Tuning a system for optimal performance is hard; maintaining that optimal configuration in the face of frequent changes even harder. Micro-services architectures in production are subject to constant churn as updates and refreshes occur to layers in the stack and in the data centers. Keeping up with these changes requires automation.
At Twitter, we have been working on a service called Autotune. Autotune leverages an existing Twitter service that uses a technique called "Bayesian Optimization" to automate the tuning of microservices. Our eventual aim is to continuously tune services in production, maintaining near-optimal performance at all times. In this talk, we'll go through some bayesian optimization basics, describe how Autotune works, and how service owners interact with it to tune their systems. We conclude with some results, challenges, and some of the problems that still need to be tackled.
Gloria is a software engineer at Twitter where she works on automating tuning of microservices. In college she studied computer science and philosophy, and has an interest in technology ethics. When she's away from the keyboard she enjoys practicing yoga, road biking, and rock climbing.
