Systems on Glass is an AI-automation side project of mine. Every episode is generated by Claude (Anthropic's AI, working through Claude Code in my own video studio): the script, the archival slides and 3D scenes, the edit, and the narration in an AI clone of my voice. I choose the topics, bring the stories from my own work, and review each episode before it goes up. Not affiliated with Anthropic or Amazon.
This June, on Prime Day, one Amazon database peaked at 192 million requests a second. A wave like that doesn't wait for your servers to wake up. When I was at Amazon, we didn't react to it. We were ready: a month before Prime Day, we tripled everything.
The short version
- A load balancer is the door. It takes every request and hands it to one of many servers, so no single server takes the whole wave. More servers behind it means more capacity: that's scaling out.
- Autoscaling reacts. It sees the load, then starts servers, and a new server still has to boot, load its code and fill its caches. AWS's docs suggest starting with a 300-second warm-up. On Prime Day, those minutes are the outage.
- So pre-scale. A traffic team forecast the peak a month out. Days before, we scaled our team's systems up about 3× and let them warm. On AWS today you can even reserve load balancer capacity ahead of a surge (LCU Reservation).
- Keep autoscaling on top. The forecast is a floor, not a promise: autoscaling stays on for whatever it missed.
- Write the last line first. An idle server still bills by the hour, so the plan ends with scaling back down after the peak.
my team's plan: forecast at T−30 days · pre-scale ~3× days before · autoscale on top · scale down after
AWS's rehearsal for 2026: 44,000+ fault-injection experiments, six times 2025's
The 3D scene and the forecast graph are my own illustrative recreations, not Amazon's systems or numbers. The Prime Day figures are from AWS's own posts.
Further reading
- All the numbers: Amazon Prime Day 2026 powered by AWS — the 192 million a second, and the 44,000 rehearsed failures.
- Set the default instance warmup (EC2 Auto Scaling) — why a new server isn't ready the moment it starts.
- How predictive scaling works — scaling from a forecast, with instances launched early.
- Load Balancer Capacity Unit Reservation — pre-warming the load balancer itself before a surge.
Next episode: How tool calling works. Previously: Cache the Model. All episodes →