Also worth mentioning that performance is much more predictable when the data fits into memory (or the working set, but that may be harder to convey).