Prime Video made one tool a monolith and cut its cost 90%

transcript

A Prime Video team turned serverless pieces into a monolith. Its cost fell by over 90 percent. Not all of Prime Video. Just one tool.

Prime Video has thousands of live streams. One tool watches them for frozen video. Blocky corruption. Sound out of sync with the picture.

Version one was serverless. Quick to build. Step Functions runs the steps. It charges per state transition. Several transitions for every second of every stream. And every frame goes up to S3. Then down again, for each detector.

The result? A hard scaling limit. At about 5 percent of expected load. And it cost too much. A huge number of paid storage requests.

Before: pay per transition. Frames parked in S3. The fix: all three parts in one process. Orchestration is just code. Frames pass in memory. No bucket at all. Outgrow a machine? Run more copies, each with different detectors.

Infrastructure cost: down by over 90 percent. And now it monitors every stream customers watch. Not just the top ones.

Every hop between services costs time. In the cloud, money too. But it's not a rule. The post says: case by case.

Full breakdown in the lecture.