mirror of
https://github.com/hanzo-ml/kuberay.git
synced 2026-10-30 19:48:47 +00:00
900 lines
35 KiB
Go
900 lines
35 KiB
Go
package common
|
|
|
|
import (
|
|
"bytes"
|
|
"fmt"
|
|
"strconv"
|
|
"strings"
|
|
|
|
"github.com/ray-project/kuberay/ray-operator/controllers/ray/utils"
|
|
|
|
rayiov1alpha1 "github.com/ray-project/kuberay/ray-operator/apis/ray/v1alpha1"
|
|
"k8s.io/apimachinery/pkg/api/errors"
|
|
"k8s.io/apimachinery/pkg/api/resource"
|
|
|
|
logf "sigs.k8s.io/controller-runtime/pkg/log"
|
|
|
|
v1 "k8s.io/api/core/v1"
|
|
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
|
|
)
|
|
|
|
const (
|
|
SharedMemoryVolumeName = "shared-mem"
|
|
SharedMemoryVolumeMountPath = "/dev/shm"
|
|
RayLogVolumeName = "ray-logs"
|
|
RayLogVolumeMountPath = "/tmp/ray"
|
|
AutoscalerContainerName = "autoscaler"
|
|
RayHeadContainer = "ray-head"
|
|
ObjectStoreMemoryKey = "object-store-memory"
|
|
// TODO (davidxia): should be a const in upstream ray-project/ray
|
|
AllowSlowStorageEnvVar = "RAY_OBJECT_STORE_ALLOW_SLOW_STORAGE"
|
|
)
|
|
|
|
var log = logf.Log.WithName("RayCluster-Controller")
|
|
|
|
// Get the port required to connect to the Ray cluster by worker nodes and drivers
|
|
// started within the cluster.
|
|
// For Ray >= 1.11.0 this is the GCS server port. For Ray < 1.11.0 it is the Redis port.
|
|
func GetHeadPort(headStartParams map[string]string) string {
|
|
var headPort string
|
|
if value, ok := headStartParams["port"]; !ok {
|
|
// using default port
|
|
headPort = strconv.Itoa(DefaultRedisPort)
|
|
} else {
|
|
// setting port from the params
|
|
headPort = value
|
|
}
|
|
return headPort
|
|
}
|
|
|
|
// rayClusterHAEnabled check if RayCluster enabled FT in annotations
|
|
func rayClusterHAEnabled(instance rayiov1alpha1.RayCluster) bool {
|
|
if instance.Annotations == nil {
|
|
return false
|
|
}
|
|
if v, ok := instance.Annotations[RayFTEnabledAnnotationKey]; ok {
|
|
if strings.ToLower(v) == "true" {
|
|
return true
|
|
}
|
|
}
|
|
return false
|
|
}
|
|
|
|
func initTemplateAnnotations(instance rayiov1alpha1.RayCluster, podTemplate *v1.PodTemplateSpec) {
|
|
if podTemplate.Annotations == nil {
|
|
podTemplate.Annotations = make(map[string]string)
|
|
}
|
|
|
|
// For now, we just set ray external storage enabled/disabled by checking if FT is enabled/disabled.
|
|
// This may need to be updated in the future.
|
|
if rayClusterHAEnabled(instance) {
|
|
podTemplate.Annotations[RayFTEnabledAnnotationKey] = "true"
|
|
// if we have FT enabled, we need to set up a default external storage namespace.
|
|
podTemplate.Annotations[RayExternalStorageNSAnnotationKey] = string(instance.UID)
|
|
} else {
|
|
podTemplate.Annotations[RayFTEnabledAnnotationKey] = "false"
|
|
}
|
|
podTemplate.Annotations[RayNodeHealthStateAnnotationKey] = ""
|
|
|
|
// set ray external storage namespace if user specified one.
|
|
if instance.Annotations != nil {
|
|
if v, ok := instance.Annotations[RayExternalStorageNSAnnotationKey]; ok {
|
|
podTemplate.Annotations[RayExternalStorageNSAnnotationKey] = v
|
|
}
|
|
}
|
|
}
|
|
|
|
// DefaultHeadPodTemplate sets the config values
|
|
func DefaultHeadPodTemplate(instance rayiov1alpha1.RayCluster, headSpec rayiov1alpha1.HeadGroupSpec, podName string, headPort string) v1.PodTemplateSpec {
|
|
// TODO (Dmitri) The argument headPort is essentially unused;
|
|
// headPort is passed into setMissingRayStartParams but unused there for the head pod.
|
|
// To mitigate this awkwardness and reduce code redundancy, unify head and worker pod configuration logic.
|
|
podTemplate := headSpec.Template
|
|
podTemplate.GenerateName = podName
|
|
if podTemplate.ObjectMeta.Namespace == "" {
|
|
podTemplate.ObjectMeta.Namespace = instance.Namespace
|
|
log.Info("Setting pod namespaces", "namespace", instance.Namespace)
|
|
}
|
|
|
|
if podTemplate.Labels == nil {
|
|
podTemplate.Labels = make(map[string]string)
|
|
}
|
|
podTemplate.Labels = labelPod(rayiov1alpha1.HeadNode, instance.Name, "headgroup", instance.Spec.HeadGroupSpec.Template.ObjectMeta.Labels)
|
|
headSpec.RayStartParams = setMissingRayStartParams(headSpec.RayStartParams, rayiov1alpha1.HeadNode, headPort, "")
|
|
headSpec.RayStartParams = setAgentListPortStartParams(instance, headSpec.RayStartParams)
|
|
|
|
initTemplateAnnotations(instance, &podTemplate)
|
|
|
|
// if in-tree autoscaling is enabled, then autoscaler container should be injected into head pod.
|
|
if instance.Spec.EnableInTreeAutoscaling != nil && *instance.Spec.EnableInTreeAutoscaling {
|
|
headSpec.RayStartParams["no-monitor"] = "true"
|
|
// set custom service account with proper roles bound.
|
|
// utils.CheckName clips the name to match the behavior of reconcileAutoscalerServiceAccount
|
|
podTemplate.Spec.ServiceAccountName = utils.CheckName(utils.GetHeadGroupServiceAccountName(&instance))
|
|
|
|
rayContainerIndex := getRayContainerIndex(podTemplate.Spec)
|
|
rayHeadImage := podTemplate.Spec.Containers[rayContainerIndex].Image
|
|
// Determine the default image to use for the Ray container.
|
|
autoscalerImage := getAutoscalerImage(rayHeadImage, instance.Spec.RayVersion)
|
|
// inject autoscaler container into head pod
|
|
autoscalerContainer := BuildAutoscalerContainer(autoscalerImage)
|
|
// Merge the user overrides from autoscalerOptions into the autoscaler container config.
|
|
mergeAutoscalerOverrides(&autoscalerContainer, instance.Spec.AutoscalerOptions)
|
|
podTemplate.Spec.Containers = append(podTemplate.Spec.Containers, autoscalerContainer)
|
|
}
|
|
|
|
isMetricsPortExists := false
|
|
for _, port := range podTemplate.Spec.Containers[0].Ports {
|
|
if port.Name == DefaultMetricsName {
|
|
isMetricsPortExists = true
|
|
break
|
|
}
|
|
}
|
|
if !isMetricsPortExists {
|
|
// add metrics port for exposing to the promethues stack.
|
|
metricsPort := v1.ContainerPort{
|
|
Name: DefaultMetricsName,
|
|
ContainerPort: int32(DefaultMetricsPort),
|
|
}
|
|
podTemplate.Spec.Containers[0].Ports = append(podTemplate.Spec.Containers[0].Ports, metricsPort)
|
|
}
|
|
|
|
return podTemplate
|
|
}
|
|
|
|
// getAutoscalerImage determines the default autoscaler image
|
|
func getAutoscalerImage(rayHeadImage string, rayVersion string) string {
|
|
if autoscalerSupportIsStable(rayVersion) {
|
|
// For Ray versions >= 2.0.0, use the Ray head's image to run the autoscaler.
|
|
return rayHeadImage
|
|
} else {
|
|
// For older Ray versions, use the Ray 2.0.0 image to run the autoscaler.
|
|
return FallbackDefaultAutoscalerImage
|
|
}
|
|
}
|
|
|
|
// Determine if autoscaler support is stable in the given rayVersion.
|
|
// Return false exactly when the major version is successfully parsed and less than 2.
|
|
// Example rayVersion inputs that return true: "2.0.0", "2.0", "2", "2.0.0rc1", "nightly", "latest", "unknown".
|
|
// Example inputs that return false: "1.13.0", "1.12", "1".
|
|
func autoscalerSupportIsStable(rayVersion string) bool {
|
|
// Try to determine major version by extracting everything that comes before the first "."
|
|
firstDotIndex := strings.Index(rayVersion, ".")
|
|
var majorVersionString string
|
|
if firstDotIndex == -1 {
|
|
// If there is no ".", try parsing the entire rayVersion as the major version.
|
|
majorVersionString = rayVersion
|
|
} else {
|
|
// Everything up to the first "."
|
|
majorVersionString = rayVersion[:firstDotIndex]
|
|
}
|
|
|
|
if majorVersion, err := strconv.Atoi(majorVersionString); err == nil {
|
|
return majorVersion >= 2
|
|
} else {
|
|
// If in doubt, just assume that the Ray version is >= 2.0.0,
|
|
// so that we use the Ray image to run the autoscaler.
|
|
// Currently, there is a lot of "doubt," since the version string is not validated.
|
|
// Users can always override the operator's choice of image with autoscalerOptions.image.
|
|
return true
|
|
}
|
|
}
|
|
|
|
// DefaultWorkerPodTemplate sets the config values
|
|
func DefaultWorkerPodTemplate(instance rayiov1alpha1.RayCluster, workerSpec rayiov1alpha1.WorkerGroupSpec, podName string, fqdnRayIP string, headPort string) v1.PodTemplateSpec {
|
|
podTemplate := workerSpec.Template
|
|
podTemplate.GenerateName = podName
|
|
if podTemplate.ObjectMeta.Namespace == "" {
|
|
podTemplate.ObjectMeta.Namespace = instance.Namespace
|
|
log.Info("Setting pod namespaces", "namespace", instance.Namespace)
|
|
}
|
|
|
|
// The Ray worker should only start once the GCS server is ready.
|
|
rayContainerIndex := getRayContainerIndex(podTemplate.Spec)
|
|
initContainer := v1.Container{
|
|
Name: "wait-gcs-ready",
|
|
Image: podTemplate.Spec.Containers[rayContainerIndex].Image,
|
|
ImagePullPolicy: v1.PullIfNotPresent,
|
|
Command: []string{"/bin/bash", "-lc", "--"},
|
|
Args: []string{
|
|
fmt.Sprintf("until ray health-check --address %s:%s > /dev/null 2>&1; do echo wait for GCS to be ready; sleep 5; done", fqdnRayIP, headPort),
|
|
},
|
|
SecurityContext: podTemplate.Spec.Containers[rayContainerIndex].SecurityContext.DeepCopy(),
|
|
// This init container requires certain environment variables to establish a secure connection with the Ray head using TLS authentication.
|
|
// Additionally, some of these environment variables may reference files stored in volumes, so we need to include both the `Env` and `VolumeMounts` fields here.
|
|
// For more details, please refer to: https://docs.ray.io/en/latest/ray-core/configure.html#tls-authentication.
|
|
Env: podTemplate.Spec.Containers[rayContainerIndex].DeepCopy().Env,
|
|
VolumeMounts: podTemplate.Spec.Containers[rayContainerIndex].DeepCopy().VolumeMounts,
|
|
}
|
|
podTemplate.Spec.InitContainers = append(podTemplate.Spec.InitContainers, initContainer)
|
|
|
|
// If the replica of workers is more than 1, `ObjectMeta.Name` may cause name conflict errors.
|
|
// Hence, we set `ObjectMeta.Name` to an empty string, and use GenerateName to prevent name conflicts.
|
|
podTemplate.ObjectMeta.Name = ""
|
|
if podTemplate.Labels == nil {
|
|
podTemplate.Labels = make(map[string]string)
|
|
}
|
|
podTemplate.Labels = labelPod(rayiov1alpha1.WorkerNode, instance.Name, workerSpec.GroupName, workerSpec.Template.ObjectMeta.Labels)
|
|
workerSpec.RayStartParams = setMissingRayStartParams(workerSpec.RayStartParams, rayiov1alpha1.WorkerNode, headPort, fqdnRayIP)
|
|
workerSpec.RayStartParams = setAgentListPortStartParams(instance, workerSpec.RayStartParams)
|
|
|
|
initTemplateAnnotations(instance, &podTemplate)
|
|
|
|
isMetricsPortExists := false
|
|
for _, port := range podTemplate.Spec.Containers[rayContainerIndex].Ports {
|
|
if port.Name == DefaultMetricsName {
|
|
isMetricsPortExists = true
|
|
break
|
|
}
|
|
}
|
|
if !isMetricsPortExists {
|
|
// add metrics port for exposing to the promethues stack.
|
|
metricsPort := v1.ContainerPort{
|
|
Name: DefaultMetricsName,
|
|
ContainerPort: int32(DefaultMetricsPort),
|
|
}
|
|
podTemplate.Spec.Containers[rayContainerIndex].Ports = append(podTemplate.Spec.Containers[rayContainerIndex].Ports, metricsPort)
|
|
}
|
|
|
|
return podTemplate
|
|
}
|
|
|
|
func initLivenessProbeHandler(probe *v1.Probe, rayNodeType rayiov1alpha1.RayNodeType) {
|
|
if probe.Exec == nil {
|
|
// we only create the probe if user did not specify any.
|
|
if rayNodeType == rayiov1alpha1.HeadNode {
|
|
// head node liveness probe
|
|
cmd := []string{
|
|
"bash", "-c", fmt.Sprintf("wget -T 2 -q -O- http://localhost:%d/%s | grep success",
|
|
DefaultDashboardAgentListenPort, RayAgentRayletHealthPath),
|
|
"&&", "bash", "-c", fmt.Sprintf("wget -T 2 -q -O- http://localhost:%d/%s | grep success",
|
|
DefaultDashboardPort, RayDashboardGCSHealthPath),
|
|
}
|
|
probe.Exec = &v1.ExecAction{Command: cmd}
|
|
} else {
|
|
// worker node liveness probe
|
|
cmd := []string{
|
|
"bash", "-c", fmt.Sprintf("wget -T 2 -q -O- http://localhost:%d/%s | grep success",
|
|
DefaultDashboardAgentListenPort, RayAgentRayletHealthPath),
|
|
}
|
|
probe.Exec = &v1.ExecAction{Command: cmd}
|
|
}
|
|
}
|
|
}
|
|
|
|
func initReadinessProbeHandler(probe *v1.Probe, rayNodeType rayiov1alpha1.RayNodeType) {
|
|
if probe.Exec == nil {
|
|
// we only create the probe if user did not specify any.
|
|
if rayNodeType == rayiov1alpha1.HeadNode {
|
|
// head node readiness probe
|
|
cmd := []string{
|
|
"bash", "-c", fmt.Sprintf("wget -T 2 -q -O- http://localhost:%d/%s | grep success",
|
|
DefaultDashboardAgentListenPort, RayAgentRayletHealthPath),
|
|
"&&", "bash", "-c", fmt.Sprintf("wget -T 2 -q -O- http://localhost:%d/%s | grep success",
|
|
DefaultDashboardPort, RayDashboardGCSHealthPath),
|
|
}
|
|
probe.Exec = &v1.ExecAction{Command: cmd}
|
|
} else {
|
|
// worker node readiness probe
|
|
cmd := []string{
|
|
"bash", "-c", fmt.Sprintf("wget -T 2 -q -O- http://localhost:%d/%s | grep success",
|
|
DefaultDashboardAgentListenPort, RayAgentRayletHealthPath),
|
|
}
|
|
probe.Exec = &v1.ExecAction{Command: cmd}
|
|
}
|
|
}
|
|
}
|
|
|
|
// BuildPod a pod config
|
|
func BuildPod(podTemplateSpec v1.PodTemplateSpec, rayNodeType rayiov1alpha1.RayNodeType, rayStartParams map[string]string, headPort string, enableRayAutoscaler *bool, creator string, fqdnRayIP string) (aPod v1.Pod) {
|
|
pod := v1.Pod{
|
|
TypeMeta: metav1.TypeMeta{
|
|
APIVersion: "v1",
|
|
Kind: "Pod",
|
|
},
|
|
ObjectMeta: podTemplateSpec.ObjectMeta,
|
|
Spec: podTemplateSpec.Spec,
|
|
}
|
|
rayContainerIndex := getRayContainerIndex(pod.Spec)
|
|
|
|
// Add /dev/shm volumeMount for the object store to avoid performance degradation.
|
|
addEmptyDir(&pod.Spec.Containers[rayContainerIndex], &pod, SharedMemoryVolumeName, SharedMemoryVolumeMountPath, v1.StorageMediumMemory)
|
|
if rayNodeType == rayiov1alpha1.HeadNode && enableRayAutoscaler != nil && *enableRayAutoscaler {
|
|
// The Ray autoscaler writes logs which are read by the Ray head.
|
|
// We need a shared log volume to enable this information flow.
|
|
// Specifically, this is required for the event-logging functionality
|
|
// introduced in https://github.com/ray-project/ray/pull/13434.
|
|
autoscalerContainerIndex := getAutoscalerContainerIndex(pod)
|
|
addEmptyDir(&pod.Spec.Containers[rayContainerIndex], &pod, RayLogVolumeName, RayLogVolumeMountPath, v1.StorageMediumDefault)
|
|
addEmptyDir(&pod.Spec.Containers[autoscalerContainerIndex], &pod, RayLogVolumeName, RayLogVolumeMountPath, v1.StorageMediumDefault)
|
|
}
|
|
cleanupInvalidVolumeMounts(&pod.Spec.Containers[rayContainerIndex], &pod)
|
|
if len(pod.Spec.InitContainers) > rayContainerIndex {
|
|
cleanupInvalidVolumeMounts(&pod.Spec.InitContainers[rayContainerIndex], &pod)
|
|
}
|
|
|
|
var cmd, args string
|
|
if len(pod.Spec.Containers[rayContainerIndex].Command) > 0 {
|
|
cmd = convertCmdToString(pod.Spec.Containers[rayContainerIndex].Command)
|
|
}
|
|
if len(pod.Spec.Containers[rayContainerIndex].Args) > 0 {
|
|
cmd += convertCmdToString(pod.Spec.Containers[rayContainerIndex].Args)
|
|
}
|
|
if !strings.Contains(cmd, "ray start") {
|
|
cont := concatenateContainerCommand(rayNodeType, rayStartParams, pod.Spec.Containers[rayContainerIndex].Resources)
|
|
// replacing the old command
|
|
pod.Spec.Containers[rayContainerIndex].Command = []string{"/bin/bash", "-lc", "--"}
|
|
if cmd != "" {
|
|
// If 'ray start' has --block specified, commands after it will not get executed.
|
|
// so we need to put cmd before cont.
|
|
args = fmt.Sprintf("%s && %s", cmd, cont)
|
|
} else {
|
|
args = cont
|
|
}
|
|
|
|
if !isRayStartWithBlock(rayStartParams) {
|
|
// sleep infinity is used to keep the pod `running` after the last command exits, and not go into `completed` state
|
|
args = args + " && sleep infinity"
|
|
}
|
|
|
|
pod.Spec.Containers[rayContainerIndex].Args = []string{args}
|
|
}
|
|
|
|
for index := range pod.Spec.InitContainers {
|
|
setInitContainerEnvVars(&pod.Spec.InitContainers[index], fqdnRayIP)
|
|
}
|
|
|
|
setContainerEnvVars(&pod, rayContainerIndex, rayNodeType, rayStartParams, fqdnRayIP, headPort, creator)
|
|
|
|
// health check only if FT enabled
|
|
if podTemplateSpec.Annotations != nil {
|
|
if enabledString, ok := podTemplateSpec.Annotations[RayFTEnabledAnnotationKey]; ok {
|
|
if strings.ToLower(enabledString) == "true" {
|
|
// Ray FT is enabled and we need to add health checks
|
|
if pod.Spec.Containers[rayContainerIndex].ReadinessProbe == nil {
|
|
// it is possible that some user have the probe parameters to override the default,
|
|
// in this case, this if condition is skipped
|
|
probe := &v1.Probe{
|
|
InitialDelaySeconds: DefaultReadinessProbeInitialDelaySeconds,
|
|
TimeoutSeconds: DefaultReadinessProbeTimeoutSeconds,
|
|
PeriodSeconds: DefaultReadinessProbePeriodSeconds,
|
|
SuccessThreshold: DefaultReadinessProbeSuccessThreshold,
|
|
FailureThreshold: DefaultReadinessProbeFailureThreshold,
|
|
}
|
|
pod.Spec.Containers[rayContainerIndex].ReadinessProbe = probe
|
|
}
|
|
// add readiness probe exec command in case missing.
|
|
initReadinessProbeHandler(pod.Spec.Containers[rayContainerIndex].ReadinessProbe, rayNodeType)
|
|
|
|
if pod.Spec.Containers[rayContainerIndex].LivenessProbe == nil {
|
|
// it is possible that some user have the probe parameters to override the default,
|
|
// in this case, this if condition is skipped
|
|
probe := &v1.Probe{
|
|
InitialDelaySeconds: DefaultLivenessProbeInitialDelaySeconds,
|
|
TimeoutSeconds: DefaultLivenessProbeTimeoutSeconds,
|
|
PeriodSeconds: DefaultLivenessProbePeriodSeconds,
|
|
SuccessThreshold: DefaultLivenessProbeSuccessThreshold,
|
|
FailureThreshold: DefaultLivenessProbeFailureThreshold,
|
|
}
|
|
pod.Spec.Containers[rayContainerIndex].LivenessProbe = probe
|
|
}
|
|
// add liveness probe exec command in case missing
|
|
initLivenessProbeHandler(pod.Spec.Containers[rayContainerIndex].LivenessProbe, rayNodeType)
|
|
}
|
|
}
|
|
}
|
|
|
|
return pod
|
|
}
|
|
|
|
// BuildAutoscalerContainer builds a Ray autoscaler container which can be appended to the head pod.
|
|
func BuildAutoscalerContainer(autoscalerImage string) v1.Container {
|
|
container := v1.Container{
|
|
Name: AutoscalerContainerName,
|
|
Image: autoscalerImage,
|
|
ImagePullPolicy: v1.PullIfNotPresent,
|
|
Env: []v1.EnvVar{
|
|
{
|
|
Name: RAY_CLUSTER_NAME,
|
|
ValueFrom: &v1.EnvVarSource{
|
|
FieldRef: &v1.ObjectFieldSelector{
|
|
FieldPath: fmt.Sprintf("metadata.labels['%s']", RayClusterLabelKey),
|
|
},
|
|
},
|
|
},
|
|
{
|
|
Name: "RAY_CLUSTER_NAMESPACE",
|
|
ValueFrom: &v1.EnvVarSource{
|
|
FieldRef: &v1.ObjectFieldSelector{
|
|
FieldPath: "metadata.namespace",
|
|
},
|
|
},
|
|
},
|
|
{
|
|
Name: "RAY_HEAD_POD_NAME",
|
|
ValueFrom: &v1.EnvVarSource{
|
|
FieldRef: &v1.ObjectFieldSelector{
|
|
FieldPath: "metadata.name",
|
|
},
|
|
},
|
|
},
|
|
},
|
|
Command: []string{
|
|
"ray",
|
|
},
|
|
Args: []string{
|
|
"kuberay-autoscaler",
|
|
"--cluster-name",
|
|
"$(RAY_CLUSTER_NAME)",
|
|
"--cluster-namespace",
|
|
"$(RAY_CLUSTER_NAMESPACE)",
|
|
},
|
|
Resources: v1.ResourceRequirements{
|
|
Limits: v1.ResourceList{
|
|
v1.ResourceCPU: resource.MustParse("500m"),
|
|
v1.ResourceMemory: resource.MustParse("512Mi"),
|
|
},
|
|
Requests: v1.ResourceList{
|
|
v1.ResourceCPU: resource.MustParse("500m"),
|
|
v1.ResourceMemory: resource.MustParse("512Mi"),
|
|
},
|
|
},
|
|
}
|
|
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
|
|
}
|
|
if autoscalerOptions.ImagePullPolicy != nil {
|
|
autoscalerContainer.ImagePullPolicy = *autoscalerOptions.ImagePullPolicy
|
|
}
|
|
if len(autoscalerOptions.Env) > 0 {
|
|
autoscalerContainer.Env = append(autoscalerContainer.Env, autoscalerOptions.Env...)
|
|
}
|
|
if len(autoscalerOptions.EnvFrom) > 0 {
|
|
autoscalerContainer.EnvFrom = append(autoscalerContainer.EnvFrom, autoscalerOptions.EnvFrom...)
|
|
}
|
|
if autoscalerOptions.SecurityContext != nil {
|
|
autoscalerContainer.SecurityContext = autoscalerOptions.SecurityContext.DeepCopy()
|
|
}
|
|
}
|
|
}
|
|
|
|
func isRayStartWithBlock(rayStartParams map[string]string) bool {
|
|
if blockValue, exist := rayStartParams["block"]; exist {
|
|
return strings.ToLower(blockValue) == "true"
|
|
}
|
|
return false
|
|
}
|
|
|
|
func convertCmdToString(cmdArr []string) (cmd string) {
|
|
cmdAggr := new(bytes.Buffer)
|
|
for _, v := range cmdArr {
|
|
fmt.Fprintf(cmdAggr, " %s ", v)
|
|
}
|
|
return cmdAggr.String()
|
|
}
|
|
|
|
func getRayContainerIndex(podSpec v1.PodSpec) (rayContainerIndex int) {
|
|
// a ray pod can have multiple containers.
|
|
// we identify the ray container based on env var: RAY=true
|
|
// if the env var is missing, we choose containers[0].
|
|
for i, container := range podSpec.Containers {
|
|
for _, env := range container.Env {
|
|
if env.Name == strings.ToLower("ray") && env.Value == strings.ToLower("true") {
|
|
log.Info("Head pod container with index " + strconv.Itoa(i) + " identified as Ray container based on env RAY=true.")
|
|
return i
|
|
}
|
|
}
|
|
}
|
|
// not found, use first container
|
|
log.Info("Head pod container with index 0 identified as Ray container.")
|
|
return 0
|
|
}
|
|
|
|
func getAutoscalerContainerIndex(pod v1.Pod) (autoscalerContainerIndex int) {
|
|
// we identify the autoscaler container based on its name
|
|
for i, container := range pod.Spec.Containers {
|
|
if container.Name == AutoscalerContainerName {
|
|
return i
|
|
}
|
|
}
|
|
|
|
// This should be unreachable.
|
|
panic("Autoscaler container not found!")
|
|
}
|
|
|
|
// labelPod returns the labels for selecting the resources
|
|
// belonging to the given RayCluster CR name.
|
|
func labelPod(rayNodeType rayiov1alpha1.RayNodeType, rayClusterName string, groupName string, labels map[string]string) (ret map[string]string) {
|
|
if labels == nil {
|
|
labels = make(map[string]string)
|
|
}
|
|
|
|
ret = map[string]string{
|
|
RayNodeLabelKey: "yes",
|
|
RayClusterLabelKey: rayClusterName,
|
|
RayNodeTypeLabelKey: string(rayNodeType),
|
|
RayNodeGroupLabelKey: groupName,
|
|
RayIDLabelKey: utils.CheckLabel(utils.GenerateIdentifier(rayClusterName, rayNodeType)),
|
|
KubernetesApplicationNameLabelKey: ApplicationName,
|
|
KubernetesCreatedByLabelKey: ComponentName,
|
|
RayClusterDashboardServiceLabelKey: utils.GenerateDashboardAgentLabel(rayClusterName),
|
|
}
|
|
|
|
for k, v := range ret {
|
|
if k == string(rayNodeType) {
|
|
// overriding invalid values for this label
|
|
if v != string(rayiov1alpha1.HeadNode) && v != string(rayiov1alpha1.WorkerNode) {
|
|
labels[k] = v
|
|
}
|
|
}
|
|
if k == RayNodeGroupLabelKey {
|
|
// overriding invalid values for this label
|
|
if v != groupName {
|
|
labels[k] = v
|
|
}
|
|
}
|
|
if _, ok := labels[k]; !ok {
|
|
labels[k] = v
|
|
}
|
|
}
|
|
|
|
return labels
|
|
}
|
|
|
|
func setInitContainerEnvVars(container *v1.Container, fqdnRayIP string) {
|
|
if container.Env == nil || len(container.Env) == 0 {
|
|
container.Env = []v1.EnvVar{}
|
|
}
|
|
// Init containers in both head and worker require FQ_RAY_IP.
|
|
// (1) The head needs FQ_RAY_IP to create a self-signed certificate for its TLS authenticate.
|
|
// (2) The worker needs FQ_RAY_IP to establish a connection with the Ray head.
|
|
container.Env = append(container.Env,
|
|
v1.EnvVar{Name: FQ_RAY_IP, Value: fqdnRayIP},
|
|
// RAY_IP is deprecated and should be kept for backward compatibility purposes only.
|
|
v1.EnvVar{Name: RAY_IP, Value: utils.ExtractRayIPFromFQDN(fqdnRayIP)},
|
|
)
|
|
}
|
|
|
|
func setContainerEnvVars(pod *v1.Pod, rayContainerIndex int, rayNodeType rayiov1alpha1.RayNodeType, rayStartParams map[string]string, fqdnRayIP string, headPort string, creator string) {
|
|
// TODO: Audit all environment variables to identify which should not be modified by users.
|
|
container := &pod.Spec.Containers[rayContainerIndex]
|
|
if container.Env == nil || len(container.Env) == 0 {
|
|
container.Env = []v1.EnvVar{}
|
|
}
|
|
|
|
// case 1: head => Use LOCAL_HOST
|
|
// case 2: worker => Use fqdnRayIP (fully qualified domain name)
|
|
ip := LOCAL_HOST
|
|
if rayNodeType == rayiov1alpha1.WorkerNode {
|
|
ip = fqdnRayIP
|
|
container.Env = append(container.Env,
|
|
v1.EnvVar{Name: FQ_RAY_IP, Value: ip},
|
|
// RAY_IP is deprecated and should be kept for backward compatibility purposes only.
|
|
v1.EnvVar{Name: RAY_IP, Value: utils.ExtractRayIPFromFQDN(ip)},
|
|
)
|
|
}
|
|
|
|
// The RAY_CLUSTER_NAME environment variable is managed by KubeRay and should not be set by the user.
|
|
clusterNameEnv := v1.EnvVar{
|
|
Name: RAY_CLUSTER_NAME,
|
|
ValueFrom: &v1.EnvVarSource{
|
|
FieldRef: &v1.ObjectFieldSelector{
|
|
FieldPath: fmt.Sprintf("metadata.labels['%s']", RayClusterLabelKey),
|
|
},
|
|
},
|
|
}
|
|
container.Env = append(container.Env, clusterNameEnv)
|
|
|
|
if !envVarExists(RAY_PORT, container.Env) {
|
|
portEnv := v1.EnvVar{Name: RAY_PORT, Value: headPort}
|
|
container.Env = append(container.Env, portEnv)
|
|
}
|
|
|
|
if strings.ToLower(creator) == RayServiceCreatorLabelValue {
|
|
// Only add this env for Ray Service cluster to improve service SLA.
|
|
if !envVarExists(RAY_TIMEOUT_MS_TASK_WAIT_FOR_DEATH_INFO, container.Env) {
|
|
deathEnv := v1.EnvVar{Name: RAY_TIMEOUT_MS_TASK_WAIT_FOR_DEATH_INFO, Value: "0"}
|
|
container.Env = append(container.Env, deathEnv)
|
|
}
|
|
if !envVarExists(RAY_GCS_SERVER_REQUEST_TIMEOUT_SECONDS, container.Env) {
|
|
gcsTimeoutEnv := v1.EnvVar{Name: RAY_GCS_SERVER_REQUEST_TIMEOUT_SECONDS, Value: "5"}
|
|
container.Env = append(container.Env, gcsTimeoutEnv)
|
|
}
|
|
if !envVarExists(RAY_SERVE_KV_TIMEOUT_S, container.Env) {
|
|
serveKvTimeoutEnv := v1.EnvVar{Name: RAY_SERVE_KV_TIMEOUT_S, Value: "5"}
|
|
container.Env = append(container.Env, serveKvTimeoutEnv)
|
|
}
|
|
if !envVarExists(SERVE_CONTROLLER_PIN_ON_NODE, container.Env) {
|
|
servePinOnNode := v1.EnvVar{Name: SERVE_CONTROLLER_PIN_ON_NODE, Value: "0"}
|
|
container.Env = append(container.Env, servePinOnNode)
|
|
}
|
|
}
|
|
// Setting the RAY_ADDRESS env allows connecting to Ray using ray.init() when connecting
|
|
// from within the cluster.
|
|
if !envVarExists(RAY_ADDRESS, container.Env) {
|
|
rayAddress := fmt.Sprintf("%s:%s", ip, headPort)
|
|
addressEnv := v1.EnvVar{Name: RAY_ADDRESS, Value: rayAddress}
|
|
container.Env = append(container.Env, addressEnv)
|
|
}
|
|
if !envVarExists(RAY_USAGE_STATS_KUBERAY_IN_USE, container.Env) {
|
|
usageEnv := v1.EnvVar{Name: RAY_USAGE_STATS_KUBERAY_IN_USE, Value: "1"}
|
|
container.Env = append(container.Env, usageEnv)
|
|
}
|
|
if !envVarExists(REDIS_PASSWORD, container.Env) {
|
|
// setting the REDIS_PASSWORD env var from the params
|
|
port := v1.EnvVar{Name: REDIS_PASSWORD}
|
|
if value, ok := rayStartParams["redis-password"]; ok {
|
|
port.Value = value
|
|
}
|
|
container.Env = append(container.Env, port)
|
|
}
|
|
if !envVarExists(RAY_EXTERNAL_STORAGE_NS, container.Env) {
|
|
// setting the RAY_EXTERNAL_STORAGE_NS env var from the params
|
|
if pod.Annotations != nil {
|
|
if v, ok := pod.Annotations[RayExternalStorageNSAnnotationKey]; ok {
|
|
storageNS := v1.EnvVar{Name: RAY_EXTERNAL_STORAGE_NS, Value: v}
|
|
container.Env = append(container.Env, storageNS)
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
func envVarExists(envName string, envVars []v1.EnvVar) bool {
|
|
if len(envVars) == 0 {
|
|
return false
|
|
}
|
|
|
|
for _, env := range envVars {
|
|
if env.Name == envName {
|
|
return true
|
|
}
|
|
}
|
|
return false
|
|
}
|
|
|
|
// TODO auto complete params
|
|
func setMissingRayStartParams(rayStartParams map[string]string, nodeType rayiov1alpha1.RayNodeType, headPort string, fqdnRayIP string) (completeStartParams map[string]string) {
|
|
// Note: The argument headPort is unused for nodeType == rayiov1alpha1.HeadNode.
|
|
if nodeType == rayiov1alpha1.WorkerNode {
|
|
if _, ok := rayStartParams["address"]; !ok {
|
|
address := fmt.Sprintf("%s:%s", fqdnRayIP, headPort)
|
|
rayStartParams["address"] = address
|
|
}
|
|
}
|
|
|
|
// add metrics port for expose the metrics to the prometheus.
|
|
if _, ok := rayStartParams["metrics-export-port"]; !ok {
|
|
rayStartParams["metrics-export-port"] = fmt.Sprint(DefaultMetricsPort)
|
|
}
|
|
|
|
// add --block option. See https://github.com/ray-project/kuberay/pull/675
|
|
if _, ok := rayStartParams["block"]; !ok {
|
|
rayStartParams["block"] = "true"
|
|
}
|
|
|
|
return rayStartParams
|
|
}
|
|
|
|
func setAgentListPortStartParams(instance rayiov1alpha1.RayCluster, rayStartParams map[string]string) (completeStartParams map[string]string) {
|
|
// add dashboard listen port for serve endpoints to RayService.
|
|
if _, ok := rayStartParams["dashboard-agent-listen-port"]; !ok {
|
|
if value, ok := instance.Annotations[EnableAgentServiceKey]; ok && value == EnableAgentServiceTrue {
|
|
rayStartParams["dashboard-agent-listen-port"] = strconv.Itoa(DefaultDashboardAgentListenPort)
|
|
}
|
|
}
|
|
|
|
return rayStartParams
|
|
}
|
|
|
|
// concatenateContainerCommand with ray start
|
|
func concatenateContainerCommand(nodeType rayiov1alpha1.RayNodeType, rayStartParams map[string]string, resource v1.ResourceRequirements) (fullCmd string) {
|
|
if _, ok := rayStartParams["num-cpus"]; !ok {
|
|
cpu := resource.Limits[v1.ResourceCPU]
|
|
if !cpu.IsZero() {
|
|
rayStartParams["num-cpus"] = strconv.FormatInt(cpu.Value(), 10)
|
|
}
|
|
}
|
|
|
|
if _, ok := rayStartParams["memory"]; !ok {
|
|
memory := resource.Limits[v1.ResourceMemory]
|
|
if !memory.IsZero() {
|
|
rayStartParams["memory"] = strconv.FormatInt(memory.Value(), 10)
|
|
}
|
|
}
|
|
|
|
if _, ok := rayStartParams["num-gpus"]; !ok {
|
|
// Scan for resource keys ending with "gpu" like "nvidia.com/gpu".
|
|
for resourceKey, resource := range resource.Limits {
|
|
if strings.HasSuffix(string(resourceKey), "gpu") && !resource.IsZero() {
|
|
rayStartParams["num-gpus"] = strconv.FormatInt(resource.Value(), 10)
|
|
// For now, only support one GPU type. Break on first match.
|
|
break
|
|
}
|
|
}
|
|
}
|
|
|
|
log.V(10).Info("concatenate container command", "ray start params", rayStartParams)
|
|
|
|
switch nodeType {
|
|
case rayiov1alpha1.HeadNode:
|
|
return fmt.Sprintf("ulimit -n 65536; ray start --head %s", convertParamMap(rayStartParams))
|
|
case rayiov1alpha1.WorkerNode:
|
|
return fmt.Sprintf("ulimit -n 65536; ray start %s", convertParamMap(rayStartParams))
|
|
default:
|
|
log.Error(fmt.Errorf("missing node type"), "a node must be either head or worker")
|
|
}
|
|
return ""
|
|
}
|
|
|
|
func convertParamMap(rayStartParams map[string]string) (s string) {
|
|
flags := new(bytes.Buffer)
|
|
// specialParameterOptions' arguments can be true or false.
|
|
// For example, --log-color can be auto | false | true.
|
|
specialParameterOptions := []string{"log-color", "include-dashboard"}
|
|
for option, argument := range rayStartParams {
|
|
if utils.Contains([]string{"true", "false"}, strings.ToLower(argument)) && !utils.Contains(specialParameterOptions, option) {
|
|
// booleanOptions: do not require any argument. Essentially represent boolean on-off switches.
|
|
if strings.ToLower(argument) == "true" {
|
|
fmt.Fprintf(flags, " --%s ", option)
|
|
}
|
|
} else {
|
|
// parameterOption: require arguments to be provided along with the option.
|
|
fmt.Fprintf(flags, " --%s=%s ", option, argument)
|
|
}
|
|
}
|
|
return flags.String()
|
|
}
|
|
|
|
// addEmptyDir adds an emptyDir volume to the pod and a corresponding volume mount to the container
|
|
// Used for a /dev/shm memory mount for object store and for a /tmp/ray disk mount for autoscaler logs.
|
|
func addEmptyDir(container *v1.Container, pod *v1.Pod, volumeName string, volumeMountPath string, storageMedium v1.StorageMedium) {
|
|
if checkIfVolumeMounted(container, pod, volumeMountPath) {
|
|
log.Info("volume already mounted", "volume", volumeName, "path", volumeMountPath)
|
|
return
|
|
}
|
|
|
|
// 1) If needed, create a Volume of type emptyDir and add it to Volumes.
|
|
if !checkIfVolumeExists(pod, volumeName) {
|
|
emptyDirVolume := makeEmptyDirVolume(container, volumeName, storageMedium)
|
|
pod.Spec.Volumes = append(pod.Spec.Volumes, emptyDirVolume)
|
|
}
|
|
|
|
// 2) Create a VolumeMount that uses the emptyDir.
|
|
mountedVolume := v1.VolumeMount{
|
|
MountPath: volumeMountPath,
|
|
Name: volumeName,
|
|
ReadOnly: false,
|
|
}
|
|
container.VolumeMounts = append(container.VolumeMounts, mountedVolume)
|
|
}
|
|
|
|
// Format an emptyDir volume.
|
|
// When the storage medium is memory, set the size limit based on container resources.
|
|
// For other media, don't set a size limit.
|
|
func makeEmptyDirVolume(container *v1.Container, volumeName string, storageMedium v1.StorageMedium) v1.Volume {
|
|
var sizeLimit *resource.Quantity
|
|
if storageMedium == v1.StorageMediumMemory {
|
|
// If using memory, set size limit based on primary container's resources.
|
|
sizeLimit = findMemoryReqOrLimit(*container)
|
|
} else {
|
|
// Otherwise, don't set a limit.
|
|
sizeLimit = nil
|
|
}
|
|
return v1.Volume{
|
|
Name: volumeName,
|
|
VolumeSource: v1.VolumeSource{
|
|
EmptyDir: &v1.EmptyDirVolumeSource{
|
|
Medium: storageMedium,
|
|
SizeLimit: sizeLimit,
|
|
},
|
|
},
|
|
}
|
|
}
|
|
|
|
// Checks if the container has a volumeMount with the given mount path and if
|
|
// the pod has a matching Volume.
|
|
func checkIfVolumeMounted(container *v1.Container, pod *v1.Pod, volumeMountPath string) bool {
|
|
for _, mountedVol := range container.VolumeMounts {
|
|
if mountedVol.MountPath == volumeMountPath {
|
|
return true
|
|
}
|
|
}
|
|
return false
|
|
}
|
|
|
|
// Checks if a volume with the given name exists.
|
|
func checkIfVolumeExists(pod *v1.Pod, volumeName string) bool {
|
|
for _, podVolume := range pod.Spec.Volumes {
|
|
if podVolume.Name == volumeName {
|
|
return true
|
|
}
|
|
}
|
|
return false
|
|
}
|
|
|
|
func cleanupInvalidVolumeMounts(container *v1.Container, pod *v1.Pod) {
|
|
// if a volumeMount is specified in the container,
|
|
// but has no corresponding pod volume, it is removed
|
|
k := 0
|
|
for _, mountedVol := range container.VolumeMounts {
|
|
for _, podVolume := range pod.Spec.Volumes {
|
|
if mountedVol.Name == podVolume.Name {
|
|
// valid mount, moving on...
|
|
container.VolumeMounts[k] = mountedVol
|
|
k++
|
|
break
|
|
}
|
|
}
|
|
}
|
|
container.VolumeMounts = container.VolumeMounts[:k]
|
|
}
|
|
|
|
func findMemoryReqOrLimit(container v1.Container) (res *resource.Quantity) {
|
|
var mem *resource.Quantity
|
|
// check the requests, if they are not set, check the limits.
|
|
if q, ok := container.Resources.Requests[v1.ResourceMemory]; ok {
|
|
mem = &q
|
|
return mem
|
|
}
|
|
if q, ok := container.Resources.Limits[v1.ResourceMemory]; ok {
|
|
mem = &q
|
|
return mem
|
|
}
|
|
return nil
|
|
}
|
|
|
|
// ValidateHeadRayStartParams will validate the head node's RayStartParams.
|
|
// Return a bool indicating the validity of RayStartParams and an err with additional information.
|
|
// If isValid is true, RayStartParams are valid. Any errors will only affect performance.
|
|
// If isValid is false, RayStartParams are invalid will result in an unhealthy or failed Ray cluster.
|
|
func ValidateHeadRayStartParams(rayHeadGroupSpec rayiov1alpha1.HeadGroupSpec) (isValid bool, err error) {
|
|
// TODO (dxia): if you add more validation, please split checks into separate subroutines.
|
|
var objectStoreMemory int64
|
|
rayStartParams := rayHeadGroupSpec.RayStartParams
|
|
// validation for the object store memory
|
|
if objectStoreMemoryStr, ok := rayStartParams[ObjectStoreMemoryKey]; ok {
|
|
objectStoreMemory, err = strconv.ParseInt(objectStoreMemoryStr, 10, 64)
|
|
if err != nil {
|
|
isValid = false
|
|
err = errors.NewBadRequest(fmt.Sprintf("Cannot parse %s %s as an integer: %s", ObjectStoreMemoryKey, objectStoreMemoryStr, err.Error()))
|
|
return
|
|
}
|
|
for _, container := range rayHeadGroupSpec.Template.Spec.Containers {
|
|
// find the ray container.
|
|
if container.Name == RayHeadContainer {
|
|
if shmSize, ok := container.Resources.Requests.Memory().AsInt64(); ok && objectStoreMemory > shmSize {
|
|
if envVarExists(AllowSlowStorageEnvVar, container.Env) {
|
|
// in ray if this env var is set, it will only affect the performance.
|
|
isValid = true
|
|
msg := fmt.Sprintf("RayStartParams: object store memory exceeds head node container's memory request, %s:%d, memory request:%d\n"+
|
|
"This will harm performance. Consider deleting files in %s or increasing head node's memory request.", ObjectStoreMemoryKey, objectStoreMemory, shmSize, SharedMemoryVolumeMountPath)
|
|
log.Info(msg)
|
|
err = errors.NewBadRequest(msg)
|
|
return
|
|
} else {
|
|
// if not set, the head node may crash and result in an unhealthy status.
|
|
isValid = false
|
|
msg := fmt.Sprintf("RayStartParams: object store memory exceeds head node container's memory request, %s:%d, memory request:%d\n"+
|
|
"This will lead to a ValueError in Ray! Consider deleting files in %s or increasing head node's memory request.\n"+
|
|
"To ignore this warning, set the following environment variable in headGroupSpec: %s=1",
|
|
ObjectStoreMemoryKey, objectStoreMemory, shmSize, SharedMemoryVolumeMountPath, AllowSlowStorageEnvVar)
|
|
err = errors.NewBadRequest(msg)
|
|
return
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
// default return
|
|
return true, nil
|
|
}
|