News › linux
By Zayden R., September 11, 2026
Kubernetes 1.37 now includes scheduler preemption for in-place Pod resizing, addressing the challenge of deferred resource requests on fully-utilized nodes. This feature, in alpha, helps prioritize critical workloads without manual intervention.
Kubernetes 1.37 brings a noteworthy advancement with the introduction of scheduler preemption for in-place Pod resizing, a feature currently in alpha. This update addresses a long-standing issue where resource requests for running Pods could be indefinitely deferred if the host node lacked sufficient allocatable capacity. With the new preemption mechanism, the Kubernetes scheduler can actively prioritize higher-priority workloads by freeing up resources on heavily-utilized nodes, preempting lower-priority tasks when necessary.
The need for this feature arose with the general availability of the core in-place Pod resize capability in Kubernetes 1.35, which allowed dynamic adjustments to CPU and memory allocations without needing disruptive restarts. However, without preemption, resize requests could remain stuck in a deferred state, posing challenges for critical applications needing immediate resource scale-ups.
By enabling the InPlacePodVerticalScalingSchedulerPreemption feature gate, cluster administrators can now ensure that essential services like in-memory databases or real-time web servers receive the resources they need promptly. This is particularly important in scenarios where an out-of-memory crash is imminent, and manual intervention to evict lower-priority Pods is no longer required.
Additionally, Kubernetes 1.37 introduces gang scheduling in beta, aiming to reduce GPU idle costs by default, further enhancing the platform's efficiency in resource management. These features together reflect Kubernetes' commitment to refining resource allocation and scheduling, making it a more robust choice for managing complex, dynamic workloads.