Presentation: Develop Hundreds of Kubernetes Services at Scale With Airbnb

Our biggest problem right now is trying to scale both our engineering productivity as well as the availability of our site. It involves breaking apart a giant application into services, and applying the best practices available. Previously we had all these bare metal instances on AWS configured with Chef. We're moving them to a more scalable SOA solution that's backed by Kubernetes, and there were some limits we hit with our old infrastructure around service discovery and stuff like that. I would consider us a fairly early adopter, and we did run into some issues. What I'm working on is solving these issues for our engineers, and more broadly for the community.

The goals of the talk are to go over broad level things that we think you should watch out for as well as very specific things. If there are specific things we got stuck on, I'll just throw it out there. There are lots of examples about that like certain ways we annotate our pods, stuff like that. I'll try to get low level when I can, when I don't think it's worth speaking in vague terms, but also address big ideas like, what happens when we store infrastructure this way or this way. What are the problems when everything is configured through a console, or as code. When we configure as code we get a bunch of validation and continuous integration for free. Also, it's really powerful for us as being able to refactor all of our projects. One thing I want to dive into is our refactoring tool and how important that has been for all kinds of migrations. Many companies have a lot of overhead, updating security, patches, software and hardware reaching end-of-life. All migrations, even big ones, can be automated in some way. We have a Kubernetes cron job script that can update every single service. We have over 500 services now. It gets to a point where you can't do this manually or if you're doing it manually you're not doing it on a very good time frame.

I believe the talk is useful both to people using Kubernetes, and for infrastructure engineers trying to empower other people. One of the other ideas I'd like to drive is that it’s very important to you to automate your own workflow. I don't want this to be about automating things for people who don't understand kubernetes or for people who don't want to be bothered. A lot of the automation that we do is stuff that I personally use. It's abstractions that are useful for everyone.

I want them to walk away with a deeper understanding of how Kubernetes works at scale, and some of the scaling problems you hit. But also not to be afraid because these problems are solvable and there are many solutions out there. Just a deeper understanding of how they might go about integrating Kubernetes with their infrastructure. It's not always straightforward. Be excited by Kubernetes, have a plan for integrating it and then know these gotchas ahead of time.

We have a bunch of different use cases, every part of our website that you might visit is powered by a Kubernetes service. If you click to make a reservation or the exchange of money is also done by a service. We also have a bunch of very data intensive machine learning jobs, and those are also done in Kubernetes. There are different use cases, payments, cron jobs that are working with banks, services that are making requests and have an API layer, and then these machine learning jobs.

One of the problems we had was we wanted to apply our Kubernetes configuration, and also all these other earlier configurations, AWS configuration. How do you deploy your service alongside all your other configuration. And other configuration such as project ownership, metadata, and documentation. All of these things need to be updated and applied somehow. Another scaling issue we hit was how do you optimize builds for a huge java monorepo, and there’s some interesting multi-stage build optimizations you can make.