Peak hour is not a traffic spike. It is a coordinated burst of geo lookups, assignment attempts, payment retries, and chat messages. If any one of those queues backs up, the whole floor feels it.
We design dispatch infrastructure around p95 latency, not average load. That means isolating realtime paths, caching zone geometry, and making payment failures observable before riders abandon.
The boring work — autoscaling policies, dead-letter queues, restore drills — is what lets a Friday night look uneventful. That is the point.