Why Amazon Prime Video Switched to a Monolith and Cut Cloud Costs

Recently, the Prime Video team published an Amazon case study that sparked lively discussion in the developer community. The team candidly evaluated how migrating from a micro‑service architecture to a monolith could save money and avoid expensive cloud services such as AWS Step Functions and Lambda serverless functions.

The original monitoring solution required a highly scalable system to track quality issues for every stream viewed by customers, handling thousands of concurrent streams. It used distributed components orchestrated by AWS Step Functions, a serverless state‑machine service.

Step Functions turned out to be a bottleneck: frequent state transitions for each second of video processing quickly hit account limits, and each transition incurred charges. Additionally, massive first‑layer calls to S3 for temporary video‑frame storage added cost.

Realizing that the distributed approach offered little benefit for their specific use case, the team repackaged all components into a single process, eliminating the need for S3 and implementing a lightweight orchestration layer within a single instance. The new solution runs on EC2 and ECS, providing a lightweight layer to dispatch customer requests.

The study concluded that while micro‑services and serverless components can handle large loads, their use in a monolithic architecture should be evaluated case‑by‑case. Consolidating services into a monolith reduced infrastructure costs by more than 90% and noticeably improved scalability. The paper also noted cost savings from EC2 Reserved Instances, showing that even internal AWS customers can benefit from such models.

Comments on Hacker News expressed surprise, noting that AWS often promotes micro‑services and serverless as the default modern deployment model. Critics, including Ruby on Rails creator David Heinemeier Hansson and consultant Sam Newman, argued that the recommended solutions are not always cost‑effective and that micro‑services are frequently over‑hyped, adding unnecessary complexity.

Hansson warned that replacing method calls and module separation with network calls and service partitioning is often “crazy,” while Newman emphasized that micro‑services should not be the default architectural choice and should be considered only after thorough value‑chain analysis.

Overall, the case study provides a rare, candid look at how simplifying architecture can dramatically lower costs and improve performance, offering valuable insights for architects beyond just AWS customers.