Peak traffic is not an exception. For real-time services, it is the product.
A goal, a delayed train, a major odds movement or a final whistle can send thousands of users to the same endpoint in seconds. Infrastructure has to absorb that burst without losing the state users care about.
Horizontal scaling helps only when sessions, queues and databases are designed to scale with it. Stateless application layers, caches with clear invalidation rules and controlled back-pressure can prevent a short spike from becoming a system-wide incident.
Observability matters just as much as capacity. Teams need to see whether the bottleneck is compute, storage, network, a third-party feed or a single slow query before customer support becomes the monitoring system.
“Reliability is the ability to stay predictable when demand stops being predictable.”