Cong thanks for the note! I am yet to come across an implementation that scales to millions while keeping latency to 10-15 ms. Also, given how hard (impossible) it is to measure one-way latency, this problem is a wonderful conundrum.