IOPS
Input/output operations per second — storage's most quoted performance figure and its easiest to mislead with.
The number means nothing without its context. A mechanical hard drive manages 75–200 random IOPS because every operation moves the arm; an NVMe SSD does millions because nothing moves at all. But the same drive reports wildly different figures depending on block size, read/write mix, queue depth and whether access is random or sequential — so "1 million IOPS" on a spec sheet always means 4 KB random reads at deep queue depth, the most flattering case available.
In practice IOPS is only half the picture. A database cares just as much about latency per operation, and about the tail: a drive with strong average performance but 50 ms at the 99th percentile produces a slow application. Cloud providers sell IOPS as a resource in its own right — AWS gp3 volumes ship with a 3000 baseline and charge beyond it — which turns the metric into a budget line rather than a hardware property.