"...service handled a peak load of 170k QPS with 40 machines running at 35% CPU usage on NYE 2015. The response time was < 5 ms at 95th percentile, and < 50 ms

"...service handled a peak load of 170k QPS with 40 machines running at 35% CPU usage on NYE 2015. The response time was < 5 ms at 95th percentile, and < 50 ms at the 99th percentile." It's ~4k per instance. I would like to know what the R/W ratio is. And how often the index is updated is the key, aka locking operations, which should be mentioned in the article. Also this article should give more comparison on other implementations, if possible. At least compare with Uber's monolithic implementation. The geo-datastructure part is very cool.