Profiling in production

Let's talk about how Odnoklassniki comes to writing profilers for a production.

In the beginning, you will hear about why you should do a performance at all. Next, the required steps: graphs, profiling locally, tests, simple profiling in the production. Kirill will tell why this wasn't enough and his team had to dive deeper.

In the end — a story about how Systrace is designed. What profit can be obtained from this knowledge? How to write your profiler similar to Systrace? And how to make Systrace work in production and most importantly why?