From the source
Amazon published a guide for administering SageMaker HyperPod clusters through SageMaker Unified Studio, focusing on governance across four control layers: organization, project, cluster, and workload.
- The post explains how infrastructure teams can manage cluster operations while offering approved compute to ML teams in their project workspaces.
- It covers identity, capacity, and observability policies, and recommends using IAM roles and resource policies rather than project membership alone for data access.
- The guide also details scheduling controls for fair sharing of accelerator pools, including quotas, priority classes, and preemption for Amazon EKS clusters, and native partitions and QoS for Slurm clusters.





