Starting containers can be done in parallel, but scheduling cannot be trivially parallelized because it needs to take into account the resources consumed by existing containers. Situations like a rack failure can cause scheduling storms where many workloads need to be rescheduled as quickly as possible.