Redis was at 50% CPU. It was already too late.

A production support question — "our self-managed Redis was at 50% CPU on a 4-core box, but latencies were already climbing, how do we know when to scale?" — turned into a real benchmark investigation for anyone self-managing a database in production, on both a standalone instance and a real primary/replica pair. Here's the metric that actually matters, the thresholds to alert on, and when to scale up vs. scale out.