It looks like Adrian Klaver had the right answer. Burstable instances are not suitable for these kinds of heavy tasks, even though they don't take hours. We temporarily moved the system to a plain EC2 and never saw the problem again. We are now in the process of migrating back to RDS.