Amazon EKS

Karpenter v1.15.0

Karpenter v1.15.0 adds node repair as a voluntary disruption method, including reboot as a node action, terminate-first repair for capacity-constrained NodePools, and a --legacy-node-repair flag to run the previous controller. It also adds metrics such as karpenter_pods_disrupted_total and labels on disruption metrics, plus AWS feature gates and CEL expressions for kubelet settings. The release notes list 41 additions and 35 fixes, with 189 maintenance lines omitted.

platform teams Platform teams running Karpenter on EKS should review repair and disruption behavior, since node repair now runs as a voluntary disruption and default repair policies changed, and the --legacy-node-repair flag offers a fallback. The new disruption and reboot metrics give more detail for monitoring node churn.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

• +Dra prioritized alternatives support (#3142) #3142 (Ryan Mistretta)added
• +Metrics: add karpenter_pods_disrupted_total counter (#3015) #3015 (praveen9354)added
• +Metrics: add consolidation_policy and termination_mode labels to disruption metrics (#3179) #3179 (Derek Frank)added
• +Add VPA prediction controller and store (#3134) #3134 (Jigisha Patil)added
• +Default pod topology spread awareness (#3205) #3205 (Andrew Mitchell)added
• +Account for instance type volume limits when provisioning (#3202) #3202 (Eddie)added
• +Reduce allocations of Requirement and Requirements (#3211) #3211 (Bevan Arps)added
• +Metrics: document metric dimensions with built-in help and values (#3295) #3295 (Derek Frank)added
• +Metrics: split the disruption decision dimension into command vs approval decisions (#3322) #3322 (Derek Frank)added
• +Metrics: name and document the unknown cloudprovider error value (#3323) #3323 (Derek Frank)added
• +Disruption: Add pod disruption tracking to perf test reporting (#2892) #2892 (Nathaniel Jones)added
• +Node repair as voluntary disruption (#3311) #3311 (Derek Frank)added

Kubernetes v1.37: Tracking When a PersistentVolumeClaim Was Last Used (Beta)

Kubernetes v1.37 promotes the PersistentVolumeClaimUnusedSinceTime feature gate to Beta and enables it by default. The PVC protection controller now sets an Unused condition on each PVC, with status True and reason NoPodsUsingPVC when no non-terminated pod references it. The condition's lastTransitionTime records when the PVC became idle.

platform teams Platform teams can list idle PVCs with kubectl and jq instead of cross-referencing pods and volumes with custom scripts, which helps find orphaned storage that drives cloud costs. The feature is on by default in v1.37, and the project intends to reach GA in a later release.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

Kubernetes v1.37: Pod-Level Resource Managers graduated to Beta

Kubernetes v1.37 graduates the Pod-Level Resource Managers feature to Beta, disabled by default and controlled by the PodLevelResourceManagers feature gate. The Topology Manager, CPU Manager and Memory Manager can now use pod-level resource declarations when making hardware placement decisions. The v1 PodResources gRPC service also adds top-level cpu_ids and memory fields for pod-level exclusive assignments.

platform teams Platform teams running latency-sensitive workloads can reserve exclusive NUMA-aligned CPU and memory for primary containers while placing non-Guaranteed sidecars in a pod-isolated shared pool. Teams that opt in should check monitoring and device plugins that read PodResources, since they can now query pod-level assignments without double-counting containers.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

Kubernetes v1.37: Scheduler Preemption for In-Place Pod Resize (Alpha)

Kubernetes v1.37 adds an Alpha feature behind the InPlacePodVerticalScalingSchedulerPreemption feature gate that lets the kube-scheduler preempt lower-priority Pods on the same node so that a running Pod's deferred in-place resize can be actuated. Previously a Deferred resize could remain pending indefinitely until node capacity freed up. The Kubelet delegates this preemption decision to the scheduler rather than evicting locally.

platform teams Platform teams bin-packing lower-priority batch workloads can allow higher-priority Pods to scale up in place without restarts, but must enable the gate across kube-apiserver, kube-scheduler and kubelet on v1.37 clusters. The feature is Alpha, and the new spec.podPreemptionPolicy node field lets controllers opt specific nodes out of resize preemption.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

Kubernetes v1.37: Introducing Node Lifecycle Conditions

Kubernetes v1.37 reserves five well-known Node condition types: DrainInProgress, Drained, MaintenancePlanned, MaintenanceInProgress, and GracefulNodeShutdownInProgress. It also adds an Alpha NodeLifecycleConditions feature gate, disabled by default, which the release describes as a no-op in v1.37 since no core component reads the conditions yet. Administrators or authorized controllers are responsible for setting and clearing them.

platform teams Platform teams can start publishing these conditions today from drain and maintenance automation to give dashboards, alerts and tooling a common status signal. No core controller changes behavior based on them in this release, so scheduling and eviction still rely on existing mechanisms.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

Kubernetes v1.37: Scale Workloads to Zero with HorizontalPodAutoscaler

Kubernetes v1.37 moves horizontal autoscaling to zero replicas to Beta and enables it by default. An HPA using an object or external metric can now scale a workload to zero and bring it back when the metric changes, a capability that previously required an add-on or the Alpha feature gate. The blog notes that Services do not buffer requests, so request-driven workloads need a separate buffering layer.

platform teams Platform teams can scale idle queue consumers and batch workers to zero to cut costs on expensive CPU or GPU Pods, but must accept cold-start latency and make sure the external metrics adapter is reliable. Before upgrading, control plane components must both have the feature enabled, and HPAs with minReplicas 0 need an object or external metric.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

Kubernetes v1.36: New Metric for Route Sync in the Cloud Controller Manager

Kubernetes v1.36 adds an alpha counter metric, route_controller_route_sync_total, to the Cloud Controller Manager route controller. The metric increments each time routes are synced with the cloud provider. It is intended to help operators A/B test the CloudControllerManagerWatchBasedRoutesReconciliation feature gate, introduced in v1.35, which reconciles routes only when nodes change instead of on a fixed interval.

platform teams Platform teams can compare sync counts with the feature gate on and off to measure reduced calls to the infrastructure provider API. The source expects a significant drop in sync rate in clusters where nodes change infrequently, but gives no measured figures.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

Kubernetes v1.36: More Drivers, New Features, and the Next Era of DRA

The Kubernetes v1.36 blog post describes DRA changes in that release, including Prioritized list and Partitionable devices graduating, plus new alpha features such as node allocatable resources via DRA and a resource pool status request object. Device taints, binding conditions, and resource health status moved to beta. The post states that the new alpha features sit behind feature gates disabled by default.

platform teams Platform teams running GPU or other accelerator workloads can use fallback device preferences, device taints, and pool status queries for scheduling and capacity planning. Alpha features like node CPU and memory via DRA should be evaluated only in test clusters for now.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules

Kubernetes v1.36: Server-Side Sharded List and Watch

Kubernetes v1.36 adds server-side sharded list and watch as an alpha feature (KEP-5866). The API server filters objects by hash range using a new shardSelector field on ListOptions, so each controller replica receives only the objects it owns. The feature requires the ShardedListAndWatch feature gate and echoes applied selectors in a shardInfo field when honored.

platform teams Platform teams running horizontally scaled controllers on very large clusters could cut per-replica deserialization, CPU and network cost, since replicas no longer receive the full event stream. Because the feature is alpha and gated, clients must fall back to client-side filtering when shardInfo is absent.

written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules