View a markdown version of this page

Scaling a development cluster - Amazon SageMaker AI

Scaling a development cluster

You scale a Ray cluster by changing the worker count directly, or by turning on autoscaling so Ray adds and removes workers with load.

Set a fixed worker count

In Studio, edit the cluster and change the replica count for a worker group. In a manifest, set replicas on the worker group.

workerGroupSpecs: - groupName: my-workers replicas: 4

Turn on autoscaling

Turn on the autoscaling toggle when you edit the cluster in Studio, then set a minimum and maximum for each worker group. In a manifest, set enableInTreeAutoscaling: true and the per-group bounds.

spec: enableInTreeAutoscaling: true workerGroupSpecs: - groupName: my-workers minReplicas: 1 maxReplicas: 8

Ray scales each worker group between its minimum and maximum based on pending tasks. The minimum sets the idle footprint, and the maximum caps cost.