170,000 people showed up. The servers stayed up.

You can load-test all year, but nothing prepares you for a real peak-demand event. Black Friday Brazil put 170,000+ live users on infrastructure my team of 10 built and operated. What held it together wasn't one heroic optimization. It was a hundred unglamorous decisions: aggressive caching, queue-based back-pressure, killing every N+1 query we could find, and rehearsing failure modes before they happened. The night of the event, the most useful thing I did was watch dashboards and do nothing. If your systems need heroes at peak, the architecture failed earlier.