Pressure Stall Information (PSI) metrics reached general availability in Kubernetes v1.36, with the announcement landing on the project blog on May 12, 2026.
What PSI actually measures
PSI surfaces percentages of time tasks are stalled on CPU, memory, or I/O — not how much resource was consumed, but how much time was lost waiting. The kernel has exposed this since 2018; Kubernetes now wires it through the kubelet at node, pod, and container granularity. Each metric ships as a cumulative total plus moving averages over 10s, 60s, and 300s windows, distinguishing transient spikes from sustained contention.
Production overhead
SIG Node measured kubelet collection overhead consistently below 2.5% of total node capacity, with kernel-level tracking between 0.925% and 3.125%, validated under high-density workloads of 80+ pods.
Why this matters
PSI gives you a first-class upstream answer to “is this pod actually being starved?” without resorting to eBPF tooling or vendor agents. If you’ve been reading CPU throttling counters to debug noisy-neighbor problems, you now have a directly applicable signal.
Source: Kubernetes v1.36: PSI Metrics for Kubernetes Graduates to GA — May 12, 2026
Stay on top of the cloud-native release wire
Kubernetes, AI infra, and CNCF moves — delivered when they matter.