Karpenter v1.15: GPU-aware provisioning, smarter node repair, and more
Karpenter v1.15 adds Dynamic Resource Allocation support for NVIDIA GPUs and EFA networking on AWS, and checks instance-level EBS volume attach limits before launching nodes. It also makes node repair a voluntary disruption that honors NodePool budgets and Pod Disruption Budgets, adds reboot as a repair action for some GPU XID errors, and adds pod deletion cost management and CEL expressions in kubelet configuration. Several of these capabilities ship as alpha or behind feature gates.
What changes for platform teams
Platform teams running GPU or EFA training fleets on EKS 1.34+ can have node sizing account for device topology before launch, and repair actions now respect disruption budgets and PDBs. Several capabilities are alpha and behind feature gates, and the upgrade guide lists breaking changes to metric labels that dashboards and alerts may depend on.
Read the original at AWS Containers Blog ↗
written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules