Kubernetes 1.37 "Garhwal": HPA scale-to-zero and gang scheduling in beta

2026-08-26 · Publisher : Kubernetes · Domain : Cloud & DevOps

On 26 August 2026, Kubernetes 1.37 ships 67 enhancements (16 stable, 23 beta, 27 alpha): gang scheduling and HPA scale-to-zero enter beta, and the metrics.k8s.io API and pod certificates become stable.

### Background On 26 August 2026, Kubernetes releases version **1.37 "Garhwal"**, named for a Himalayan region of Uttarakhand, India: **67 enhancements**, including 16 stable, 23 beta, 27 alpha and one deprecation or removal. It continues the focus on AI workloads, with the scheduler, DRA and autoscaling. ### Key facts - **Gang scheduling** in beta: all-or-nothing scheduling of a pod group through the Workload and PodGroup APIs; workload-aware preemption also enters beta. - **Scale to zero** for the HorizontalPodAutoscaler in beta. - **metrics.k8s.io API** stable, after nearly nine years in beta; it feeds `kubectl top` and CPU and memory autoscaling. - **Pod certificates** and ClusterTrustBundles stable: the kubelet distributes keys, X.509 certificates and trust anchors to pods. - DRA: ResourceClaim device status, DRA extended resources and device taints/tolerations reach stable. ### Why it matters The release consolidates scheduling for distributed AI workloads (gang scheduling, DRA, preemption) and reduces external dependencies: native workload identity can avoid a separate component such as cert-manager in some use cases. Scale-to-zero interests teams looking to cut the cost of inference or rarely used services. As always, beta and stable do not carry the same support coverage: qualify before production. The release has 67 enhancements for a single removal or deprecation, a sign of a cycle made mostly of additions and graduations. ### Takeaway Kubernetes 1.37 consolidates AI workload scheduling and native pod identity, with scale-to-zero in beta.

Tags : Kubernetes, Cloud natif, IA, FinOps, Open Source