End of Microservices Era? — Amazon Prime Case Study
Prime Video offers thousands of live streams to customers. To guarantee a seamless viewing experience, Video Quality Analysis (VQA) at Prime uses a monitoring tool that identifies quality issues like block corruption or audio/video sync problems. Once detected, they swiftly resolve these issues to ensure high-quality content delivery.
What Went Wrong?
Originally this solution is not designed for high-scale operation, the old architecture was based on Amazon Lambda, which was good for building services quickly. they encountered scaling bottlenecks and high expenses while onboarding more streams.
The two major challenges in their existing serverless approach are:
- Orchestration management using AWS Step Functions — Multiple state transitions per second caused account limits to be quickly reached, and AWS Step Functions charged users per state transition.
- The intermediate data transformation — The defect detectors, which operate as individual microservices, download and analyze images simultaneously using AWS Lambda. However, the significant cost arose from the large number of expensive requests made to the S3 bucket’s primary storage.
To overcome the challenges mentioned above, the company made a bold decision to re-architect its infrastructure. The following statement quotes Amazon’s CTO on this controversial move.
Building evolvable software systems is a strategy, not a religion. And revisiting your architectures with an open mind is a must.
- Werner Vogels, CTO, Amazon.com
How Monolithic Solution Saved 90% of Cost
After realizing that a distributed approach was not providing significant benefits in their specific use case, Prime Video consolidated components into a single process, eliminating the need for an S3 bucket as intermediate storage. By transferring data in memory, they streamlined operations. Additionally, they implemented orchestration to manage components within a single instance.
The below diagrams depict the before & after architectures;
