View a markdown version of this page

Quota and scheduling behavior for Ray workloads - Amazon SageMaker AI

Quota and scheduling behavior for Ray workloads

Task Governance accounts quota at RayCluster granularity. A cluster reserves quota for its full declared size when it is admitted, and holds that quota for its whole lifetime regardless of load. A long-lived cluster that sits idle still counts against your team quota until you delete it.

Gang scheduling

Task Governance admits the head and all workers together. A RayCluster is not admitted until quota exists for the entire declared size, so a partially scheduled cluster never starts.

Preemption

When higher-priority work needs capacity, Task Governance preempts a lower-priority Ray workload. Preemption deletes the whole cluster, including the head.