When doing a deploy_start, we read the state of all dependent resources from the DB. We could also read this from the scheduler itself. Once the new code is proven sufficiently stable, we should also make the switch here.