LoginSign Up
← Back to all sparks
Kubernetes logo

Kubernetes

DEVOPSINFRA · APIS
Velocity7.5

Container orchestration

kubernetes.io ↗

Kubernetes v1.37 matures its memory management and scheduling stack for AI/ML workloads.

resource-managementai-workloadsschedulingobservabilityfeature-graduation
◆Current state
Kubernetes v1.37 is completing a systematic maturation pass across resource management, scheduling, and observability. Memory QoS is now enabled by default on cgroup v2 nodes; native histogram support lands in beta; the Node Lifecycle Conditions API gives operators a structured vocabulary for node health beyond readiness taints. This is a hardening release, not a surface-area expansion.
◆Where it's heading
v1.37 signals a deliberate push to make Kubernetes a first-class substrate for AI/ML workloads: DRA Extended Resource support at GA, workload-aware scheduling advances, and in-place pod resize preemption all address the scheduling and resource isolation patterns that large training and inference jobs require. The next cycle will focus on pushing these features from beta to GA and expanding their scope.
◆Prediction
DRA and rootless mode will both reach GA in v1.38, closing the current AI-workload resource isolation wave; HPA scale-to-zero will advance toward stable API status.

◆Recent moves

  1. 18h ago

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

    Pod-Level Resource Managers graduate to Beta in v1.37, disabled by default. This builds on the alpha introduced in v1.36, allowing coordinated resource management at the pod level rather than requiring per-container workarounds—relevant for AI/ML jobs requiring tightly coupled resource allocation.

    View source ↗
  2. 1d ago

    Kubernetes v1.37: Memory QoS Graduates to Beta

    Memory QoS graduates to Beta and is now enabled by default in v1.37 on cgroup v2 Linux nodes. The memory controller gives the kernel better guidance on allocation priority, which reduces the chance of well-behaved workloads being OOM-killed when the node is under pressure.

    View source ↗
  3. 1d ago

    Kubernetes Changed Block Tracking API - Beta Differences

    Changed Block Tracking (CBT) support for CSI drivers moves to Beta with the external-snapshot-metadata v1.0.0 release, enabling incremental backup scenarios that only transfer changed blocks rather than full volume snapshots—a meaningful efficiency gain for stateful workloads.

    View source ↗
  4. 4d ago

    Kubernetes v1.37: Native Histograms Graduates to Beta

    Native histogram support for Kubernetes metrics graduates to Beta and is enabled by default in v1.37. Native histograms consume significantly less storage and network than classic histograms at equivalent precision, which matters at the cardinality levels typical Kubernetes clusters generate.

    View source ↗
  5. 5d ago

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

    Scheduler preemption for in-place pod resize ships as Alpha in v1.37. When a pod needs more resources and the scheduler can't find a node with sufficient free capacity, it can now preempt lower-priority pods to make room rather than blocking the resize—completing the in-place resize story for priority-aware clusters.

    View source ↗
  6. 6d ago

    Kubernetes v1.37: Introducing Node Lifecycle Conditions

    Node Lifecycle Conditions introduces a new structured API for expressing fine-grained node health states—covering hardware faults, OS-level conditions, and provider-specific signals—alongside the existing readiness and taint mechanisms. This consolidates node observability into a single queryable surface rather than scattering signals across taints, labels, and annotations.

    View source ↗