Fix Kubernetes CPU Throttling & CFS Quotas
The P99 latency alert fires while Grafana reports 20% CPU usage. Welcome to Kubernetes CPU throttling , where Linux CFS Quotas forcibly freeze containers in microscopic 100ms windows despite low average usage. Here is the complete SRE blueprint to detect micro-freezes, tune runtimes, and bypass CFS limits using bare-metal static core pinning. Phase 1: The Dashboard Illusion (100ms CFS Quota)…
Kubernetes CPU throttling occurs when Linux's Completely Fair Scheduler (CFS) forcefully freezes containers during 100-millisecond windows, even if average CPU usage is low. Dashboards often show misleadingly low average CPU usage, masking the microscopic freezes. The CFS limits are enforced using 100ms windows, granting execution time based on the limit set (10ms for 100m, 50ms for 500m).
When a bursty, multi-threaded application uses its allocation within the first 10ms, the kernel freezes the container for the remaining 90ms. To detect throttling, track the percentage of 100ms windows where the kernel parks the application using PromQL. If the Throttling Ratio exceeds 15–25%, artificial latency spikes are likely.
Java and Go runtimes can amplify the issue by inspecting the host OS for cores and spawning a large number of threads. On bare metal nodes, use dedicated physical CPU core pinning for containers to bypass the CFS quota system. Set CPU requests equal to limits in the Guaranteed QoS class to ensure containers are pinned to dedicated cores, effectively eradicating throttling.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
