How AI workloads are reshaping Kubernetes autoscaling
The article describes how Kubernetes autoscaling is being reworked for AI workloads, where the scheduler and the cluster autoscaler often disagree about what a usable node is. It covers Dynamic Resource Allocation for topology and device sharing, and notes that DRA support in autoscalers still varies by implementation. It also says workload-aware scheduling is beta and disabled by default, and that the autoscaler does not yet take part in gang scheduling, with a SIG-Autoscaling proposal under discussion.
What changes for platform teams
Platform teams running GPU node pools should check whether their autoscaler supports DRA claims and provisioning requests, since without that a pending pod may trigger an incompatible node launch or wasted capacity. Gang scheduling for large jobs still depends on a bolt-on provisioning request path that is best effort and provider dependent.
Read the original at CNCF Blog ↗
written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules