Kubernetes in 2026: What changed for platform teams

The major Kubernetes changes through November 2026, from Ingress NGINX retirement and Gateway API to autoscaling, scheduling, and upgrade planning.

Ingress NGINX reached retirement, Kubernetes gained new scaling and scheduling capabilities, and several older networking features are on their way out. For anyone running a cluster in 2026, two questions matter. Which changes make day-to-day operations easier, and which dependencies need attention before the next upgrade?

As of November 20, 2026. Kubernetes 1.36 was released on April 22 and 1.37 on August 26. The release team is targeting December 16 for Kubernetes 1.38. This review of shipped features therefore covers releases through 1.37. Kubernetes 1.36 · Kubernetes 1.37 · Kubernetes 1.38 release schedule

Networking: Ingress NGINX retirement changes upgrade planning

Maintenance of the community Ingress NGINX project ended on March 24. Existing installations continue to run, but they receive no bug fixes or security updates. Teams switching controllers should check which features they currently configure with controller-specific annotations and how to reproduce that behavior in the target environment. The project released Ingress2Gateway 1.0 as a starting point for migration. Kubernetes 1.36 · Ingress2Gateway 1.0

The Gateway API also matured. Version 1.5 brought TLSRoute into the Standard channel, and version 1.6 followed with TCPRoute and UDPRoute. That gives teams standardized resources for more types of traffic. The routes available in a particular cluster depend on its Gateway controller. Gateway API 1.5 · Gateway API 1.6

Two Service routing changes are worth planning for. Kubernetes 1.36 deprecated the Service field .spec.externalIPs. The earliest version in which kube-proxy is expected to disable its support by default is 1.40. In November, SIG Network set out a more specific plan for kube-proxy's IPVS mode, with removal currently planned for 1.40. Teams using IPVS have time to test nftables or iptables with their networking configuration. Service ExternalIPs · IPVS

Scaling: Workers can go down to zero

In Kubernetes 1.37, scale-to-zero for the Horizontal Pod Autoscaler is a beta feature enabled by default. It needs a suitable object or external metric that remains available while no Pods are running. Queue length is one example. When the queue is empty, a worker Deployment can scale down to zero replicas. When new tasks arrive, the autoscaler can bring replicas back. Test how reliably the metric is available and how long the first Pod takes to start processing work. HPA Scale-to-zero

Vertical scaling also moved forward. Changing CPU and memory for individual containers in a running Pod became stable in 1.35 at the end of 2025. 1.36 brought the corresponding capability for shared Pod-level resources into beta. Platform teams can test those two cases separately when evaluating live resource changes. In-Place Pod Resize in 1.35 · Pod-Level Resources in 1.36

Scheduling: Kubernetes can schedule related Pods as a group

A distributed training job may need all its workers at once. When some Pods start and the rest wait for capacity, the job can occupy resources without making progress. In Kubernetes 1.37, the Workload and PodGroup APIs for gang scheduling reached beta. The scheduler can wait until there is sufficient capacity to place the entire group. The core features of this scheduling model require explicit enablement in 1.37. Workload-Aware Scheduling

Support for traditional Extended Resources in Dynamic Resource Allocation (DRA) also became stable in 1.37. DRA drivers can now fulfill those requests. This is useful for teams that want to request GPU resources through the familiar interface and manage allocation with DRA. DRA

Linux nodes, security, and storage

For Linux nodes, cgroup v2 is now a practical upgrade requirement under the default configuration. Since Kubernetes 1.35, the kubelet has refused to start on cgroup v1 nodes by default. The project explained the transition in detail in October. Memory QoS for cgroup v2 nodes reached beta in 1.37, giving operators more control over how the kernel protects memory under pressure and when it throttles workloads. cgroup v2 · Memory QoS

Several workload isolation and identity features became stable. User namespaces and fine-grained kubelet API authorization reached that stage in 1.36. Pod certificates and ClusterTrustBundles followed in 1.37, making it possible to deliver certificates and trust anchors to Pods. Teams sharing a platform should check which of these features their cluster and runtime support. User Namespaces · Kubelet API Authorization · Pod Certificates and Cluster Trust Bundles

Volume Group Snapshots became stable in 1.36. With a compatible CSI driver, related volumes can be snapshotted as a group. Kubernetes 1.37 added beta tracking for how long PersistentVolumeClaims have been unused. That helps operators review old volumes. Whether a volume can be deleted still depends on what the business uses it for. Volume Group Snapshots · Kubernetes 1.37

Control plane and tooling

The Metrics API for CPU and memory usage became stable in 1.37. The control plane also gained beta support for etcd RangeStream, which reduces memory use for large list requests. Teams using the classic Kubernetes Dashboard have a migration to plan as well. The project was archived, and the Kubernetes blog explains how to move to Headlamp. Metrics API · etcd RangeStream · Headlamp

What to check before the next upgrade

For most platform teams, the first step is to inventory what is running. Is Ingress NGINX still installed? Does kube-proxy use IPVS? Do any Services set .spec.externalIPs? Do any Linux nodes still use cgroup v1? The answers tell you which changes need testing first. You can then try new features on specific workloads, such as scale-to-zero for a queue worker or gang scheduling for a batch job that needs its Pods to start together.

Through November, the common thread is that Kubernetes can express more precisely what resources a workload needs, when its Pods need to start together, and how far it can scale down. Upgrades also require a closer look at the older components a cluster still relies on. The December release is the next point at which to revisit that inventory against the changes that have shipped. Kubernetes 1.38 release schedule