Your team runs a streaming Dataflow job that performs windowed aggregations over a high-volume Pub/Sub stream. The job uses large amounts of persistent disk on each worker to hold windowing and shuffle state, and you have noticed that autoscaling is slow because new workers must transfer state before contributing to processing. You want to reduce disk costs, speed up autoscaling responsiveness, and offload state management from the worker VMs without rewriting your pipeline logic. What should you do?