OCI Kubernetes Engine (OKE) support for GPU maintenance event repair

You can now configure OKE to coordinate maintenance event repair for eligible bare metal GPU worker nodes in enhanced clusters.

When the GPU Health Monitoring add-on reports an eligible critical HPC fault, OKE correlates the fault with an OCI Compute maintenance event. For nodes selected for repair, OKE makes the node unavailable for new workloads and attempts to drain workloads from the node before coordinating the action specified by the maintenance event:

• STOP: OKE coordinates processing of the maintenance event and returns the same worker node to service.
• TERMINATE: OKE terminates the affected Compute instance. For nodes in managed node pools, normal node pool capacity management creates a replacement node.

You can monitor the correlated maintenance event and repair status by using the node’s NodeHealthReport resource and Kubernetes node conditions.

For more information, see Repairing GPU Hosts.