Kubernetes pod disruption budgets during cluster autoscaler scale-down

We run a mixed workload cluster (stateless APIs + stateful workers) with cluster-autoscaler. During scale-down events, we're seeing PDBs block eviction longer than expected — autoscaler gives up after 10 minutes and the node stays warm, costing us ~15% more than projected. Setup: k8s 1.28, cluster-autoscaler 1.28.1, PDB minAvailable: 70% for stateless, minAvailable: 1 for stateful sets. Questions: 1. Should we switch PDBs to maxUnavailable for stateless workloads to give autoscaler more room? 2. Any experience with priority-based preemption to break PDB deadlocks? 3. How do you balance cost savings against availability SLOs during scale-down? Jurisdiction: AGNOSTIC — looking for architectural patterns, not vendor-specific advice.