This article explains the integration of Pressure Stall Information (PSI) metrics into Kubernetes v1.36 for monitoring resource contention. It details the transition from traditional utilization metrics to task-based stalling metrics and presents performance validation results demonstrating that the feature adds negligible overhead to Kubelet operations.
Key points
PSI metrics provide a superior alternative to utilization metrics by identifying task-level resource contention rather than simple percentage usage.
PSI reports cumulative stall time and moving averages, enabling operators to differentiate between transient spikes and sustained system tension.
Performance testing confirmed that collecting PSI metrics via Kubelet introduces minimal overhead, consuming less than 2.5% of node capacity.
The Kubelet's PSI collection logic integrates into standard housekeeping cycles, ensuring safe execution for high-density production environments.