March 27, 2017

Slaying Monoliths at Netflix with Node.js


Yunong Xiao
Yunong Xiao, Principal Software Engineer at Netflix, describes scaling challenges the company has encountered and explains how the company delivers content to a global audience on an ever-growing number of platforms.

The growing number of Netflix subscribers -- nearing 85 million at the time of this Node.js Interactive talk -- has generated a number of scaling challenges for the company. In his talk, Yunong Xiao, Principal Software Engineer at Netflix, describes these challenges and explains how the company went from delivering content to a global audience on an ever-growing number of platforms, to supporting all modern browsers, gaming consoles, smart TVs, and beyond. He also looks at how this led to radically modifying their delivery framework to make it more flexible and resilient.

One of the first steps Netflix took to cope with their swelling subscriber base was to migrate all their infrastructure to the cloud. But somehow, Xiao says, that didn't mean that once the migration complete, the developers could "just sit around and watch TV shows." The cloud, after all, is just somebody else's computer, and scaling for the number of users is just part of the problem. As the number of users increased, so did the number of platforms they had to deliver to. In its first iteration, Netflix only worked on the browsers, and the framework was simply a Java web server that managed everything. The server did more or less everything, both rendering the UI and accessing the data.

Netflix relies on microservices to provide a diverse range of features. For each microservice there is a team of developers that more or less owns the service and provides a client to the Java server to use. The Java server -- the monolith in this story -- suffered from several issues. To begin with, it was very slow to push and innovate. Every time a new show launched and they wanted to add a new roll title to the UI, they had to push the service. If one of the development teams launched a new and improved version of a client, they had to push the service. When a new microservice was added to the existing ones, they had to push the service. Furthermore, increasing the number of supported devices was nearly impossible in any practical sense.

So in the next iteration, the development team migrated to a REST API. This unlocked the ability to support more devices. The new framework also separated the rendering of the UI and the accessing of data processes. However, the REST API also came with its fair share of disadvantages. For one it was inflexible, as it was originally designed for one kind of device and adding new devices was painful. Also, as a different team owned the REST API, the microservices teams were often waiting weeks for API changes to support their own new services.

It also proved inefficient. REST is resource based and every little element on the Netflix UI is a resource. So, in order to, for example, fetch all of a customer's favorite movies, the services had to make multiple round trips to the back end. Ultimately, it proved difficult to maintain, because the API became more complex and bloated as developers tried to retrofit it with more features.

The different developer teams needed flexibility to innovate for the platforms they were supporting, and the resulting REST API was too clunky and restrictive for this. Another evolution of the Netflix framework was required.

The API.NEXT allowed each team to upload their own custom APIs to the servers. The teams could change the scripts (written in Groovy) as much as they liked without affecting other teams. The API service itself could also be updated independently from the APIs that it was serving. The problem was the Monolith was back again, and that led to scaling problems. Netflix has literally thousands of scripts sharing the same space, serving millions of clients. It was common, says, Xiao, to "run out of headspace," be that memory, CPU, or I/O bandwidth. This led to expensive upgrades when more resources were needed. Another thing that even led to outages were errors in the scripts themselves: If a script had a memory leak, for example, it could bring down the system for everyone.

Another problem was what Xiao calls "Developer Ergonomics." The NEXT.API server was a very complex piece of software with multiple moving parts. Scripts could not be tested locally. To test a script, the team had to upload it to a test site, run it, test it, and, if there were any problems, go through the whole process again after troubleshooting the issues. This process was slow and inconvenient and led to the current iteration of the Netflix framework, one in which scalability and availability, and developer productivity are taken into account.

While designing the new framework, it was established that, on the scalability/availability front, one of the goals was to achieve process isolation to avoid the problems the NEXT.API suffered from. It also required that the data access scripts and API servers were kept separate to reduce infrastructure costs. The designers also wanted to reduce the startup time and have immutable deployment artifacts, which would allow to reproduce the different builds.

As for developer productivity, most developers wanted to use the same language (preferably JavaScript) on the server and the client, and not deal with two distinct technologies. They also needed to be able to run tests locally, have faster incremental builds, and an environment that as closely mirrored the production as possible.

The new framework, called New Generation Data Access API, has moved all the data accessing APIs into separate apps running Node.js. Each app is now isolated running in a Docker container. The back-end services are now contained within a Java-based server the Netflix development team calls the Remote Service Layer. The RSL integrates all back-end services under one consistent API. Whenever developers want to deploy a new APIs, they push JavaScript to the server in the form of a Node.js container.

Overall, Netflix's current combined Java/Node-based platform allows for a quicker and easier deployment, with fewer of the issues that plagued prior monolithic approaches.

Watch the complete presentation below:

If you're interested in speaking at or attending Node.js Interactive North America 2017 - happening October 4-6 in Vancouver, Canada - please subscribe to the Node.js community newsletter to keep abreast of dates and deadlines.

Click Here!