This is implemented using the same concurrency mechanism as for the
replication: the work is split in slices that should be small enough
to avoid slowing down network requests significantly.