Democratizing Machine Learning at Netflix: Building the Model Lifecycle Graph

Netflix's machine learning infrastructure spans multiple business domains including personalization, studio workflows, payments, and advertising, but fragmented ML tools and silos prevent effective cross-domain collaboration and asset discovery.

• •Netflix applies machine learning across multiple business domains (personalization, studio, payments, ads) with diverse tech stacks and organizational structures
• •The core problem: ML tools exist in isolated silos with separate backends and interfaces, making it difficult for practitioners to discover model dependencies and reuse assets across teams
• •Content embeddings created by Studio teams for post-production could benefit Ads and Personalization teams but cross-pollination is extremely difficult due to fragmented infrastructure
• •Netflix built the Metadata Service (MDS) and Model Lifecycle Graph to connect heterogeneous metadata from dozens of sources (pipeline orchestration, model registry, feature store, experimentation platform, data