A load balancer in front of an auto-scaling group. Drive traffic at it, kill instances under it, and watch what the autoscaler can and cannot catch in time.
Why the autoscaler is always late. Scaling reacts to CPU that is already high, then a new instance boots before it serves anything. Set warm-up to 45 s, hit Spike, and watch the 5xx hole open and stay open for most of a minute. Drop warm-up to 0 s and the same spike barely marks the chart. Capacity you have to wait for is capacity you do not have.
Model. Each instance serves 24 concurrent requests at 45 ms of service time, so about 530 req/s before queueing. Latency is service time plus queue delay; requests past a half-second of queue are shed as 503s. CPU is smoothed over 10 s, which is part of why scaling lags. Health checks run every 5 s and evict after two failures. Compute is costed at $0.085 per instance-hour, billed from boot. Figures are simulated, not measured from a real fleet.