Democratizing Machine Learning at Netflix: Building a Model Lifecycle Graph

As Netflix's ML use expanded into personalization, studio, payments, and ads, it faced the problem of models and data becoming isolated within each domain. There were assets with high reuse value, like Studio's content embeddings, but the model registry, pipeline orchestrator, experimentation platform, feature store, AI Dataset, and identity platform were all separate, making cross-domain reuse and collaboration difficult.

Practitioners need answers to three questions. Discovery asks what features and data sources exist, Lineage asks which pipeline produced which model, and Impact asks which A/B tests and downstream models are affected. With practitioners having to navigate separate UIs for each system as before, these connections couldn't be tracked all at once.

The solution is the Metadata Service (MDS) and the Model Lifecycle Graph. Every ML asset is identified with an AIP URI in the form aip://<componentType>/<platformId>/<resourceId>, components are normalized into entities, and Domain is separated from Provider. Here, Component is a uniquely addressable object, Entity is an ML asset with attributes like name, description, and owner, Domain is an abstract interface grouping assets that share the same data shape, such as models and pipelines, and Provider is the implementation of that interface tailored to an actual source system.

Data flows through four stages.

• Event ingestion: Thin events are received via Kafka and AWS SNS/SQS to signal changes.
• Entity enrichment: When an event arrives, the source system API is called to hydrate the latest state.
• Normalization: Different schemas are converted into standard entities.
• Graph materialization: Relationships are connected to form a navigable graph.

This design allows the latest state to be recovered on the next hydration even if event ordering gets scrambled or some events are lost. However, since this increases read load on source systems, rate limiting, caching, and backoff become important. Ultimately, MDS ties models, features, pipelines, experiments, and datasets into a single connected graph, enabling reuse and collaboration across domains.