feat(helm/raySpec): allow customizing the Ray head node readiness probe

The multi-node Ray head pod hard-codes its readiness probe to
failureThreshold: 1 / periodSeconds: 10. On slow, large multi-node
deployments the head serving process can take longer than a single 10s
period to start answering /health, so the pod is marked unready after the
first failed probe and never receives traffic. Users had no way to relax
it -- setting readinessProbe under servingEngineSpec, modelSpec, or
raySpec.headNode had no effect on the head node.

Honor raySpec.headNode.readinessProbe when set (rendered verbatim via
toYaml), falling back to the existing default otherwise, so behavior is
unchanged unless the field is provided. Document the new field in
values.yaml and add it to values.schema.json.

Fixes #742.

Signed-off-by: Anai-Guo <antai12232931@outlook.com>

Author: Anai-Guo

helm/templates/ray-cluster.yaml

@@ -167,11 +167,15 @@ spec:
               - name: {{ include "chart.container-port-name" . }}
                 containerPort: {{ include "chart.container-port" . }}
             readinessProbe:
+              {{- if $modelSpec.raySpec.headNode.readinessProbe }}
+              {{- toYaml $modelSpec.raySpec.headNode.readinessProbe | nindent 14 }}
+              {{- else }}
               httpGet:
                 path: /health
                 port: {{ include "chart.container-port" . }}
               failureThreshold: 1
               periodSeconds: 10
+              {{- end }}
             livenessProbe:
               exec:
                 command: ["/bin/bash", "-c", "echo TBD"]

helm/values.schema.json

@@ -1454,6 +1454,10 @@
                                             "requestMemory": {
                                                 "description": "Memory request for the head node",
                                                 "type": "string"
+                                            },
+                                            "readinessProbe": {
+                                                "description": "Readiness probe configuration for the Ray head node. If set, overrides the built-in default (failureThreshold: 1, periodSeconds: 10). Accepts a standard Kubernetes probe spec.",
+                                                "type": "object"
                                             }
                                         }
                                     }

helm/values.yaml

@@ -275,6 +275,18 @@ servingEngineSpec:
         requestMemory: "1Gi"
         # -- GPU request for the head node
         requestGPU: 1
+        # Readiness probe for the Ray head node (raySpec.headNode.readinessProbe).
+        # If unset, defaults to an httpGet /health probe with failureThreshold 1
+        # and periodSeconds 10. Override it to relax the probe when the model
+        # takes a long time to come up on a multi-node setup, otherwise the head
+        # pod may be marked unready and never receive traffic. Accepts a standard
+        # Kubernetes probe spec, e.g.:
+        # readinessProbe:
+        #   httpGet:
+        #     path: /health
+        #     port: 8000
+        #   failureThreshold: 10
+        #   periodSeconds: 30
 
   # -- Container port
   containerPort: 8000