SIG Architecture has completed a review of feature gates on master (v1.38 development) against the feature gates documentation. The review identified overdue feature gates, gates due for removal in v1.38, GA/deprecated gates that are not locked to their default, and several stale alpha/beta gates across SIG Node, API Machinery, Auth, Scheduling, Network, Windows, Storage, Instrumentation, Apps, and Cloud Provider.
SIG UI has nominated Kahiro Okina (@kahirokunn) as a new SIG UI Lead, following a recommendation from current SIG UI Chairs. Kahiro has been a contributor and reviewer to Headlamp, contributing to its multi-cluster UI, plugin ecosystem, mentoring, and community growth. The nomination is now in a one-week lazy consensus period.
The Steering Committee election voting will remain open through October 1, with exception requests due by September 29. If you are marked as not eligible despite your contributions, please request an exception.
Next Deadline: KEP Readiness Deadline, September 22, 2026 (AoE) / September 23, 2026 at 12:00 UTC
The Kubernetes v1.38 release cycle began on August 31, with the final release targeted for December 16, 2026. The KEP Readiness Deadline is scheduled for September 22, 2026 (AoE) / September 23, 2026 at 12:00 UTC. Enhancements are expected to complete their Production Readiness Review by this deadline; enhancements that miss it require an exception. The Enhancements Freeze is scheduled for September 29 (AoE) / September 30 at 12:00 UTC. If you’d like to include your KEP in the v1.38 release cycle, please reach out to your SIG leads to get your KEP tracked for the release. Find more information in the Call for Enhancements announcement.
luxas advanced KEP-5681: Conditional Authorization by adding conditional authorization support to the kube-apiserver. Conditional authorization lets an authorizer return requirements that depend on request-object content, which the API server can evaluate later during admission instead of making a decision solely from request metadata. This PR is the fourth part of the implementation series and builds on earlier core machinery, enabling more context-aware authorization policies for Kubernetes APIs.
NAVEENKUMARRK777 implemented the alpha allowDisruptionByPriorityGreaterThanOrEqual capability described in KEP-3280. The change adds API and scheduler support for controlling which lower-priority workloads may be disrupted, based on a priority threshold. This gives workload authors and cluster operators more precise control over disruption behavior for priority-sensitive applications.
jdizkowski implemented recursive status handling for CompositePodGroups, part of the CompositePodGroup API proposal (KEP-6012). CompositePodGroups are intended to represent hierarchical, heterogeneous workloads whose components may have different scheduling requirements. This PR propagates status through the hierarchy, allowing users and controllers to observe the aggregate state of complex workload groups instead of inspecting each child PodGroup separately.
This enhancement extends KEP-3756 to recover CSI volumes after a kubelet restart when pod-local vol_data.json metadata is missing or corrupt. Without the fallback, a global mount can remain active after its Pod directory disappears, causing the volume to leave node.status.volumesInUse and potentially allowing a ReadWriteOnce filesystem to be attached to another node. The proposal lets kubelet reconstruct the volume from metadata stored beside the global mount, then verify or unstage it through the normal volume-manager flow.
The work builds on a long-standing reconstruction problem tracked in kubernetes/kubernetes#101791 and the manual-cleanup procedure documented by KEP-3756. cniackz, with SIG Storage and SIG Node contributors including jsafrane and gnufied, is implementing the fallback in kubernetes/kubernetes#138454. The KEP is implementable, and its KEP and implementation updates are currently under review.
KEP 6058 is targeting Alpha in Kubernetes v1.38.
The change is opt-in behind a new alpha feature gate and affects kubelet CSI reconstruction and unmounting. It is intended to prevent leaked global mounts and reduce the risk of double-mounting and data corruption on storage systems such as Fibre Channel, iSCSI, and EBS.
--restart flag values on kubectl run has been improved.+k8s:immutable tag to validation-gen so declarative validation can express immutable field semantics.PostBind extension point under KEP-6021 for plugins that run after a Pod has been bound to a Node.--warnings-as-errors support to kubectl apply and kubectl apply --dry-run=server, so CI pipelines can fail on server-side warning conditions.memory.swap.max so allocated swap is now correctly reflected in Pod status.kube-controller-manager adds a new _total counter for controller reconciliation errors.PodSandboxConfig.resource_claims is now propagated to the runtime for DRA-enabled pods.+k8s:required codegen so pointer-typed required fields don’t incorrectly reject zero values.scheduler_perf so DRA benchmarks can measure Extended Resource pathways.PrioritizedList requests.SchedulingResult to Pod events so future scheduling events can reference the prior decision.events_dispatched_total was incremented for undispatched bookmarks.endpoint_slice_controller_changes to endpoint_slice_controller_changes_total to add the conventional _total suffix; the old name is retained temporarily for backward compatibility./flagz endpoint reflects flag layer, not effective config.ImageVolume to GA.The Maintainer Summit schedule for Kubecon North America is live; it includes two SIG/Subproject meetings and Ask The Kubernetes Steering Committee. The same week, the Kubecon Europe CfP opened.
Voting is now open for the 2026 Kubernetes Steering Committee election. Six candidates are running this year, including Sean McGinnis (stmcginnis), Michael McCune (elmiko), Priyanka Saggu (priyankasaggu11929), Janet Kuo (janetkuo), Kaslin Fields (kaslin), and Jack Francis (jackfrancis).
Voting will remain open through October 1, with exception requests due by September 29. If you are marked as not eligible despite your contributions, please request an exception.
KCD Around the World: San Francisco Bay Area 2026 Organizers explores how KCD San Francisco Bay Area is bringing together a vast and often siloed cloud native community. Kashish Verma and Rey Lejano sit down with organizers Lisa-Marie Namphy and Jason (Jay) Smith to talk about how the event came together, what they learned from its first edition, and how this year’s KCD is evolving with the growing intersection of Kubernetes, AI, and open source.
Next Deadline: v1.38 Release Cycle Begins, August 31
The Kubernetes v1.38 release cycle began on August 31, with the final release currently targeted for December 16, 2026.
Starting with v1.38, the release cycle has been adjusted to give contributors more time between Enhancements Freeze and Code Freeze. The change is intended to give KEPs more time to be reviewed and merged before Code Freeze. The schedule is still being finalized, so more details will follow as the cycle progresses.
BitWeaverDev added terminating-state
visibility for StatefulSets in kubectl get. This makes it easier for users
and operators to identify StatefulSets that are being deleted but have not
finished terminating, without needing to inspect individual Pods or controller details. The change improves the day-to-day observability of StatefulSet lifecycle operations and helps users investigate deletions that are taking longer than expected.
harche fixed extended-resource accounting when
DRANodeAllocatableResources is enabled. Incorrect handling in this path could cause the node’s available extended resources to be represented inaccurately, affecting scheduling decisions for Pods that use Dynamic Resource Allocation. This PR corrects the resource information exposed for DRA, improving coordination between node resource reporting and scheduling for workloads that use device or other extended-resource claims.
Kubernetes v1.37.0 was released on August 26; see the Release Schedule below for details.
SIG Windows has added @jrvaldes as co-chair, with Aravindh Puthiyaparambil moving to emeritus.
SIG Node has published a new developer guide for node declared features, documenting how contributors should declare and use node-level feature signals.
Voting for the Kubernetes Steering Committee continues. If you are a Kubernetes Org member, please vote before October 1.
Kubernetes v1.37: Garhwal has been released on 26th August! The Release Announcement and Release Notes are now available. A few contributor-focused changes worth calling out:
Starting with v1.38, the release cycle will also have more time between Enhancements Freeze and Code Freeze, giving contributors more time to get their work reviewed and merged and helping reduce the need for Code Freeze exceptions.
harche fixed extended resource accounting when the DRANodeAllocatableResources feature is enabled. The scheduler and kubelet rely on node allocatable information to determine whether a Pod’s Dynamic Resource Allocation claims can be satisfied. This PR corrects the handling of extended resources in that path, preventing inaccurate node capacity information from affecting scheduling decisions for workloads that use DRA.
grosser fixed the OpenAPI schema for ParamRef.parameterNotFoundAction in admission registration APIs. Although the API requires callers to specify how admission policy evaluation should behave when a referenced parameter object is absent, the published schema did not mark the field as required. This PR aligns the OpenAPI contract with the API’s validation requirements, improving generated clients, schema validation, and tooling that consumes Kubernetes API definitions.
This enhancement, authored by @pohly and @everpeace, extends Dynamic Resource Allocation (DRA) with a tainting mechanism for individual hardware devices, modeled on how Kubernetes already taints Nodes. DRA drivers or admins can mark a specific device, such as a degraded GPU, as tainted, and workloads declare tolerations on their DeviceRequests to decide whether they’re still willing to use it. A new DeviceTaintRule API lets admins patch taints onto matching devices without needing driver support, and a dedicated controller evicts non-tolerating pods much like the existing node taint-eviction-controller.
KEP 5055 enhancement is stable in v1.37.
containerd is deprecating the implicit restore-via-CreateContainer codepath introduced by KEP-2008 (Forensic Container Checkpointing) after multiple CVEs; containerd 2.3.4 and 2.2.7 retain it as an experimental opt-in behind enable_experimental_restore_via_create, and containerd 2.4 removes it in favor of the explicit RestorePod API from KEP-5823.
SIG ContribEx has added AI notetaker guidance to Kubernetes Zoom meeting guidelines. AI notetaker bots (such as Granola, Otter.ai, Fireflies.ai, Fathom, or Read.ai) are not permitted on Kubernetes Zoom calls. Unauthorized third-party recording poses risks around consent, privacy, and moderation.
Voting is now open for the 2026 Kubernetes Steering Committee election. Six candidates are running this year, including Sean McGinnis (stmcginnis), Michael McCune (elmiko), Priyanka Saggu (priyankasaggu11929), Janet Kuo (janetkuo), Kaslin Fields (kaslin), and Jack Francis (jackfrancis).
Voting will remain open through October 1, with exception requests due by September 29. If you are marked as not eligible despite your contributions, please request an exception.
Next Deadline: v1.37.0 release, 26th August
The v1.37 release cycle enters its final week. Kubernetes v1.37.0-rc.1 has been released, built with Go 1.26.5, following v1.37.0-rc.0 on August 6. The final v1.37.0 release, and release blog publication are all targeted for August 26.
August Kubernetes patch releases have been delayed to August 19 to allow additional time for cherry-picks affected by recent GitHub disruptions and automation failures.
Kubernetes v1.38 Release Team Shadow application is open through August 28, with the v1.38 release cycle expected to begin around August 31.
pravk03 fixed DRA resource accounting for init containers. Resource overhead from DRA claims used by init containers should contribute to a Pod’s peak resource usage while those containers are running, but should not continue reducing the node’s steady-state allocatable capacity after they finish. This PR corrects that distinction, helping prevent nodes from appearing to have less capacity than they actually do and improving scheduling decisions for workloads that use Dynamic Resource Allocation.
RafPe fixed kubelet cache invalidation for name-keyed ClusterTrustBundle objects. Without invalidating cached entries when a bundle changes, kubelet could continue using stale trust-bundle data instead of observing the latest object contents. This PR ensures that updates to named ClusterTrustBundles invalidate the corresponding cache entries, allowing kubelet consumers to use current trust information.
This enhancement extends the Workload API and PodGroup API to allow ResourceClaims and ResourceClaimTemplates to be associated with an entire PodGroup, making it easier to share DRA resources across multiple related Pods. A ResourceClaim referenced by a PodGroup is reserved for the whole group rather than each individual Pod, avoiding the current 256-entry limit in status.reservedFor. Similarly, a ResourceClaimTemplate can automatically generate one ResourceClaim for the PodGroup, eliminating the need to manually create and name shared claims. This is particularly useful for large AI/ML workloads, where closely related Pods need to be placed within the same network or topology boundary for better performance. The enhancement also makes resource sharing more practical for workload APIs such as JobSet and LeaderWorkerSet, allowing Kubernetes to manage shared DRA resources generically through the common PodGroup abstraction.
This enhancement has been proposed to become Beta for Kubernetes v1.37.
SIG Autoscaling has clarified its subproject donation requirements, formalizing what external projects need to satisfy to be adopted under the SIG.
Dims has published a post-Code-Freeze CI recovery writeup showing that contributors across SIGs fixed roughly 27 CI jobs through 60 PRs since July 22, with the sig-release-master-blocking board now green; the full gist lists about 30 open failing-test and flake issues that still need owners.
The KubeCon + CloudNativeCon NA 2026 schedule is now live.
Voting is now open for the 2026 Kubernetes Steering Committee election. Six candidates are running this year, including Sean McGinnis (stmcginnis), Michael McCune (elmiko), Priyanka Saggu (priyankasaggu11929), Janet Kuo (janetkuo), Kaslin Fields (kaslin), and Jack Francis (jackfrancis). Voting will remain open until October 1.
All current Kubernetes org members are eligible to vote this year due to GitHub contributions not being accurately tracked during the first four months of 2026. Members of the SRC or CoCC are also eligible. Eligibility can be checked at elections.k8s.io or in the full voter list in the election repository. If you believe you should be eligible, but are not, please file an exception.
Next Deadline: RC1 release, August 19
We are in the final stretch for v1.37, which looks like it’s going out on time. If you’re working on Release Notes or the blog you have plenty to do, but most contributors just need be ready to respond quickly to any test failures. This means you can start thinking about v1.38. Speaking of which, the v1.38 Shadow Application is already open, and due August 28th.
vshkrabkov fixed Pod event handling for updates where the Pod UID changes. Because a UID identifies a specific Pod instance, an update with a new UID represents a different Pod rather than a normal modification. Treating it as an ordinary update could leave consumers with stale state associated with the previous Pod. This PR represents the transition as a delete event for the old Pod followed by an add event for the new Pod, allowing API machinery and scheduling consumers to process the identity change correctly.
macsko fixed scheduler state tracking for Pods rejected during the PlacementFeasible stage. The scheduler uses UnschedulablePlugins to identify which plugins made a Pod unschedulable and which cluster events should cause it to be retried. This PR ensures that the relevant plugin information is recorded for PlacementFeasible rejections, improving requeue behavior for workload-aware scheduling and Dynamic Resource Allocation scenarios where placement can depend on device, resource, or topology changes.
This KEP proposes exposing device health information (such as GPU failures) in Pod Status so Kubernetes can indicate when a Pod is using a device that is unhealthy or has failed. Currently, device failures are difficult to diagnose because Pods often enter CrashLoopBackOff without any indication that the assigned device is the root cause, and Kubernetes may keep restarting the Pod on the same faulty device. By surfacing unhealthy devices directly in Pod Status, the proposal provides a generic mechanism for easier troubleshooting and better handling of temporary or permanent device failures, while keeping the initial scope small and extensible for future enhancements.
PodLevelResourceManagers is disabled by default.The Kubernetes Steering Committee held a Q&A session for election candidates at the public Steering meeting on Wednesday, August 5.
SIG ContribEx has merged documentation for the Kubernetes social media automation process, formalizing how community posts across project accounts are triaged, drafted, and published.
KCD Around the World is a new series exploring Kubernetes Community Days through the volunteers, contributors, organizers, and attendees who make them happen. In the first edition, Kashish Verma interviews KCD New York 2026 organizer Christopher Tineo about what makes the New York community unique and the moments, conversations, and people that made this year’s event special.
Next Deadline: Docs Freeze, August 5; Release blog ready to review, August 13
We are in Docs Freeze. The Kubernetes v1.37 release blog will reach the “Ready for Review” milestone on August 13 as the release team continues preparing for the v1.37 release.
The target date for this month’s patch releases is next week, August 11.
lukaszwojciechowski fixed a kubelet CPU Manager bug that could re-allocate CPUs incorrectly for pods using pod-level resources after a kubelet restart. The bug affected pods with pod-level exclusive CPU assignments under the static CPU Manager policy. Before this fix, kubelet could either reject a restored pod with “not enough cpus available” or move it to different CPUs, leaking the original exclusive CPUs. The PR adds a restore path that reuses the checkpointed pod-level CPU set and container assignments instead of blindly allocating again.
suii2210 fixed a StatefulSet reconciliation deadlock that could leave podManagementPolicy: Parallel workloads stuck after a failed scale-up. A partially successful rollout could leave a missing desired Pod and an out-of-range condemned Pod at the same time. If creation of the missing Pod failed because of ResourceQuota, the controller previously returned before deleting the condemned Pod, so the quota was never released and reconciliation could not make progress. This PR changes the ordering for Parallel StatefulSets so condemned Pods are processed before missing desired replicas. The existing OrderedReady behavior is unchanged.
This proposal extends the use of Kubernetes NominatedNodeName so the scheduler can record a pod’s expected node placement at the beginning of the binding cycle, rather than using it only for preemption. The goal is to make the scheduler’s internal node reservation visible to other components, such as the Cluster Autoscaler, which may otherwise incorrectly treat the node as underutilized and delete it while the pod is still in the binding phase (for example, during PreBind volume or device allocation). By exposing the expected placement through NominatedNodeName, external components can make better decisions, and the scheduler can also retain scheduling decisions across restarts, allowing PreBind operations to resume on the same node. This improves coordination between the scheduler, autoscaling, and resource allocation mechanisms such as DRA ResourceClaims. NominatedNodeName is in Beta since v1.35.
PodLevelResourceManagers is disabled by default.fabrizio.pandini: Shoutout to @Karthik-K-N for implementing a series of changes that reduce Cluster API E2E test duration by about 30 minutes, benefiting all Cluster API contributors.
jberkus: Kudos to @Furkan AYDIN for helping us quickly move to the new inviter.co for this Slack and troubleshooting permission issues.
Kubernetes v1.36.3 has been released, built with Go 1.26.5. This patch release includes fixes for DRA scheduling and allocator bugs, a kubelet memory leak regression introduced in 1.36, server-side apply regressions, kubeadm etcd learner promotion and join resiliency improvements, and reduced kubelet log noise for missing optional container annotations.
Kubernetes v1.37.0-beta.0 has been released, built with Go 1.26.5. Highlights include a new Recreate strategy for StatefulSets, ClusterTrustBundle moving to stable, metrics.k8s.io promoted to v1, PersistentVolumeClaimUnusedSinceTime enabled by default in beta, new scheduler and workload-aware preemption capabilities, default etcd 3.7.0, and numerous fixes across DRA, scheduling, CEL, kubelet, kube-proxy, and API machinery.
Nominations are now open for the 2026 Kubernetes Steering Committee election, with three seats available for two-year terms. Self-nominations and community nominations must be submitted via a GitHub issue by August 3, with candidate bios due August 4. Prospective candidates are also invited to a public Steering Committee Q&A session on August 5, where current committee members will discuss what serving on the Steering Committee is like and answer questions from candidates.
client-go enhancements in Kubernetes v1.37 introduce a new type-safe informer event handler API, reducing boilerplate and making deletion event handling safer and easier. The release also adds a client-go CHANGELOG.md for documenting breaking Go API changes, expands context handling and contextual logging support, and introduces tooling to help contributors track and document API compatibility changes.
Next Deadline: Feature Blogs Ready, August 3; Docs Freeze August 5
We are in Code Freeze. If you are working on a feature blog post for v1.37, remember that you need that ready for editing by Monday. If you are responsible for an Enhancement, please have your documentation finalized by Wednesday.
Cherry-pick deadline for the next patch releases is next week, August 7.
In this pull request, tallclair introduced the KubeletAllocatedPodsEndpoint feature gate and the corresponding /allocatedPods endpoint on the Kubelet. Linked to the Dynamic Containers proposal (KEP-6030), this endpoint allows the API server to query an on-demand /allocated subresource directly from the node without inflating storage overhead in etcd. By serving the Kubelet’s locally assigned pod specifications, this change provides a lightweight way to track when a running pod’s allocated spec diverges from its desired state, laying critical groundwork for in-place resizing and dynamic container features in v1.37.
In this pull request, rst0git introduced pod-level checkpoint and restore RPCs (CheckpointPod and RestorePod) to the v1.RuntimeService CRI interface. Driven by KEP-5823, this establishes a standardized contract between the Kubelet and container runtimes (such as CRI-O and containerd) to save and recover entire pod sandbox states without requiring Kubernetes to manage or interpret runtime-specific state files. This update lays the core API foundation for native stateful pod migration and fast-booting workloads across Kubernetes clusters.
In this pull request, tallclair promoted PLEGOnDemandRelist to General Availability, making the optimized kubelet relist behavior the permanent default in Kubernetes v1.37 after proving stable throughout v1.36. This graduation removes the feature gate and confirms the implementation is production-ready for all users. The change improves kubelet reliability and simplifies long-term maintenance for SIG Node while signaling confidence that no emergency rollback path is needed.
KEP-6164, authored by @jpbetz, @michaelasp, and @yedou37 in June 2026, eliminates Kubernetes’ long-standing internal
(__internal) API types in favor of stable versioned API types. By removing unnecessary conversions inside the kube-apiserver, the enhancement reduces conversion overhead, with benchmarks showing up to 5.8x faster conversions and significantly lower memory allocations for large Pod list operations. The implementation follows a phased approach: first making internal and versioned types memory-identical, then replacing internal types with Go type aliases before eventually removing them entirely. Targeting Stable in Kubernetes v1.37, the enhancement improves control plane performance and simplifies API maintenance without changing any user-facing APIs or wire formats.
allocationMode: All requests did not properly fail when a matching device was capacity-exhausted under DRAConsumableCapacity.watch_list_duration_seconds metric for better observability of watch-list operations.PodGroupPostFilter plugins which incorrectly referenced “PostFilter plugins” instead of “PodGroupPostFilter plugins”.storage_to_cache stage to the watch dispatch metric for finer-grained latency observability of the watch pipeline.allocatedPods Kubelet endpoint gated by the KubeletAllocatedPodsEndpoint feature gate is added; this surfaces allocated-pod information from the Kubelet and backs the /allocated Pod subresource proposed under KEP-5972 Dynamic Containers.cache_to_watcher stage to the watch dispatch metric, complementing the new storage_to_cache stage for end-to-end latency visibility.Unused condition evaluation for unbound PVCs, since the condition only applies to bound PVCs under the PersistentVolumeClaimUnusedSinceTime feature.ResourceClaim events, ensuring pods are re-queued when a ResourceClaim allocation is removed.error outcome to the cloud-provider route_sync_total metric, which previously only counted successful syncs.numgen map for single-endpoint service DNAT, reducing per-rule scan cost in clusters with many Services.AllowUnsafeMalformedObjectDeletion to Beta under KEP-3926.PLEGOnDemandRelist to GA.InPlacePodVerticalScalingInitContainers to GA.Maintainer Summit North America 2026 will take place on November 8 in Salt Lake City ahead of KubeCon + CloudNativeCon North America. Registration is open to Kubernetes and CNCF project maintainers, organization members, and community leaders.
Janet Kuo announced her self-nomination for the 2026 Kubernetes Steering Committee election, highlighting her long-standing contributions to Kubernetes, leadership in SIG Apps, and work on AI initiatives including Agent Sandbox and AI Conformance.
Node Feature Discovery v0.19.0 has been released. Highlights include immediate node re-labeling after node rebuilds, configurable NodeFeature owner references, new x86-64 microarchitecture labels, topology updater scalability improvements, and reconciliation performance enhancements. Contributors upgrading should review the release’s action-required notes, including the required RBAC update.
Next Deadline: Code Freeze and Test Freeze, July 22nd
Kubernetes v1.37 has entered Code Freeze and Test Freeze. Enhancements targeting this release should now be merge-ready with the required approvals, passing tests, and documentation PRs opened. From this point onward, only release-blocking changes should proceed.
Patch releases v1.36.3, v1.35.7, and v1.34.10 are now available. These releases include Go updates and bug fixes.
In this pull request, Jefftree enabled the EtcdRangeStream feature gate by default, advancing the feature to Beta as part of KEP-5966: Etcd Range Streaming. Instead of retrieving entire response sets from etcd in a single, memory-intensive operation, the kube-apiserver can now stream large LIST responses incrementally in smaller chunks. This significantly reduces peak memory usage during large list operations and watch cache initialization while improving performance. Benchmark results on 5,000-node clusters show approximately a 1.4× improvement in watch cache initialization time, making this an important scalability enhancement for large production Kubernetes clusters. See also the related implementation work in the ongoing Watch/List streaming effort, including PRs such as #139692.
In this pull request, stlaz promoted ClusterTrustBundles and ClusterTrustBundleProjection to General Availability (GA), completing the work tracked by KEP-3257. The stable certificates.k8s.io/v1 API enables cluster administrators to distribute cluster-wide X.509 trust anchors (CA certificates) to workloads through projected volumes, providing a standardized and production-ready mechanism for managing trust bundles. This eliminates the need for custom ConfigMaps or out-of-band certificate distribution methods, simplifying certificate management while improving security and portability for workloads that depend on dynamically managed trust anchors.
KEP-5972, authored by Tim Allclair (@tallclair) in June 2026, proposes allowing main containers to be added to and removed from running Pods, making containers mutable after Pod creation. It is intended to support modern AI and agentic workloads, including high-performance batch jobs and low-latency AI agents, but would also change Kubernetes’ long-standing assumption that Pods are immutable execution envelopes.
If adopted, it would require updates to existing ecosystem components such as third-party controllers, service meshes, and logging tools that assume .spec.containers never changes after Pod creation. It also introduces two new Pod subresources: /dynamic for adding and removing containers in running Pods, and /allocated, which returns the allocated version of the Pod directly from the Kubelet. The KEP is currently undergoing broader design review, with discussion focused on its architectural scope and the impact such a fundamental change could have across the Kubernetes ecosystem.
As this proposal changes one of Kubernetes’ core assumptions about Pod immutability, it remains under active discussion for a future Kubernetes release. Follow the ongoing discussion in the KEP pull request or the SIG Architecture mailing list.
CapacityRequestPolicyRange did not support fractional quantities, aligning capacity-range behaviour with resource.Quantity semantics.--metrics-bind-address flag documentation, which incorrectly described the default binding behaviour.Recreate update strategy for StatefulSets, giving operators a StatefulSet analog to the Deployment Recreate strategy.MutatingAdmissionPolicy and MutatingAdmissionPolicyBinding (GA in 1.36) are now stored as v1, fixing autocalculation of the emulated storage version.EtcdRangeStream feature gate is enabled by default (Beta); watch-cache initialization is ~1.4× faster in unit benchmarks.resource.Quantity mutation bug in DRA counter caches that could corrupt device accounting.DistinctAttribute handling for DRA requests with multiple devices in the consumable-capacity path.GetResetFields implementations slip through undetected.resource.kubernetes.io/numaNode standard device attribute (KEP-6072), enabling cross-driver NUMA co-placement via matchAttribute.--restart or --image-pull-policy (e.g. --restart=never instead of Never).quantity.Add would mutate the receiver.distroless-iptables v0.9.4 (nft 1.0.6.1) to fix crashes on systems with nft 1.1.3.PreemptionPolicy field to PodGroupTemplate under KEP-5710.SelectionPredicate.Empty() to be aware of sharding, correcting selector-emptiness reporting when sharding is in use.matchAttribute constraint bug under KEP-5491 so the intersection is properly restored on backtrack when DRAListTypeAttributes is enabled.Maintainer Summit North America 2026 CFP CFP closes on July 19. Maintainers, project leaders, and contributors are encouraged to submit session proposals before the deadline. Registration is open for it as well.
Steering Committee election nominations continue, with Janet Kuo announcing her candidacy and Sean McGinnis receiving a community nomination. Nominations close on August 3.
Next Deadline: Code Freeze, July 23th
If your enhancement is unlikely to be ready by Code Freeze, now is the time to request a Code Freeze Exception. Review the Exceptions process to ensure your request meets the requirements and is submitted before the upcoming Code Freeze.
Patch releases for all supported Kubernetes versions are targeted for July 14th.
In this pull request Jefftree updated Kubernetes to use etcd v3.7.0, replacing the previous release candidate with the final upstream version and updating both the embedded client libraries and default etcd images. This brings Kubernetes in sync with the latest stable etcd release while enabling new storage capabilities that future Kubernetes features can build upon, including ongoing work around efficient Watch/List operations. See the related upstream discussion in the etcd v3.7.0 release and the related Watch/List streaming work in Kubernetes.
In this pull request pohly advanced the Dynamic Resource Allocation (DRA) Device Taints and Tolerations feature to General Availability via the resource.k8s.io/v1 API. Linked to KEP-5055, this milestone stabilizes how clusters restrict or allow specialized hardware access, such as GPUs, using dedicated device-level taints. Moving out of the experimental phases means cluster operators and hardware vendors can confidently implement production-ready, fine-grained resource scheduling without worrying about feature gate maintenance.
rthallisey introduced several critical, well-known node conditions into the core API, including GracefulNodeShutdownInProgress, DrainInProgress, Drained, MaintenancePlanned, and MaintenanceInProgress. This foundational change implements the initial phase of KEP-6161, designed to build explicit ecosystem awareness around structural cluster disruptions. By exposing these standardized node states, downstream automation tools, operators, and core controllers can programmatically coordinate edge cases during maintenance workflows to improve overall workload uptime.
In this pull request, luxas implemented the core engine logic for the highly anticipated Conditional Authorization feature Part of KEP-5681, this foundational work builds directly upon earlier API refinements to establish the machinery required for context-aware access control. By providing these framework hooks, the API server will soon be able to evaluate complex runtime variables and policies before granting client permissions, significantly upgrading security capabilities across multi-tenant clusters.
vshkrabkov enhanced the default scheduling flow to support PodGroups as atomic preemption victims Linked to the Workload-Aware Preemption design under KEP-5710, this change introduces clean interface abstractions for Victim and DomainVictim to group node-level state. Rather than evaluating and evicting pods individually, the scheduler can now remove an entire group as a single unit, avoiding partial preemption failures and heavily optimizing gang-scheduled AI or batch workloads.
In this pull request, GFilipek introduced the new PodGroupPostFilter extension point to the scheduling framework API. Associated with KEP-5710, this clean architectural change replaces previously hardcoded Workload-Aware Preemption logic with a fully pluggable interface. This enables out-of-tree and custom plugin authors to seamlessly build and configure their own complex preemption and scheduling rules when dealing with multi-pod groups and batch scheduling behaviors.
KEP-5963: DRA Device Compatibility Groups
This KEP extends the Dynamic Resource Allocation (DRA) ResourceSlice API to let device drivers declare mutually exclusive allocation constraints between groups of devices. This addresses hardware (such as GPUs) that support multiple incompatible partitioning or virtualization modes, where selecting one configuration prevents others from being used simultaneously.
Today, DRA cannot express these constraints, causing the scheduler to make invalid allocation decisions. As a result, conflicts are detected only during resource preparation, leading to pod startup failures, repeated scheduling retries, and a poor user experience. This proposal enables the scheduler to understand these constraints upfront, preventing incompatible allocations before scheduling.
This KEP is currently in Alpha stage.
memory.low cgroup value was ineffective due to missing ancestor cgroup coverage.PodGroupPostFilter scheduler extension point that runs after standard PostFilter for gang-scheduled workloads.--max-depth flag to truncate schema output for deeply nested resources like CRDs and Pod specs.cluster-info dump --output-directory creates output files with owner-only permissions, preventing world-readable exposure of potentially sensitive cluster state.PodGroup.Status was not updated when pods in the group had mismatched .spec.schedulerName or priority values.metrics.k8s.io API promoted to v1 GA.github.com/google/cadvisor/lib bumped from v0.60.3 to v0.60.4, which pulls in updated golang.org/x/* module versions across the tree.etcd v3.7 was released last week. Highlights include RangeStream (to be supported by Kubernetes v1.37), bootstrapping from v3store, a protobuf overhaul, and multiple performance improvements. Kubernetes v1.37 will test against etcd v3.7.
Kubernetes v1.37 has reached its mid-cycle milestone with 86 tracked enhancements. Feature blog placeholders are now due, while Code Freeze begins on July 22 ahead of the planned August 26 release.
SIG etcd has released etcd v3.7.0, introducing the long-awaited RangeStream feature, performance improvements, the removal of the legacy v2store, and a major protobuf overhaul.
The Kubernetes Steering Committee election process is underway. Community member Tim Hockin has opened a self-nomination and is seeking endorsements from eligible voters to qualify for the 2026 election.
The CFP for the Maintainer Track, ContribFest, and Lightning Talks at KubeCon North America 2026 closes on July 12. Kubernetes SIGs and Working Groups are encouraged to submit session proposals before the deadline.
Next Deadline: Feature Blog Freeze, 10th July 2026
Feature Blog Freeze placeholders for Kubernetes v1.37 are due July 10th (AoE). If your enhancement is getting a feature blog, open a placeholder PR before the deadline; the Release Comms team will follow up on delivery and review from there.
Cherry-picks for the July patch releases are due July 10th, with releases targeted for July 14th.
In this pull request hakuna-matatah introduced the HPAOptimizedSelectorStore feature gate to replace the legacy, lock-heavy bimultimap datastore inside the HorizontalPodAutoscaler controller. This architectural update mitigates extreme Read/Write mutex contention, reducing average lock wait times by 214×. As a result, reconciliation throughput exploded by up to 195× under massive concurrent workloads. This optimization allows the HPA controller to scale linearly with high CPU core counts, eliminating a critical performance bottleneck for dense enterprise clusters.
In this pull request p0lyn0mial enabled the WatchListCompression feature gate by default, introducing transparent gzip compression for WatchList responses. By utilizing a recycled pool of writers via perFlushGzipWriter, this PR resolves previous memory regression bugs. Large-scale benchmarks demonstrate that this change slashes P99 latency by up to 54% and reduces baseline API server memory by 7%. It is a major scalability milestone that trades minor CPU cycles for optimized network throughput.
KEP-2021: HPA supports scaling to/from zero pods for object/external metrics
This KEP enables the Horizontal Pod Autoscaler (HPA) to scale workloads from zero to many replicas and back to zero when using object or external metrics. Since CPU and memory metrics require running pods, scale-to-zero is limited to these metric types. The feature is intended to reduce costs and energy consumption for intermittently idle, resource-intensive workloads (such as GPU-based applications) while still allowing HPA to automatically scale back up when demand returns (for example, when a queue receives new messages). At time of writing, this KEP is listed as Beta in v1.37.
PreemptionPolicy field to PodGroupSpec as part of KEP-5710 workload-aware preemption; the field is gated behind the PodGroupPreemptionPolicy feature gate and follows the same Never / PreemptLowerPriority semantics as per-pod preemption.PodGroup API object to PodGroupInfo so scheduler plugins can obtain a consistent PodGroup state throughout the scheduling cycle.bindAddress in KubeProxyConfiguration; explicit wildcard binds (0.0.0.0 / ::) no longer trigger contradictory recommendations, invalid addresses now get their own warning, and specific non-wildcard IPs still recommend the per-family wildcard.