Kubernetes 1.37: Workload-Aware Scheduling and native histograms reach Beta; HPA scale-to-zero

Kubernetes 1.37 just shipped a set of changes that will force platform teams to touch two things they normally treat as orthogonal: scheduling semantics and telemetry.

The single most consequential runtime change is Workload-Aware Scheduling (WAS) moving important primitives into Beta: a PodGroup-like scheduling unit, workload-aware preemption semantics, and shared ResourceClaims for grouped pods. This isn’t a neat feature checkbox — it rewrites how you reason about multi-pod jobs and preemption in clusters where pods must start together.

Why this matters

Gang scheduling has always been an awkward fit for Kubernetes. Libraries and operators built their own PodGroup-like CRDs and crude admission gates to assemble collections of pods. With WAS Beta, those patterns become first-class: a PodGroup-like object can represent a scheduling unit, ResourceClaims can be shared across that unit, and the scheduler can preempt with group semantics rather than on a per-pod basis. That gives you coordinated startup, more predictable placement for MPI/parameter-server workloads, and consistent preemption behavior without brittle operator hacks.

If you run ML training, distributed inference, or any workload that requires N-out-of-M pods to be scheduled together, this is the right call from upstream. The alternative — ad-hoc gang-scheduling implementations spread across repos — was unsustainable. But it also means existing workload operators should be audited: controllers that assume per-pod preemption or that inject ephemeral ResourceClaims will behave differently once clusters adopt WAS.

Native histograms: control-plane telemetry that scales

Kubernetes 1.37 also promotes native histograms to Beta. Practically, that changes how high-cardinality metrics and bucketed histograms are represented and transmitted in the control plane and Kubelet metrics pipeline. Instead of shoehorning bucketed values into label-heavy Prometheus metrics, native histograms allow more compact encoding and server-side efficiency.

This improves control-plane performance and reduces metric scraping overhead, but it will break assumptions in monitoring stacks that expect the old encoding. Dashboards that aggregate by label or rely on histogram-to-summary conversions will need updates. If you haven’t audited your exporters and alerting rules for histogram semantics, do it now; native histograms are the right long-term move, but they’re not a drop-in replacement.

HPA can scale to zero (Beta)

Also in 1.37: HorizontalPodAutoscaler scaling to zero is Beta. That’s a practical win for event-driven and bursty services that want cost-efficient zero-to-one lifecycles without a separate scale-to-zero controller. It’s limited to HPA configurations using object or external metrics, so don’t expect replacement of Knative-like activation semantics for every case, but it reduces one class of custom autoscaling glue.

Runtime security: containerd patching and credential hardening

Parallel to the Kubernetes churn, containerd maintainers published security and hardening patches across supported branches to address credential-persistence and related issues in distribution and CRI flows. The fixes avoid persisting registry credentials during image pulls and CRI interactions. If you run containerd on any supported branch, prioritize upgrading to the latest patch release; leaked registry credentials are a trivial escalation path.

Operational takeaways

Start mapping which teams run gang-like workloads and which controllers assume per-pod preemption. Expect behavioral differences once WAS is enabled.
Audit dashboards, exporters, and alerting rules for histogram semantics; update queries to use native histograms where appropriate.
Roll the latest containerd patch releases across clusters — the auth-data hardening fixes are low-risk and high-payoff.

The v1.38 cycle is already rolling toward KEP readiness and enhancement freezes, so expect faster stabilization on these APIs. The deeper signal here isn’t just three features landing: it’s that the project is treating scheduling primitives and control-plane telemetry as first-class system-level concerns, not optional add-ons. That’s overdue, and it’ll make clusters more predictable — but only if platform teams actually update their controllers, dashboards, and runtimes instead of pretending these changes are purely cosmetic.

Kubernetes 1.37: Workload-Aware Scheduling and native histograms reach Beta; HPA scale-to-zero

Sources

Kubernetes 1.37.0: Scheduler Preemption for In-Place Pod Resize (Alpha) and patch-maintenance

Kubernetes v1.37: KubeletInUserNamespace (rootless), HPA Scale-to-Zero, and Workload-Aware Scheduling to Beta

Kubernetes DRA GA: Dynamic Resource Allocation Reaches GA and Workload-Aware Scheduling Advances