Removed: the orchestration layer — Step Functions driving the workflow, a Lambda function per defect detector, and the S3 bucket passing video frames between stages.
Outcome: infrastructure cost down by more than 90%, and better scaling than the distributed version.
Amazon ran the same collapse deliberately. Prime Video’s video quality analysis service was assembled from serverless parts: Step Functions orchestrating the workflow, Lambda functions running each defect detector, an S3 bucket passing video frames between stages. It “hit a hard scaling limit at around 5% of the expected load”, and paying per state transition and per cross-component byte made it unaffordable regardless. The team “moved all components into a single process to keep the data transfer within the process memory.” Their own conclusion: “Moving our service to a monolith reduced our infrastructure cost by over 90%. It also increased our scaling capabilities.”
The orchestration layer was the cost and the attack surface at the same time, so it went. (The original post is no longer served from its old address; the archived copy is linked above.)
Source: Prime Video Tech, Scaling up the Prime Video audio/video monitoring service and reducing costs by 90% (archived) (archived 30 December 2023)