What happens at the stroke of midnight on December 31st? Other than the fireworks 🙂 In this talk, we’ll dive into the Lamba service that powers Videos@ Meta and we’ll see whats really happening behind the scenes on New Years Eve. We’ll explore the system design, and we’ll explore our test principles, for an end-to-end look at the biggest day of the year.

This talk covers the growth and evolution of Meta’s video backend infrastructure that supports VOD processing and delivery. Over the past few years, we rearchitected our backend to transition from app-specific vertical stacks to horizontal layers shared across Meta’s apps such as Instagram, Facebook, Messenger and Oculus. This defragmentation of our stack has significantly improved our resource usage efficiency and engineering velocity.

Working with video data in a large scale, distributed environment presents many challenges in storage, efficiency, and usability. This talk unveils a general purpose tool to build complex video processing applications: virtual video files. Virtual video files save storage space, and simplify complex operations like Distributed Encoding, Live to VOD conversion, and more. Furthermore, we discuss how Meta integrates it's own network file stores and abstractions with ffmpeg - strategies to optimize for ease of use and efficient data access.

This talk covers how we measure visual quality at Meta and then how we use these metrics to make smarter trade-offs in our player when optimizing for the users' quality of experience. Quality metrics are used to decide the best ceiling to use when playing videos on cellular networks. Quality metrics are also leveraged when making tradeoffs between visual quality and risk of in play stalls in our adaptive streaming algorithms.

Understanding the media capabilities of the end-user devices can help service providers gain valuable insights on playback experience. These insights can be used to craft the content delivery workflows to optimize the right encoding recipes available for these devices. Especially in the case of user generated content which can have a multitude of ingress formats and several hundreds of end user device types, this presents a large scale data association challenge - choosing the right codec formats, resolutions, frame rates and bitrates that can deliver the best viewing experience. This presentation covers one such at-scale problem of finding the right operating points or video recipes to be delivered to the end users enabling rich experience.