Managing Ray workloads with Studio
Amazon SageMaker Studio manages Ray workloads from a web interface. Open your cluster in the
HyperPod console, then choose the Tasks tab. The
tab lists the Ray resources in the namespaces you have access to, and the Task type filter switches between RayCluster,
RayJob, RayCronJob, and RayService.
Studio reads and writes the same KubeRay custom resources that the operator
reconciles. A resource you create in Studio is identical to one you apply with
kubectl. You can move between the two surfaces on the same workload.
Before anyone uses the Tasks tab, configure a SageMaker AI domain and grant it access to your cluster. For more information, see Setting up Studio for Ray.
Creating a cluster
Choose Create Ray cluster and either complete the form
or paste a RayCluster manifest into the YAML view. The form covers the
common fields. The YAML view accepts any field the KubeRay operator supports. Use it when
you need a field the form does not expose.
The namespace selector in the form shows the compute allocation and current utilization for each namespace you have access to. Use it to size the cluster against the capacity that is actually free before you submit, instead of finding out afterward that the cluster is waiting on quota. This matters most in a namespace that HyperPod Task Governance manages, where the cluster is not admitted until quota exists for its full declared size. For more information, see Quota and scheduling behavior for Ray workloads.
Attach a space to the cluster when you want to develop against it interactively. A
space is a JupyterLab or Code Editor environment in Amazon SageMaker Studio. Code you run there,
in a notebook or a terminal, reaches the cluster through ray.init().
A space carries its own Ray version, from the SageMaker AI Distribution image it runs. Match that version to the Ray version of the cluster. A mismatch produces runtime errors that are hard to diagnose. For more information, see Attaching Ray cluster to Space.
Actions on a cluster
Choose Actions on a row to act on that resource without leaving Studio.
| Action | What it does |
|---|---|
| Interactive development | Opens a JupyterLab or Code Editor space attached to the cluster, so
ray.init() in that space connects to it. Requires the
SageMaker AI Spaces add-on. For more information, see Attaching Ray cluster to Space. |
| Submit job | Submits a job to the running cluster. |
| Open Ray Dashboard | Opens the Ray Dashboard for the cluster. Requires the Ray Endpoint Operator. For more information, see Installing the HyperPod Ray Endpoint Operator. |
| Open Grafana | Opens the Grafana dashboards for the cluster. Requires the observability add-on. For more information, see Setting up Ray metrics collection. |
| Edit Ray cluster | Changes the cluster, including worker replica counts. |
| Suspend Ray cluster | Stops the cluster's pods and keeps the resource. Releases the compute. |
| Resume Ray cluster | Restarts the pods of a suspended cluster with its existing configuration. Available when the cluster is suspended. |
| Delete Ray cluster | Deletes the cluster and its pods. |
Viewing details and events
Choose a resource name to open it. The Events tab lists the pods of the Ray cluster together with their Kubernetes events. Start there when a cluster does not reach a running state. It also helps when a pod restarts unexpectedly, or a job fails for a reason its logs do not explain. The events name the failing pod and the reason, such as an image pull failure or insufficient allocatable capacity. For more information, see Troubleshooting Ray on HyperPod.
To see utilization for the cluster, choose Actions, then Open Grafana. For quota, priority, and preemption behavior in a namespace that HyperPod Task Governance manages, see Quota and scheduling behavior for Ray workloads.
