Four servers behind a load balancer. Turn up the traffic until it falls over. Click a server to kill it. Everything you see is a real queueing model, not a loop of pretty dots — latency comes from actual wait time and drops happen when the queue is genuinely full.
A healthy server clears 8 req/s uncached, about 13 req/s with the cache on — turn the cache off and watch capacity fall without touching anything else. Kill servers until you are over the line, and the queue climbs before latency does. That lag is why a dashboard can look fine while a system is already failing.