Kubernetes has introduced significant changes to resource management over the last three releases. Since version 1.35, container-level resource updates are stable, and as of version 1.37, released on August 26, 2026, pod-level resource management has been stabilized. These updates allow for modifying resource requests and limits for running pods without requiring a restart.
The Kubelet now natively supports Pressure Stall Information (PSI), which reached general availability in version 1.36. Unlike traditional utilization metrics, PSI measures the time tasks spend waiting for resources like CPU, memory, or I/O. This provides a more accurate signal for performance bottlenecks, as high utilization does not always correlate with application latency or throttling.
Resource changes are managed via the resizePolicy field, which determines whether an update requires a container restart. While CPU limits can often be adjusted on the fly, memory adjustments may require a restart depending on the runtime, such as with JVM or CPython, where heap sizes are fixed at startup. Changes are applied via a specific subresource patch.
The platform now supports defining resource budgets at the pod level, allowing containers to share a common pool. This replaces the previous requirement of setting individual limits for every container, which often led to OOM errors or inefficient resource allocation. However, static CPU managers and Windows nodes do not support in-place resizing.
To implement these features, clusters must run on Cgroup v2 with a Linux kernel version 4.20 or newer that supports PSI. Users should verify their runtime compatibility, as outdated versions of containerd or CRI-O may fail to apply updates. Monitoring should focus on PSI metrics rather than utilization to better align with user experience.
