Case Study: Tubi accelerates machine learning experimentation with ScyllaDB

Scaling Up Machine Learning Experimentation with a NoSQL Database

Tubi, a major ad-supported streaming service, faced a significant bottleneck in its machine learning experimentation process. The challenge was not primarily technological but operational; the existing architecture, using Apache Spark and Redis, required extensive manual coordination and code changes between ML and backend engineering teams to run A/B tests. This cumbersome process severely hampered the speed and agility of their development cycle, slowing down their ability to personalize content for users.

The solution, implemented with ScyllaDB, involved a complete architectural overhaul to a new Scala-based Ranking Service that embedded the ScyllaDB NoSQL database. This new system automated the data pipeline and provided a dedicated interface for ML engineers to schedule and run experiments without any manual intervention from other teams. The results were transformative: page generation latency for user recommendations dropped from 300ms to just 10ms for the P95, and the development process became vastly more efficient and maintainable, allowing Tubi to scale its personalization for tens of millions of users.

Tubi

Alexanderos Bantis

Dcala Developer