Online auctions produce one of the most punishing traffic patterns in production: a flat baseline that erupts eighteen-fold in a single five-minute window the moment a high-value lot goes
live, then collapses just as fast. When a high-traffic auction platform hit this pattern during a live event, its Kubernetes cluster scaled, but not fast enough. Pods queued, backend
connections stormed, and requests timed out during the twenty minutes that mattered most.

This is a production post-mortem told with the real telemetry: how the HPA and cluster autoscaler responded, why reactive scaling lost the race against near-instant demand, and how the
stateless tier collided with a connection-limited stateful backend.

Then: what actually worked. Event-driven pre-scaling with KEDA, connection pooling to protect the data tier, warm-capacity trade-offs, and spike-shaped load testing. Attendees running
event-driven workloads will leave with patterns to survive the moment the hammer falls.