Enriching the Customer Experience Using AWS Lambda with Prime Video

With more than 200 million Prime members worldwide, Prime Video offers a vast catalog of movies and television shows tailored to individual tastes. Categories like “movies we think you will like” or “new release movies” are more than just friendly suggestions: they are outcomes of a machine learning (ML) model designed using Machine Learning on AWS—a comprehensive set of artificial intelligence and ML services—that assesses a customer’s likelihood of engagement based on past streaming behavior.

“With our offline personalization service, the home page often wouldn’t reflect a customer’s latest viewing history,” says Hani Suleiman, senior principal engineer at Prime Video. “We wanted to create a solution where the home page understands what you just watched and provides other recommendations that are suited to your tastes.”

Running offline jobs so frequently was inefficient and created unnecessary waste for Prime Video. At the same time, the offline workflows didn’t deliver recommendations as close to real time as the team would have liked. It was clear that a more dynamic runtime solution was needed. So, Prime Video embarked on a mission to keep pace with customers’ actions and preferences.

To bring its recommendations as close to real time as possible, Prime Video devised an event-driven architecture based on AWS Lambda; this workflow scales based on the number of customer playback events rather than the number of requests to the recommendation service. With an event-driven architecture, the team could deliver recommendations in near real time, improving engagement.

An AWS Lambda function initiates the personalization workflow whenever a customer performs a key action on Prime Video, such as streaming a new movie or television show. To ingest these events, Prime Video uses an existing event stream on Amazon Kinesis Video Streams, which captures, processes, and stores media streams for playback, analytics, and ML.

Then, ML inferences are processed in near real time while the customer is watching the content. When customers return to the Prime Video home page, they can explore content recommendations based on what they just watched.

With serverless architecture, Prime Video can scale up by orders of magnitude throughout the day, meeting demand during peak hours in a given territory. “The amount of traffic that we get at 3:00 a.m. is very different from the traffic at 8:00 p.m.,” says Suleiman. “Using AWS Lambda, we do not have to worry about accommodating these fluctuations. It automatically scales based on our traffic needs.”

On AWS, Prime Video has dramatically reduced development work and simplified scaling and operations. “We no longer have to manage security campaigns, perform operating system upgrades, or replace hosts,” says Suleiman. “AWS Lambda takes care of all that for us. Even though the team now owns new code and a new service, we haven’t increased our operational burden.” The event-driven workflow is also 25 percent cheaper in compute costs compared with alternative architecture designs evaluated by the team.