In 2023 the Prime Video technology blog published a post with a headline that travelled fast: the team had moved a video-quality monitoring service from a serverless, distributed design to a single process, and cut its infrastructure cost by over 90%. Within days it was being cited as proof that serverless and microservices had failed, then as proof that they hadn't.
The original post has since been taken down. Its address on primevideotech.com now redirects to a general Amazon page, and Amazon's CTO notes on his own post that the site is no longer available. An archived copy survives, and this note is built from it, plus the public responses.
What they had
Prime Video monitors every live stream its customers watch for perceptual quality problems such as block corruption or audio/video sync issues, then triggers a fix. The team already owned a quality-inspection tool, but by its own account it was never designed to run at high scale. The target was thousands of concurrent streams, growing over time.
The service has three parts: a media converter that turns streams into frames and audio buffers, defect detectors that analyse them in real time, and orchestration that controls the flow. The first version was built, in the post's words, "using serverless components (for example, AWS Step Functions or AWS Lambda)", which it calls "a good choice for building the service quickly":
- Step Functions orchestrated the flow.
- A microservice split videos into frames and uploaded the images to an S3 bucket as temporary storage.
- Each detector ran as a separate microservice that downloaded the images and processed them concurrently using Lambda.