The failure I've seen multiple times is that the database is returning data within normal latencies, and then there is a traffic tipping point and the latencies go up 1000x for all requests.