Predictions should come from **empirical observation of your own site's real usage**, not from benchmarks made in artificial environments. Benchmarking and performance research have value, but must not be the sole indicators of capacity — real-world observations outweigh any theoretical measurement. The corollary is relentless instrumentation: *measure, measure, measure*. Measurement is a necessity, not an option — you put a fuel gauge on your car's dashboard for a reason; don't run your systems without one. Industry-tested tools for CPU, memory, and disk usage can be repurposed to measure application-level and business metrics too. This is why a metric must be treated as a *proxy* and validated against reality: a benchmark that doesn't reflect how your system is actually used will mislead the plan. --- *Source: [[The Art of Capacity Planning]] (John Allspaw, O'Reilly 2008) — Ch 1 — Goals, Issues, and Processes in Capacity Planning*