A ceiling has two parts that can both go stale: its *value* and the *metric* it's measured in. Everyone expects the value to shift with hardware. The subtler trap is that the **defining metric** can change entirely. Allspaw's example: a database's limiting factor is disk I/O — until you upgrade to a faster disk subsystem, after which the bottleneck becomes the single gigabit network card. The old red-line metric is no longer the one that fails first. So picking the right metric to follow is itself hard (not all bottlenecks are obvious), and it must be re-validated after any hardware or architecture change. This is the forecasting-side companion to "adding capacity moves the bottleneck" — and a reason diagonal-scaling decisions (more spindles, more RAM, SSDs) should be re-checked against *which* resource now binds, not just how much of the old one you have. Re-confirm the binding metric whenever the platform changes. --- *Source: [[The Art of Capacity Planning]] (John Allspaw, O'Reilly 2008) — Ch 4 — Predicting Trends*