I ported a service from Go and the median request latency barely moved, but the tail did, and here is a week of flamegraphs explaining exactly which part of the old runtime was responsible for it

p50 went from 11ms to 9ms, which is noise. p99 went from 340ms to 24ms, which is the garbage collector, and it is the only number anybody on the support rota ever felt.

2 comments

Counterpoint: the old way still wins if you only ever have one machine.

0

This matches what I saw, though my numbers were about half as dramatic.

0