Merge remote-tracking branch 'upstream/master' into move-unmarshall-envyaml

Author: owenowenisme

apiserver/pkg/server/ray_job_submission_service_server.go

@@ -20,6 +20,7 @@ import (
 	api "github.com/ray-project/kuberay/proto/go_client"
 	rayv1 "github.com/ray-project/kuberay/ray-operator/apis/ray/v1"
 	"github.com/ray-project/kuberay/ray-operator/controllers/ray/utils"
+	utiltypes "github.com/ray-project/kuberay/ray-operator/controllers/ray/utils/types"
 )
 
 type RayJobSubmissionServiceServerOptions struct {
@@ -54,7 +55,7 @@ func (s *RayJobSubmissionServiceServer) SubmitRayJob(ctx context.Context, req *a
 	if err != nil {
 		return nil, err
 	}
-	request := &utils.RayJobRequest{Entrypoint: req.Jobsubmission.Entrypoint}
+	request := &utiltypes.RayJobRequest{Entrypoint: req.Jobsubmission.Entrypoint}
 	if req.Jobsubmission.SubmissionId != "" {
 		request.SubmissionId = req.Jobsubmission.SubmissionId
 	}
@@ -89,7 +90,7 @@ func (s *RayJobSubmissionServiceServer) SubmitRayJob(ctx context.Context, req *a
 		}
 	}
 
-	sid, err := rayDashboardClient.SubmitJobReq(ctx, request, nil)
+	sid, err := rayDashboardClient.SubmitJobReq(ctx, request)
 	if err != nil {
 		return nil, err
 	}
@@ -216,7 +217,7 @@ func (s *RayJobSubmissionServiceServer) getRayClusterURL(ctx context.Context, re
 }
 
 // Internal method to convert RayJobInfo to JobSubmissionInfo
-func convertNodeInfo(info *utils.RayJobInfo) *api.JobSubmissionInfo {
+func convertNodeInfo(info *utiltypes.RayJobInfo) *api.JobSubmissionInfo {
 	jsi := api.JobSubmissionInfo{
 		Entrypoint:   info.Entrypoint,
 		JobId:        info.JobId,

apiserver/pkg/server/ray_job_submission_service_server_test.go

@@ -17,7 +17,6 @@ import (
 	"github.com/ray-project/kuberay/apiserver/pkg/util"
 	api "github.com/ray-project/kuberay/proto/go_client"
 	rayv1 "github.com/ray-project/kuberay/ray-operator/apis/ray/v1"
-	"github.com/ray-project/kuberay/ray-operator/controllers/ray/utils"
 	utiltypes "github.com/ray-project/kuberay/ray-operator/controllers/ray/utils/types"
 	fakeclientset "github.com/ray-project/kuberay/ray-operator/pkg/client/clientset/versioned/fake"
 )
@@ -189,7 +188,7 @@ func TestConvertNodeInfo(t *testing.T) {
 		"pip":         "[numpy pandas]",
 	}
 
-	rayJobInfo := utils.RayJobInfo{
+	rayJobInfo := utiltypes.RayJobInfo{
 		Entrypoint:   entrypoint,
 		JobId:        jobID,
 		SubmissionId: submissionID,

docs/reference/api.md

@@ -246,7 +246,7 @@ _Appears in:_
 | `entrypoint` _string_ | Entrypoint represents the command to start execution. |  |  |
 | `runtimeEnvYAML` _string_ | RuntimeEnvYAML represents the runtime environment configuration<br />provided as a multi-line YAML string. |  |  |
 | `jobId` _string_ | If jobId is not set, a new jobId will be auto-generated. |  |  |
-| `submissionMode` _[JobSubmissionMode](#jobsubmissionmode)_ | SubmissionMode specifies how RayJob submits the Ray job to the RayCluster.<br />In "K8sJobMode", the KubeRay operator creates a submitter Kubernetes Job to submit the Ray job.<br />In "HTTPMode", the KubeRay operator sends a request to the RayCluster to create a Ray job.<br />In "InteractiveMode", the KubeRay operator waits for a user to submit a job to the Ray cluster. | K8sJobMode |  |
+| `submissionMode` _[JobSubmissionMode](#jobsubmissionmode)_ | SubmissionMode specifies how RayJob submits the Ray job to the RayCluster.<br />In "K8sJobMode", the KubeRay operator creates a submitter Kubernetes Job to submit the Ray job.<br />In "HTTPMode", the KubeRay operator sends a request to the RayCluster to create a Ray job.<br />In "InteractiveMode", the KubeRay operator waits for a user to submit a job to the Ray cluster.<br />In "SidecarMode", the KubeRay operator injects a container into the Ray head Pod that acts as the job submitter to submit the Ray job. | K8sJobMode |  |
 | `entrypointResources` _string_ | EntrypointResources specifies the custom resources and quantities to reserve for the<br />entrypoint command. |  |  |
 | `entrypointNumCpus` _float_ | EntrypointNumCpus specifies the number of cpus to reserve for the entrypoint command. |  |  |
 | `entrypointNumGpus` _float_ | EntrypointNumGpus specifies the number of gpus to reserve for the entrypoint command. |  |  |

helm-chart/kuberay-operator/README.md

@@ -174,6 +174,9 @@ spec:
 | operatorCommand | string | `"/manager"` | Path to the operator binary |
 | leaderElectionEnabled | bool | `true` | If leaderElectionEnabled is set to true, the KubeRay operator will use leader election for high availability. |
 | reconcileConcurrency | int | `1` | The maximum number of reconcile operations that can be performed simultaneously. This setting controls the concurrency of the controller reconciliation loops. Higher values can improve throughput in clusters with many resources, but may increase resource consumption. |
+| kubeClient | object | `{"burst":200,"qps":100}` | Kube Client configuration for QPS and burst settings. This setting controls the QPS and burst rate of the kube client when sending requests to the Kubernetes API server. If the QPS and burst values are too low, we may easily hit rate limits on the API server and slow down the controller reconciliation loops. |
+| kubeClient.qps | float | `100` | The QPS value for the client communicating with the Kubernetes API server. Must be a float number. |
+| kubeClient.burst | int | `200` | The maximum burst for throttling requests from this client to the Kubernetes API server. Must be a non-negative integer. |
 | rbacEnable | bool | `true` | If rbacEnable is set to false, no RBAC resources will be created, including the Role for leader election, the Role for Pods and Services, and so on. |
 | crNamespacedRbacEnable | bool | `true` | When crNamespacedRbacEnable is set to true, the KubeRay operator will create a Role for RayCluster preparation (e.g., Pods, Services) and a corresponding RoleBinding for each namespace listed in the "watchNamespace" parameter. Please note that even if crNamespacedRbacEnable is set to false, the Role and RoleBinding for leader election will still be created.  Note: (1) This variable is only effective when rbacEnable and singleNamespaceInstall are both set to true. (2) In most cases, it should be set to true, unless you are using a Kubernetes cluster managed by GitOps tools such as ArgoCD. |
 | singleNamespaceInstall | bool | `false` | When singleNamespaceInstall is true: - Install namespaced RBAC resources such as Role and RoleBinding instead of cluster-scoped ones like ClusterRole and ClusterRoleBinding so that   the chart can be installed by users with permissions restricted to a single namespace.   (Please note that this excludes the CRDs, which can only be installed at the cluster scope.) - If "watchNamespace" is not set, the KubeRay operator will, by default, only listen   to resource events within its own namespace. |

helm-chart/kuberay-operator/templates/_helpers.tpl

@@ -163,6 +163,12 @@ rules:
   - get
   - patch
   - update
+- apiGroups:
+  - ""
+  resources:
+  - pods/resize
+  verbs:
+  - patch
 - apiGroups:
   - ""
   resources:
@@ -301,12 +307,6 @@ rules:
   - list
   - update
   - watch
-- apiGroups:
-  - apiextensions.k8s.io
-  resources:
-  - customresourcedefinitions
-  verbs:
-  - get
 {{- end -}}
 {{- if or .batchSchedulerEnabled (eq .batchSchedulerName "scheduler-plugins") }}
 - apiGroups:

helm-chart/kuberay-operator/templates/deployment.yaml

@@ -115,6 +115,22 @@ spec:
             {{- end }}
             {{- $argList = append $argList (printf "--reconcile-concurrency=%v" .Values.reconcileConcurrency) -}}
             {{- end -}}
+            {{- if hasKey .Values "kubeClient" -}}
+            {{- if hasKey .Values.kubeClient "qps" -}}
+            {{- $qps := toString .Values.kubeClient.qps }}
+            {{- if not (regexMatch "^[+-]?[0-9]+(\\.[0-9]+)?$" $qps) }}
+            {{- fail (printf "values.kubeClient.qps must be a valid float number, got %q" $qps) }}
+            {{- end }}
+            {{- $argList = append $argList (printf "--qps=%v" .Values.kubeClient.qps) -}}
+            {{- end -}}
+            {{- if hasKey .Values.kubeClient "burst" -}}
+            {{- $burst := toString .Values.kubeClient.burst }}
+            {{- if not (regexMatch "^[0-9]+$" $burst) }}
+            {{- fail (printf "values.kubeClient.burst must be a non-negative integer, got %q" $burst) }}
+            {{- end }}
+            {{- $argList = append $argList (printf "--burst=%v" .Values.kubeClient.burst) -}}
+            {{- end -}}
+            {{- end -}}
             {{- (printf "\n") -}}
             {{- $argList | toYaml | indent 12 }}
           ports:

helm-chart/kuberay-operator/tests/deployment_test.yaml

@@ -275,3 +275,16 @@ tests:
       - contains:
           path: spec.template.spec.containers[?(@.name=="kuberay-operator")].args
           content: "--reconcile-concurrency=5"
+
+  - it: Should use custom kube client qps and burst when set
+    set:
+      kubeClient:
+        qps: 75.5
+        burst: 150
+    asserts:
+    - contains:
+        path: spec.template.spec.containers[?(@.name=="kuberay-operator")].args
+        content: "--qps=75.5"
+    - contains:
+        path: spec.template.spec.containers[?(@.name=="kuberay-operator")].args
+        content: "--burst=150"

helm-chart/kuberay-operator/values.yaml

@@ -121,6 +121,17 @@ leaderElectionEnabled: true
 # Higher values can improve throughput in clusters with many resources, but may increase resource consumption.
 reconcileConcurrency: 1
 
+# -- Kube Client configuration for QPS and burst settings.
+# This setting controls the QPS and burst rate of the kube client when sending requests to the Kubernetes API server.
+# If the QPS and burst values are too low, we may easily hit rate limits on the API server and slow down the controller reconciliation loops.
+kubeClient:
+  # -- The QPS value for the client communicating with the Kubernetes API server.
+  # Must be a float number.
+  qps: 100.0
+  # -- The maximum burst for throttling requests from this client to the Kubernetes API server.
+  # Must be a non-negative integer.
+  burst: 200
+
 # -- If rbacEnable is set to false, no RBAC resources will be created, including the Role for leader election, the Role for Pods and Services, and so on.
 rbacEnable: true
 

ray-operator/apis/config/v1alpha1/configuration_types.go

@@ -88,6 +88,6 @@ func (config Configuration) GetDashboardClient(mgr manager.Manager) func(rayClus
 	return utils.GetRayDashboardClientFunc(mgr, config.UseKubernetesProxy)
 }
 
-func (config Configuration) GetHttpProxyClient(mgr manager.Manager) func() utils.RayHttpProxyClientInterface {
+func (config Configuration) GetHttpProxyClient(mgr manager.Manager) func(hostIp, podNamespace, podName string, port int) utils.RayHttpProxyClientInterface {
 	return utils.GetRayHttpProxyClientFunc(mgr, config.UseKubernetesProxy)
 }

ray-operator/apis/ray/v1/rayjob_types.go

@@ -82,6 +82,7 @@ const (
 	K8sJobMode      JobSubmissionMode = "K8sJobMode"      // Submit job via Kubernetes Job
 	HTTPMode        JobSubmissionMode = "HTTPMode"        // Submit job via HTTP request
 	InteractiveMode JobSubmissionMode = "InteractiveMode" // Don't submit job in KubeRay. Instead, wait for user to submit job and provide the job submission ID.
+	SidecarMode     JobSubmissionMode = "SidecarMode"     // Submit job via a sidecar container in the Ray head Pod
 )
 
 type DeletionPolicyType string
@@ -174,6 +175,7 @@ type RayJobSpec struct {
 	// In "K8sJobMode", the KubeRay operator creates a submitter Kubernetes Job to submit the Ray job.
 	// In "HTTPMode", the KubeRay operator sends a request to the RayCluster to create a Ray job.
 	// In "InteractiveMode", the KubeRay operator waits for a user to submit a job to the Ray cluster.
+	// In "SidecarMode", the KubeRay operator injects a container into the Ray head Pod that acts as the job submitter to submit the Ray job.
 	// +kubebuilder:default:=K8sJobMode
 	// +optional
 	SubmissionMode JobSubmissionMode `json:"submissionMode,omitempty"`