Client-Side Load Balancing at a Million Requests Per Second

How we built an in-process client-side load balancer for a million requests per second of internal fan-out traffic, what we layered on top (N-ring fade-in, occupancy-based bounded load, and AZ-aware routing with a latency health factor), and how hardening that path cut cost and made the service resilient to the infrastructure underneath it.