What's this about?
An engineering writeup claiming a 99th-percentile latency of 0ms for autocomplete across 240 million domain names, achieved by prefetching: when you type a character, it fetches results for every plausible *next* character, so the following keystroke is answered locally. The asterisk in the title is doing work, and the comments found it.
What is P99 and why is the measurement disputed?
P99 means 99% of requests complete faster than this — the standard way to describe latency, because averages hide the slow tail users actually notice. The dispute is where the clock starts. The implementation renders on keyup, when you release the key. Commenters argue perceived latency starts at keydown, when you press it, and that measuring from keyup both understates the number and, worse, delays the render.
What's the community saying?
- "Clever, but that's not how we measure latency" is the top comment and sets the whole thread.
- The defence is that perceived latency is what matters for UX, which the critic turns around: rendering at keyup makes perceived latency strictly worse.
- The strongest counter-argument is practical — keydown rendering breaks on Ctrl+V, accented characters, several Asian input methods, backspace and held keys.
- The rebuttal that ends it: every UI you can test types the letter on keydown, and Google's autocomplete updates while you hold a key down, so the edge cases are clearly solvable.