Microservices

Amazon Prime Video cut the infrastructure cost of one of its monitoring tools by more than 90% by taking a distributed design apart and running its components in one process, which then scaled further than the distributed version ever had. The tool spotted audio and video defects in streams. As first built, AWS Step Functions orchestrated the work, one component split each video into frames and uploaded them to Amazon S3, and defect detectors running as AWS Lambda functions downloaded the frames and analysed them. The design was meant to scale out, and it supported about 5% of the expected load. The orchestration performed several state transitions for every second of every stream, so it quickly reached account limits, and Step Functions charges per state transition. Passing every frame through S3 generated a high number of Tier-1 calls, S3 requests billed one by one, and those were expensive too.