Kubernetes v1.37 Release: Key Updates and Changes
Gang Scheduling (KEP-4671) — Promoted to Beta
This is the biggest news of this release. It has been promoted from 1.35 Alpha to 1.37 Beta, with 1.38 as the target for Stable.
Core Mechanism: kube-scheduler identifies pods belonging to a group and holds them until all pods reach the same stage in the scheduling/binding cycle; if not all pods reach that stage within the timeout, it halts scheduling for the entire group and releases all resources. To achieve this, a new core type called `Workload` has been introduced, and the pod spec includes an object reference to the `Workload` to which it belongs. The key change in Beta is the scheduling method itself: the Workload Scheduling Cycle has been introduced, which processes all pods belonging to a PodGroup as a single batch rather than using the existing sequential, pod-by-pod method; these pods do not enter the standard scheduling queue individually. Additionally, to support Elastic Jobs, the `minCount` field in `PodGroup` and `PodGroupTemplate` became mutable starting with version 1.37.
The concurrent promotion of workload-aware preemption (KEP-5710) is also significant. When the scheduler preempts low-priority workloads, it makes decisions at the PodGroup level rather than the individual Pod level, eliminating the waste of taking down the entire group by terminating only a portion of it.
The boundaries from a QKS perspective are also clear: KEP-4671’s “Non-Goals” explicitly state that bringing fairness or multi-workload queuing into the kube-scheduler is not the goal, and that Kueue and Volcano will continue to handle those roles. In other words, the core takes on only the “mechanism for atomically scheduling this group of pods,” while quota and queue management across tenants remains the responsibility of the upper layers.
DRA — Reaching Stable + Large-Scale Scaling
These changes are of significant practical importance for GPU infrastructure:
Device Taints/Tolerations (KEP-5055) Stable: This feature marks failed or underperforming devices to prevent them from being assigned to new workloads. It allows administrators to drain devices much like isolating nodes, and to exclude devices based on criteria such as which driver manages them. This feature is immediately applicable to managing nodes with GPU failures (excluding only GPUs that have encountered XID errors).
Group Claim Sharing (KEP-5729) Beta: This feature enables multiple pods to share a single resource claim, allowing for the execution of large-scale, multi-node jobs. This feature works in tandem with gang scheduling, and the Workload API and DRA ResourceClaim are integrated via the `DRAWorkloadResourceClaims` gate.
Network Interface Standardization (KEP-4817): This enables DRA drivers to describe connected network interfaces in a consistent manner, which is becoming increasingly important for GPU and RDMA workloads.
CompositePodGroup API (KEP-6012) Alpha: Provides a way to describe the complex scheduling requirements of heterogeneous workloads (e.g., combinations of groups with different configurations, such as a training job and a parameter server).
Watchcache Initialization Resilience (Stable):
Watchcache initialization during API server startup or recovery no longer causes traffic spikes in etcd. Instead of overloading etcd with costly `list/watch` requests, only a limited number of requests are delegated, while the rest are rejected with an HTTP 429 status code, thereby reducing the risk of control plane failures in large clusters. Custom controllers must implement exponential backoff that respects the `Retry-After` header. This change will be particularly noticeable in architectures where the control plane is operated in a consolidated manner, such as the Kamaji-hosted control plane.
HPA scale-to-zero Beta (enabled by default):
For workloads based on `object/external` metrics, this feature allows scaling down to zero pods during idle periods and restoring them when demand returns, thereby reducing costs for queue consumers, batch jobs, and GPU workloads. `minReplicas` must be set to 0; CPU/memory metric-based scaling is not supported.
kube-proxy Direction Determination:
A deprecation warning is displayed for clusters that do not specify a proxy mode (KEP-5343), and the IPVS backend has been officially deprecated, with a roadmap to disable it in 1.40 and remove it in 1.43.
This v1.37 release includes significant changes in terms of AI and GPU workload scheduling, device resource management, and the stability of large-scale cluster operations.
From a QKS perspective, it is also necessary to evaluate the feasibility of implementing key features such as gang scheduling, DRA, and HPA scale-to-zero.
Source
: Kubernetes Blog · Kubernetes v1.37 Release
, August 26, 2026.
View the official Kubernetes v1.37 release notes →