Continuous Profiling in Production: What, Why and How

What is continuous profiling?

Performance is crucial to make the system much cheaper. We always found profiling in production an incredibly laborious process. Every time we did it it was a huge win. But we do it very frequently because of how laborious and how risky it is. Engineers don't like going into production doing things because basically, they get the wrong people shouting at them.

Are you going to be talking about specific tools or strategies for implementing tooling?

We talk about different information coming out of existing monitoring systems in terms of metrics and how instrumentation works. We look at some profiling techniques which are JVM specific but would apply to any language that's running on the JVM, Java, Scala and Clojure.

How focused on the JVM is this talk?

A lot of our discussions about different metrics and implementation tradeoffs are pretty much applicable to anyone running a software system. JVM specific is when we talk about how you can do low overhead JVM production profiling.

What do you want people to leave the talk with?

One thing that people often do when they diagnose performance problems is basically reproduce those performance problems in a development environment and then try to fix that. People do that because they don't have enough information to see what their production system is doing. One of the big takeaways is that you don't want to be doing that. It can often be very difficult to do, very time consuming to get an accurate reproduction. Your software stack, your hardware and software, can all be very different in your development environment than in production. You might fix the wrong problem.

What do you feel is the next potentially disruptive technology for software?

Where we are on the AI adoption process in terms of technologies, we're very very early on. Even though it's a current disruptor it feels like it will still be a huge disruptive factor for the next 10 to 15 years.