Replying to @sayrer
@sayrer We don’t go into it in detail in this talk, but it’s less ML and more the fine-grained nature of Twitter’s services.
Ian Brown @igb
@sayrer Even a random parameter value generator would have some success.
View on Twitter ↗
0 1

Replies

@sayrer Because the small scope and focus of each service makes it easy to create an objective measurement of performance.
@sayrer Example: "number of user lookups per minute of GC"
@sayrer That plus systems like Aurora/Mesos make it easier to spawn experiments and clone/mutate service configs.
@sayrer Anyway, check out @ysr1729’s work in this area for more detail. He’s been the driver of the bulk of our ML investigations.