Honestly, the main reason is that I previously worked at Dropbox and knew we could reliably throw ~petabytes at MySQL. Or at the very least, when we started scaling I could call my friends from Dropbox and Facebook to help out. :)

Most of our complexity is in the sync later, so we spend more time tuning our Python sync server. For example, we wrote a custom statistical profiler: https://nylas.com/blog/performance