We recently added a requirement to rollback_to_stable that a checkpoint needs to be completed prior to it being called. It turns out that's not a reasonable requirement - for in-memory deployments need to be able to run rollback_to_stable and never create checkpoints.