Running containers at scale has never been simple. Kubernetes 1.37, released this September under the codename “Garhwal,” makes the job more manageable in ways that platform teams will notice almost immediately.
The release packs 67 enhancements, making it one of the larger Kubernetes updates in recent memory. That’s substantial. But the total count matters less than what’s actually in there, and two capabilities in particular stand out for teams running production infrastructure.
Memory QoS is now Beta and enabled by default. Quality of Service for memory, which is the ability to guarantee that critical workloads get the memory they need without being killed or throttled by lower-priority processes, has been a persistent pain point in shared cluster environments. When multiple applications compete for memory on the same nodes, the wrong thing gets terminated at the worst time. Memory QoS gives operators real control over that hierarchy, and enabling it by default means teams benefit immediately without custom configuration.
Pod-Level Resource Managers moved to Beta as well. Previously, resource allocation in Kubernetes required setting CPU and memory limits at the individual container level, which got cumbersome for pods running multiple containers. This feature lets you set resources at the pod level instead, simplifying configuration and eliminating a common source of mismatched declarations that cause silent performance issues. Cleaner config, fewer surprises.
For organizations running AI or machine learning workloads on Kubernetes, this release is worth a careful look. Pay attention here. The 1.37 improvements include targeted enhancements for GPU-intensive and memory-heavy training jobs, and as more engineering teams move their ML pipelines into Kubernetes-managed environments, native controls for resource allocation and scheduling stop being optional.
The release also stabilizes the Metrics API. Metrics collection, meaning the process of gathering performance data from across your cluster, has historically produced inconsistent results across Kubernetes versions. A stable Metrics API gives monitoring tools like Prometheus and Grafana a reliable interface that won’t shift underneath them during an upgrade. That matters for any team trying to maintain observability through a version transition.
On top of the 1.37 release, Karmada graduated as a CNCF project this month. Big deal for multi-cloud teams. Karmada lets you manage Kubernetes workloads across multiple clusters and cloud environments from a single control plane, and that graduation comes with stronger community backing and a long-term support signal that makes it easier to justify in production architecture decisions.
If you haven’t started your 1.37 upgrade planning, the Memory QoS and Pod-Level Resource Manager improvements together make a strong case for moving soon. Start planning now. Your cluster stability will thank you.
Want to explore how Kubernetes and modern DevOps practices could benefit your business? Let’s talk.