Breaking the monolith: Improving velocity by migrating to ML platform - Vaidehi Thete

The New York Times reimagined its targeted email program by transitioning a monolithic, multi-hour machine learning job into a series of decoupled, composable services. Platform Engineer Vaidehi Thete details how the original Airflow-orchestrated process suffered from resource constraints, stale user data, and operational overhead that hindered newsroom flexibility. By leveraging a centralized machine learning platform featuring Go-based microservices, RabbitMQ for data persistence, and Trident for model serving, the team reduced processing time for 4 million users from hours to minutes and successfully scaled the system to 40 million users. Key technical integrations include Bigtable for real-time user history and Redis for article features, which improved prediction quality. This architectural shift eliminated 14 hours of manual curation work for editors, increased weekly active users, and established a reusable framework that allows practitioners to focus on model development rather than infrastructure management.