Node Resource Pressure, Eviction, and Recovery

A node hits 95% memory usage. The kubelet starts killing pods. Which ones die first, and can you control the order? The answers live in the kubelet's eviction manager, a control loop that monitors resource signals, compares them against configured thresholds, and evicts pods in a predictable sequence when a node runs low on memory, disk, or PIDs.

The eviction manager is the last line of defense before the Linux OOM killer intervenes. When it works correctly, the right pods get terminated gracefully and the node recovers. When it is misconfigured, critical workloads die while idle pods survive. The CKA Troubleshooting domain (30% of the exam) expects you to understand this mechanism, diagnose evictions, and configure thresholds.