Running Spot Instances with Karpenter can cut Kubernetes compute costs by 59–77%, but requires correct configuration to avoid reliability pitfalls. Key practices include setting capacity-type requirements in NodePool specs, allowing multiple instance families for wider fallback options, and using EC2 Fleet's Price Capacity Optimized strategy. Karpenter handles interruptions natively via an SQS queue and five EventBridge rules — do not run AWS Node Termination Handler alongside it. A two-NodePool pattern (spot at weight 100, on-demand at weight 1) provides automatic fallback when spot capacity is exhausted. Additional guidance covers PodDisruptionBudgets, terminationGracePeriod alignment with the 2-minute reclaim window, multi-AZ topology spread, and the do-not-disrupt annotation for long-running tasks. ML-based spot selection (Cast AI's Intelligent Spot) can proactively reduce interruptions by 23–94% by routing workloads to historically stable pools.
