View a markdown version of this page

Troubleshooting Ray on HyperPod - Amazon SageMaker AI

Troubleshooting Ray on HyperPod

Use this page to diagnose common problems with Ray clusters, dashboards, jobs, and observability on HyperPod. Each entry gives the symptom, the cause, and the resolution.

Ray cluster stuck in Pending

Symptom

The head or worker pods stay Pending and the cluster never reaches Running.

Cause

The requested vCPU or memory is higher than the Kubernetes allocatable capacity on the chosen instance type. Allocatable is lower than the instance nameplate because kube-system reserves and DaemonSets consume part of each node.

Resolution

Lower the vCPU and memory requests so they fit allocatable capacity, or choose a larger instance type. Check allocatable with kubectl describe node my-node and set requests below the reported values.

ray.init() fails with a version error

Symptom

ray.init() fails, or the cluster starts but clients cannot connect.

Cause

The Ray version in the space or client image does not match spec.rayVersion in the cluster manifest.

Resolution

Set spec.rayVersion to the Ray version in the container image, and use the same Ray version in the space or client. For more information, see Installing KubeRay on HyperPod Amazon EKS.

Dashboard link returns an authorization error

Symptom

The Ray Dashboard link opens but returns an authorization error.

Cause

Under the private access strategy, the dashboard is scoped to the identity that created the connection. A different identity is denied.

Resolution

Open the dashboard as the identity that created the cluster, or use the public access strategy to grant access to anyone with connect permission in the namespace. For more information, see Access strategies and security best practices.

Ray workload never admitted

Symptom

A RayCluster or RayJob stays unadmitted and no pods are created.

Cause

The namespace has no compute allocation in HyperPod Task Governance, so Task Governance holds the workload with no quota to admit it against.

Resolution

Assign a compute allocation to the namespace, or submit to a namespace that has one. For more information, see Setting up task governance for Ray.

Grafana dashboards are empty

Symptom

The Ray Grafana dashboards load but show no data for a running cluster.

Cause

Ray metrics are not turned on in the HyperPod Observability add-on, so no Ray metrics are scraped.

Resolution

Turn on Ray metrics in the Observability add-on and confirm the add-on version supports Ray metric scraping. For more information, see Setting up Ray metrics collection.

Job submission fails to reach the endpoint

Symptom

ray job submit fails to connect, or returns an authorization error.

Cause

The job submission endpoint is unreachable, or the session token expired. A dashboard or endpoint session is valid for up to six hours.

Resolution

Confirm the endpoint address, then generate a new connection URL to refresh the token and retry. For more information, see Generating a dashboard connection URL.