Karpenter's disruption system uses three mechanisms to recover wasted cluster capacity: consolidation (removes underutilized/empty nodes), drift (replaces nodes whose spec diverges from NodePool/EC2NodeClass), and expiration (TTL-based drain that bypasses disruption budgets). Drift detection is event-driven and typically resolves within ~5 minutes. Disruption budgets limit concurrent node replacements per NodePool, defaulting to 10% if unset; setting nodes to '0' blocks voluntary disruption but NOT expiration. The do-not-disrupt annotation protects individual pods or nodes from voluntary disruption, but a NodePool terminationGracePeriod can override it. For stateful workloads, combining PodDisruptionBudgets, WhenEmpty consolidation policy, and AZ-aware topology constraints prevents EBS reattachment failures and unexpected pod migrations. expireAfter is unsuitable as a primary AMI rotation strategy for workloads with strict maintenance windows — drift-triggered rotation with conservative budgets is safer. Practical YAML examples and kubectl commands for verifying and pausing disruption are included.
