Agreed. Scaling based on the length of the queue, up to some maximum.
Even better, based on queue latency instead of length