Netflix on AWS stack

Netflix’s cloud stack is back in focus after engineer Shalini Goyal sketched the core pieces publicly: Amazon load balancers at the edge, Zuul as the gateway, Cassandra for data, and Kafka for event streams. (x.com) A load balancer is traffic control for servers: Netflix has long used Amazon Web Services regions and Elastic Load Balancing to spread requests and avoid single choke points. Netflix said in 2016 that it had completed its seven-year cloud migration and could shift capacity across multiple AWS regions; AWS now says Netflix operates across four AWS Regions. (about.netflix.com) (aws.amazon.com) A gateway is the front door for app requests: Netflix’s Zuul service sits between devices and backend services, deciding where requests go and applying checks before traffic moves deeper into the system. Netflix introduced Zuul in June 2013 and said its API was then handling more than 50,000 requests per second at peak; the current Zuul project page still describes it as the front door for requests from devices and websites. (netflixtechblog.com) (github.com) A distributed database is a data store spread across many machines instead of one big server: Netflix says Apache Cassandra remains central to that layer. In a September 2024 engineering post, Netflix called Cassandra “the backbone” for use cases including user sign-ups, viewing histories, real-time analytics, and live streaming. (netflixtechblog.com) An event stream is a running feed of updates that lets services react without waiting on one another: Kafka fills that role in several Netflix systems. Netflix engineers wrote in 2018 that Keystone, one of the company’s real-time data platforms, included a Kafka-enabled messaging service for producing, collecting, processing, and moving microservice events in near real time. (netflixtechblog.com) Those pieces fit the shape of Netflix’s larger architecture on AWS: stateless services at the edge, distributed data stores underneath, and messaging systems linking microservices without turning every request into a direct call. Netflix said the move away from vertically scaled single points of failure followed a 2008 database corruption incident, and said cloud elasticity let it add thousands of virtual servers and petabytes of storage within minutes. (about.netflix.com) The video itself does not ride through that same request path end to end. Netflix said in its cloud-migration post that video is delivered through Open Connect, its own global content delivery network, while AWS handles business logic, distributed databases, analytics, recommendations, transcoding, and other application functions. (about.netflix.com) Netflix’s public architecture notes also show why engineers keep studying this stack more than a decade later. The company tied Zuul to dynamic routing and multiregion resiliency, described Cassandra as a high-availability store for globally distributed workloads, and said AWS infrastructure gives it room to scale for more than 280 million members in over 190 countries. (github.com) (netflixtechblog.com) (aws.amazon.com) Goyal’s outline landed because it turned an abstract “Netflix on AWS” story into four concrete building blocks engineers can name, compare, and copy: traffic in through load balancers, requests through Zuul, state in Cassandra, and events through Kafka. (x.com)