A load balancer is usually the first box you draw after the clients, and the depth you can give on it is a fast proxy for how senior you sound. This covers L4 vs L7, the algorithms and their failure modes, health checks, and the mistake of moving your single point of failure instead of removing it.

Every system wants to be fast, scalable, and reliable, which is exactly why saying so is worthless. This post is about the sharper version: p99 under 200 ms, 99.99% availability, eventual consistency on the timeline — and the architectural lever each of those numbers forces you to pull.