
return nil
}
func (r *RayClusterReconciler) createService(ctx context.Context, raySvc *corev1.Service, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
// making sure the name is valid
raySvc.Name = utils.CheckName(raySvc.Name)
// Set controller reference
if err := controllerutil.SetControllerReference(instance, raySvc, r.Scheme); err != nil {
return err
}
if err := r.Create(ctx, raySvc); err != nil {
if errors.IsAlreadyExists(err) {
logger.Info("Pod service already exist, no need to create")
return nil
}
return err
}
logger.Info("Pod Service created successfully", "service name", raySvc.Name)
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Created", "Created service %s", raySvc.Name)
return nil
}
func (r *RayClusterReconciler) createHeadPod(ctx context.Context, instance rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
// build the pod then create it
pod := r.buildHeadPod(ctx, instance)
if EnableBatchScheduler {
if scheduler, err := r.BatchSchedulerMgr.GetSchedulerForCluster(&instance); err == nil {
scheduler.AddMetadataToPod(&instance, utils.RayNodeHeadGroupLabelValue, &pod)
} else {
return err
}
}
logger.Info("createHeadPod", "head pod with name", pod.GenerateName)
if err := r.Create(ctx, &pod); err != nil {
return err
}
r.Recorder.Eventf(&instance, corev1.EventTypeNormal, "Created", "Created head pod %s", pod.Name)
return nil
}
func (r *RayClusterReconciler) createWorkerPod(ctx context.Context, instance rayv1.RayCluster, worker rayv1.WorkerGroupSpec) error {
logger := ctrl.LoggerFrom(ctx)
// build the pod then create it
pod := r.buildWorkerPod(ctx, instance, worker)
if EnableBatchScheduler {
if scheduler, err := r.BatchSchedulerMgr.GetSchedulerForCluster(&instance); err == nil {
scheduler.AddMetadataToPod(&instance, worker.GroupName, &pod)
} else {
return err
}
}
if err := r.Create(ctx, &pod); err != nil {
return err
}
logger.Info("Created pod", "Pod ", pod.GenerateName)
r.Recorder.Eventf(&instance, corev1.EventTypeNormal, "Created", "Created worker pod %s", pod.Name)
return nil
}
// Build head instance pod(s).
func (r *RayClusterReconciler) buildHeadPod(ctx context.Context, instance rayv1.RayCluster) corev1.Pod {
logger := ctrl.LoggerFrom(ctx)
podName := utils.PodGenerateName(instance.Name, rayv1.HeadNode)
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)
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...)
}
logger.Info("head pod labels", "labels", podConf.Labels)
creatorCRDType := getCreatorCRDType(instance)
pod := common.BuildPod(ctx, podConf, rayv1.HeadNode, instance.Spec.HeadGroupSpec.RayStartParams, headPort, autoscalingEnabled, creatorCRDType, fqdnRayIP)
// Set raycluster instance as the owner and controller
if err := controllerutil.SetControllerReference(&instance, &pod, r.Scheme); err != nil {
logger.Error(err, "Failed to set controller reference for raycluster pod")
}
return pod
}
func getCreatorCRDType(instance rayv1.RayCluster) utils.CRDType {
return utils.GetCRDType(instance.Labels[utils.RayOriginatedFromCRDLabelKey])
}
// Build worker instance pods.
func (r *RayClusterReconciler) buildWorkerPod(ctx context.Context, instance rayv1.RayCluster, worker rayv1.WorkerGroupSpec) corev1.Pod {
logger := ctrl.LoggerFrom(ctx)
podName := utils.PodGenerateName(fmt.Sprintf("%s-%s", instance.Name, worker.GroupName), rayv1.WorkerNode)
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)
autoscalingEnabled := instance.Spec.EnableInTreeAutoscaling
podTemplateSpec := common.DefaultWorkerPodTemplate(ctx, instance, worker, podName, fqdnRayIP, headPort)
if len(r.workerSidecarContainers) > 0 {
podTemplateSpec.Spec.Containers = append(podTemplateSpec.Spec.Containers, r.workerSidecarContainers...)
}
creatorCRDType := getCreatorCRDType(instance)
pod := common.BuildPod(ctx, podTemplateSpec, rayv1.WorkerNode, worker.RayStartParams, headPort, autoscalingEnabled, creatorCRDType, fqdnRayIP)
// Set raycluster instance as the owner and controller
if err := controllerutil.SetControllerReference(&instance, &pod, r.Scheme); err != nil {
logger.Error(err, "Failed to set controller reference for raycluster pod")
}
return pod
}
func (r *RayClusterReconciler) buildRedisCleanupJob(ctx context.Context, instance rayv1.RayCluster) batchv1.Job {
logger := ctrl.LoggerFrom(ctx)
pod := r.buildHeadPod(ctx, instance)
pod.Labels[utils.RayNodeTypeLabelKey] = string(rayv1.RedisCleanupNode)
// Only keep the Ray container in the Redis cleanup Job.
pod.Spec.Containers = []corev1.Container{pod.Spec.Containers[utils.RayContainerIndex]}
pod.Spec.Containers[utils.RayContainerIndex].Command = []string{"/bin/bash", "-lc", "--"}
pod.Spec.Containers[utils.RayContainerIndex].Args = []string{
"echo \"To get more information about manually delete the storage namespace in Redis and remove the RayCluster's finalizer, please check https://docs.ray.io/en/master/cluster/kubernetes/user-guides/kuberay-gcs-ft.html for more details.\" && " +
"python -c " +
"\"from ray._private.gcs_utils import cleanup_redis_storage; " +
"from urllib.parse import urlparse; " +
"import os; " +
"import sys; " +
"redis_address = os.getenv('RAY_REDIS_ADDRESS', '').split(',')[0]; " +
"redis_address = redis_address if '://' in redis_address else 'redis://' + redis_address; " +
"parsed = urlparse(redis_address); " +
"sys.exit(1) if not cleanup_redis_storage(host=parsed.hostname, port=parsed.port, password=os.getenv('REDIS_PASSWORD', parsed.password), use_ssl=parsed.scheme=='rediss', storage_namespace=os.getenv('RAY_external_storage_namespace')) else None\"",
}
// Disable liveness and readiness probes because the Job will not launch processes like Raylet and GCS.
pod.Spec.Containers[utils.RayContainerIndex].LivenessProbe = nil
pod.Spec.Containers[utils.RayContainerIndex].ReadinessProbe = nil
// Set the environment variables to ensure that the cleanup Job has at least 60s.
pod.Spec.Containers[utils.RayContainerIndex].Env = append(pod.Spec.Containers[utils.RayContainerIndex].Env, corev1.EnvVar{
Name: "RAY_redis_db_connect_retries",
Value: "120",
})
pod.Spec.Containers[utils.RayContainerIndex].Env = append(pod.Spec.Containers[utils.RayContainerIndex].Env, corev1.EnvVar{
Name: "RAY_redis_db_connect_wait_milliseconds",
Value: "500",
})
// The container's resource consumption remains constant. so hard-coding the resources is acceptable.
// In addition, avoid using the GPU for the Redis cleanup Job.
pod.Spec.Containers[utils.RayContainerIndex].Resources = corev1.ResourceRequirements{
Limits: corev1.ResourceList{
corev1.ResourceCPU: resource.MustParse("200m"),
corev1.ResourceMemory: resource.MustParse("256Mi"),
},
Requests: corev1.ResourceList{
corev1.ResourceCPU: resource.MustParse("200m"),
corev1.ResourceMemory: resource.MustParse("256Mi"),
},
}
// For Kubernetes Job, the valid values for Pod's `RestartPolicy` are `Never` and `OnFailure`.
pod.Spec.RestartPolicy = corev1.RestartPolicyNever
redisCleanupJob := batchv1.Job{
ObjectMeta: metav1.ObjectMeta{
Name: fmt.Sprintf("%s-%s", instance.Name, "redis-cleanup"),
Namespace: instance.Namespace,
Labels: pod.Labels,
Annotations: pod.Annotations,
},
Spec: batchv1.JobSpec{
BackoffLimit: ptr.To[int32](0),
Template: corev1.PodTemplateSpec{
ObjectMeta: pod.ObjectMeta,
Spec: pod.Spec,
},
// make this job be best-effort only for 5 minutes.
ActiveDeadlineSeconds: ptr.To[int64](300),
},
}
if err := controllerutil.SetControllerReference(&instance, &redisCleanupJob, r.Scheme); err != nil {
logger.Error(err, "Failed to set controller reference for the Redis cleanup Job.")
}
return redisCleanupJob
}
// SetupWithManager builds the reconciler.
func (r *RayClusterReconciler) SetupWithManager(mgr ctrl.Manager, reconcileConcurrency int) error {
b := ctrl.NewControllerManagedBy(mgr).
For(&rayv1.RayCluster{}, builder.WithPredicates(predicate.Or(
predicate.GenerationChangedPredicate{},
predicate.LabelChangedPredicate{},
predicate.AnnotationChangedPredicate{},
))).
Owns(&corev1.Pod{}).
Owns(&corev1.Service{})
if EnableBatchScheduler {
b = batchscheduler.ConfigureReconciler(b)
}
return b.
WithOptions(controller.Options{
MaxConcurrentReconciles: reconcileConcurrency,
LogConstructor: func(request *reconcile.Request) logr.Logger {
logger := ctrl.Log.WithName("controllers").WithName("RayCluster")
if request != nil {
logger = logger.WithValues("RayCluster", request.NamespacedName)
}
return logger
},
}).
Complete(r)
}
func (r *RayClusterReconciler) calculateStatus(ctx context.Context, instance *rayv1.RayCluster, reconcileErr error) (*rayv1.RayCluster, error) {
// TODO: Replace this log and use reconcileErr to set the condition field.
logger := ctrl.LoggerFrom(ctx)
if reconcileErr != nil {
logger.Info("Reconciliation error", "error", reconcileErr)
}
// Deep copy the instance, so we don't mutate the original object.
newInstance := instance.DeepCopy()
if features.Enabled(features.RayClusterStatusConditions) {
if reconcileErr != nil {
if reason := utils.RayClusterReplicaFailureReason(reconcileErr); reason != "" {
meta.SetStatusCondition(&newInstance.Status.Conditions, metav1.Condition{
Type: string(rayv1.RayClusterReplicaFailure),
Status: metav1.ConditionTrue,
Reason: reason,
Message: reconcileErr.Error(),
})
}
} else {
// if reconcileErr == nil, we can safely remove the RayClusterReplicaFailure condition.
meta.RemoveStatusCondition(&newInstance.Status.Conditions, string(rayv1.RayClusterReplicaFailure))
}
}
// TODO (kevin85421): ObservedGeneration should be used to determine whether to update this CR or not.
newInstance.Status.ObservedGeneration = newInstance.ObjectMeta.Generation
runtimePods := corev1.PodList{}
filterLabels := client.MatchingLabels{utils.RayClusterLabelKey: newInstance.Name}
if err := r.List(ctx, &runtimePods, client.InNamespace(newInstance.Namespace), filterLabels); err != nil {
return nil, err
}
newInstance.Status.ReadyWorkerReplicas = utils.CalculateReadyReplicas(runtimePods)
newInstance.Status.AvailableWorkerReplicas = utils.CalculateAvailableReplicas(runtimePods)
newInstance.Status.DesiredWorkerReplicas = utils.CalculateDesiredReplicas(ctx, newInstance)
newInstance.Status.MinWorkerReplicas = utils.CalculateMinReplicas(newInstance)
newInstance.Status.MaxWorkerReplicas = utils.CalculateMaxReplicas(newInstance)
totalResources := utils.CalculateDesiredResources(newInstance)
newInstance.Status.DesiredCPU = totalResources[corev1.ResourceCPU]
newInstance.Status.DesiredMemory = totalResources[corev1.ResourceMemory]
newInstance.Status.DesiredGPU = sumGPUs(totalResources)
newInstance.Status.DesiredTPU = totalResources[corev1.ResourceName("google.com/tpu")]
if utils.CheckAllPodsRunning(ctx, runtimePods) {
newInstance.Status.State = rayv1.Ready
}
// Check if the head node is running and ready by checking the head pod's status.
if features.Enabled(features.RayClusterStatusConditions) {
headPod, err := common.GetRayClusterHeadPod(ctx, r, newInstance)
if err != nil {
return nil, err
}
// GetRayClusterHeadPod can return nil, nil when pod is not found, we handle it separately.
if headPod == nil {
meta.SetStatusCondition(&newInstance.Status.Conditions, metav1.Condition{
Type: string(rayv1.HeadPodReady),
Status: metav1.ConditionFalse,
Reason: rayv1.HeadPodNotFound,
Message: "Head Pod not found",
})
} else {
headPodReadyCondition := utils.FindHeadPodReadyCondition(headPod)
meta.SetStatusCondition(&newInstance.Status.Conditions, headPodReadyCondition)
}
if !meta.IsStatusConditionTrue(newInstance.Status.Conditions, string(rayv1.RayClusterProvisioned)) {
// RayClusterProvisioned indicates whether all Ray Pods are ready when the RayCluster is first created.
// Note RayClusterProvisioned StatusCondition will not be updated after all Ray Pods are ready for the first time.
if utils.CheckAllPodsRunning(ctx, runtimePods) {
meta.SetStatusCondition(&newInstance.Status.Conditions, metav1.Condition{
Type: string(rayv1.RayClusterProvisioned),
Status: metav1.ConditionTrue,
Reason: rayv1.AllPodRunningAndReadyFirstTime,
Message: "All Ray Pods are ready for the first time",
})
} else {
meta.SetStatusCondition(&newInstance.Status.Conditions, metav1.Condition{
Type: string(rayv1.RayClusterProvisioned),
Status: metav1.ConditionFalse,
Reason: rayv1.RayClusterPodsProvisioning,
Message: "RayCluster Pods are being provisioned for first time",
})
}
}
}
if newInstance.Spec.Suspend != nil && *newInstance.Spec.Suspend && len(runtimePods.Items) == 0 {
newInstance.Status.State = rayv1.Suspended
}
if err := r.updateEndpoints(ctx, newInstance); err != nil {
return nil, err
}
if err := r.updateHeadInfo(ctx, newInstance); err != nil {
return nil, err
}
timeNow := metav1.Now()
newInstance.Status.LastUpdateTime = &timeNow
if instance.Status.State != newInstance.Status.State {
if newInstance.Status.StateTransitionTimes == nil {
newInstance.Status.StateTransitionTimes = make(map[rayv1.ClusterState]*metav1.Time)
}
newInstance.Status.StateTransitionTimes[newInstance.Status.State] = &timeNow
}
return newInstance, nil
}
func (r *RayClusterReconciler) getHeadServiceIPAndName(ctx context.Context, instance *rayv1.RayCluster) (string, string, error) {
runtimeServices := corev1.ServiceList{}
if err := r.List(ctx, &runtimeServices, common.RayClusterHeadServiceListOptions(instance)...); err != nil {
return "", "", err
}
if len(runtimeServices.Items) < 1 {
return "", "", fmt.Errorf("unable to find head service. cluster name %s, filter labels %v", instance.Name, common.RayClusterHeadServiceListOptions(instance))
} else if len(runtimeServices.Items) > 1 {
return "", "", fmt.Errorf("found multiple head services. cluster name %s, filter labels %v", instance.Name, common.RayClusterHeadServiceListOptions(instance))
} else if runtimeServices.Items[0].Spec.ClusterIP == "" {
return "", "", fmt.Errorf("head service IP is empty. cluster name %s, filter labels %v", instance.Name, common.RayClusterHeadServiceListOptions(instance))
} else if runtimeServices.Items[0].Spec.ClusterIP == corev1.ClusterIPNone {
// We return Head Pod IP if the Head service is headless.
headPod, err := common.GetRayClusterHeadPod(ctx, r, instance)
if err != nil {
return "", "", err
}
if headPod != nil {
return headPod.Status.PodIP, runtimeServices.Items[0].Name, nil
}
return "", runtimeServices.Items[0].Name, nil
}
return runtimeServices.Items[0].Spec.ClusterIP, runtimeServices.Items[0].Name, nil
}
func (r *RayClusterReconciler) updateEndpoints(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
// TODO: (@scarlet25151) There may be several K8s Services for a RayCluster.
// We assume we can find the right one by filtering Services with appropriate label selectors
// and picking the first one. We may need to select by name in the future if the Service naming is stable.
rayHeadSvc := corev1.ServiceList{}
filterLabels := client.MatchingLabels{
utils.RayClusterLabelKey: instance.Name,
utils.RayNodeTypeLabelKey: "head",
}
if err := r.List(ctx, &rayHeadSvc, client.InNamespace(instance.Namespace), filterLabels); err != nil {
return err
}
if len(rayHeadSvc.Items) != 0 {
svc := rayHeadSvc.Items[0]
if instance.Status.Endpoints == nil {
instance.Status.Endpoints = map[string]string{}
}
for _, port := range svc.Spec.Ports {
if len(port.Name) == 0 {
logger.Info("updateStatus", "service port's name is empty. Not adding it to RayCluster status.endpoints", port)
continue
}
if port.NodePort != 0 {
instance.Status.Endpoints[port.Name] = fmt.Sprintf("%d", port.NodePort)
} else if port.TargetPort.IntVal != 0 {
instance.Status.Endpoints[port.Name] = fmt.Sprintf("%d", port.TargetPort.IntVal)
} else if port.TargetPort.StrVal != "" {
instance.Status.Endpoints[port.Name] = port.TargetPort.StrVal
} else {
logger.Info("updateStatus", "service port's targetPort is empty. Not adding it to RayCluster status.endpoints", port)
}
}
} else {
logger.Info("updateEndpoints", "unable to find a Service for this RayCluster. Not adding RayCluster status.endpoints", instance.Name, "Service selectors", filterLabels)
}
return nil
}
func (r *RayClusterReconciler) updateHeadInfo(ctx context.Context, instance *rayv1.RayCluster) error {
headPod, err := common.GetRayClusterHeadPod(ctx, r, instance)
if err != nil {
return err
}
if headPod != nil {
instance.Status.Head.PodIP = headPod.Status.PodIP
instance.Status.Head.PodName = headPod.Name
}
ip, name, err := r.getHeadServiceIPAndName(ctx, instance)
if err != nil {
return err
}
instance.Status.Head.ServiceIP = ip
instance.Status.Head.ServiceName = name
return nil
}
func (r *RayClusterReconciler) reconcileAutoscalerServiceAccount(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
if instance.Spec.EnableInTreeAutoscaling == nil || !*instance.Spec.EnableInTreeAutoscaling {
return nil
}
serviceAccount := &corev1.ServiceAccount{}
namespacedName := common.RayClusterAutoscalerServiceAccountNamespacedName(instance)
if err := r.Get(ctx, namespacedName, serviceAccount); err != nil {
if !errors.IsNotFound(err) {
return err
}
// If users specify ServiceAccountName for the head Pod, they need to create a ServiceAccount themselves.
// However, if KubeRay creates a ServiceAccount for users, the autoscaler may encounter permission issues during
// zero-downtime rolling updates when RayService is performed. See https://github.com/ray-project/kuberay/issues/1123
// for more details.
if instance.Spec.HeadGroupSpec.Template.Spec.ServiceAccountName == namespacedName.Name {
logger.Error(err, fmt.Sprintf(
"If users specify ServiceAccountName for the head Pod, they need to create a ServiceAccount themselves. "+
"However, ServiceAccount %s is not found. Please create one. "+
"See the PR description of https://github.com/ray-project/kuberay/pull/1128 for more details.", namespacedName.Name), "ServiceAccount", namespacedName)
return err
}
// Create service account for autoscaler if there's no existing one in the cluster.
serviceAccount, err := common.BuildServiceAccount(instance)
if err != nil {
return err
}
// making sure the name is valid
serviceAccount.Name = utils.CheckName(serviceAccount.Name)
// Set controller reference
if err := controllerutil.SetControllerReference(instance, serviceAccount, r.Scheme); err != nil {
return err
}
if err := r.Create(ctx, serviceAccount); err != nil {
if errors.IsAlreadyExists(err) {
logger.Info("Pod service account already exist, no need to create")
return nil
}
return err
}
logger.Info("Pod ServiceAccount created successfully", "service account name", serviceAccount.Name)
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Created", "Created service account %s", serviceAccount.Name)
return nil
}
return nil
}
func (r *RayClusterReconciler) reconcileAutoscalerRole(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
if instance.Spec.EnableInTreeAutoscaling == nil || !*instance.Spec.EnableInTreeAutoscaling {
return nil
}
role := &rbacv1.Role{}
namespacedName := common.RayClusterAutoscalerRoleNamespacedName(instance)
if err := r.Get(ctx, namespacedName, role); err != nil {
if !errors.IsNotFound(err) {
return err
}
// Create role for autoscaler if there's no existing one in the cluster.
role, err := common.BuildRole(instance)
if err != nil {
return err
}
// making sure the name is valid
role.Name = utils.CheckName(role.Name)
// Set controller reference
if err := controllerutil.SetControllerReference(instance, role, r.Scheme); err != nil {
return err
}
if err := r.Create(ctx, role); err != nil {
if errors.IsAlreadyExists(err) {
logger.Info("role already exist, no need to create")
return nil
}
return err
}
logger.Info("Role created successfully", "role name", role.Name)
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Created", "Created role %s", role.Name)
return nil
}
return nil
}
func (r *RayClusterReconciler) reconcileAutoscalerRoleBinding(ctx context.Context, instance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
if instance.Spec.EnableInTreeAutoscaling == nil || !*instance.Spec.EnableInTreeAutoscaling {
return nil
}
roleBinding := &rbacv1.RoleBinding{}
namespacedName := common.RayClusterAutoscalerRoleBindingNamespacedName(instance)
if err := r.Get(ctx, namespacedName, roleBinding); err != nil {
if !errors.IsNotFound(err) {
return err
}
// Create role bindings for autoscaler if there's no existing one in the cluster.
roleBinding, err := common.BuildRoleBinding(instance)
if err != nil {
return err
}
// making sure the name is valid
roleBinding.Name = utils.CheckName(roleBinding.Name)
// Set controller reference
if err := controllerutil.SetControllerReference(instance, roleBinding, r.Scheme); err != nil {
return err
}
if err := r.Create(ctx, roleBinding); err != nil {
if errors.IsAlreadyExists(err) {
logger.Info("role binding already exist, no need to create")
return nil
}
return err
}
logger.Info("RoleBinding created successfully", "role binding name", roleBinding.Name)
r.Recorder.Eventf(instance, corev1.EventTypeNormal, "Created", "Created role binding %s", roleBinding.Name)
return nil
}
return nil
}
func (r *RayClusterReconciler) updateRayClusterStatus(ctx context.Context, originalRayClusterInstance, newInstance *rayv1.RayCluster) error {
logger := ctrl.LoggerFrom(ctx)
if !r.inconsistentRayClusterStatus(ctx, originalRayClusterInstance.Status, newInstance.Status) {
return nil
}
logger.Info("updateRayClusterStatus", "name", originalRayClusterInstance.Name, "old status", originalRayClusterInstance.Status, "new status", newInstance.Status)
err := r.Status().Update(ctx, newInstance)
if err != nil {
logger.Info("Error updating status", "name", originalRayClusterInstance.Name, "error", err, "RayCluster", newInstance)
}
return err
}
// sumGPUs sums the GPUs in the given resource list.
func sumGPUs(resources map[corev1.ResourceName]resource.Quantity) resource.Quantity {
totalGPUs := resource.Quantity{}
for key, val := range resources {
if strings.HasSuffix(string(key), "gpu") && !val.IsZero() {
totalGPUs.Add(val)
}
}
return totalGPUs
}