Recommendations for using interruptible/spot nodes

We are currently evaluating Ray Cluster on Kubernetes on AWS and would like to know if there are recommendations around using spot nodes. Specifically, are there any known methods for spilling objects to S3 or external storage and ensuring that the Object Ref is owned by the Head node or any other Actor on a small on-demand node? We are currently using the RayDatasets API (map_batches()) which is creating objects and we cannot find a way to specify their owner.