Batch inference workloads can be configured to tolerate preemption on one minute to one hour of advance notice