If a user scales to 100 and then down to 50 immediately, prefer to "dequeue" the pending replicas instead of tear down the running ones.