Model Registry
Definition
Central repository for storing, versioning, and managing trained machine learning models with their metadata, ensuring easy access and organization.
Use Cases
- Netflix: Managing frequent updates to recommendation and personalization models while ensuring only approved versions reach production — Uses an internal ML platform with a centralized model catalog/registry concept to track model versions, metadata, and deployment readiness as part of their MLOps workflows (Improved operational reliability for model rollouts, clearer lineage and reproducibility, and faster iteration with controlled promotion of models)
- Uber: Coordinating many ML models (pricing, matching, fraud, ETA) across teams with consistent versioning and deployment governance — Uses an internal ML platform (Michelangelo) that includes a model repository/registry capability to manage model artifacts, metadata, and productionization workflows (More standardized model deployment processes, reduced friction between training and serving, and better traceability of what model version is running in production)
- Airbnb: Tracking and deploying multiple iterations of ranking and search models with auditability and rollback capability — Uses an internal ML platform that supports model management (registry-like cataloging of model artifacts and metadata) integrated with training pipelines and deployment (Safer deployments with easier rollback, improved reproducibility, and better collaboration across data science and engineering)
Provider Equivalents
- AWS: Amazon SageMaker Model Registry
- Azure: Azure Machine Learning model registry
- GCP: Vertex AI Model Registry
- OCI: OCI Data Science Model Catalog
Frequently Asked Questions
- What's the difference between a Model Registry and a Feature Store?
- A model registry manages trained models: versions, metadata, approvals, and which model is deployed. A feature store manages input data features used to train and serve models, ensuring the same feature definitions are used consistently in training and production.
- When should I use a Model Registry?
- Use a model registry when you have more than one model version to manage, need controlled promotion to staging/production, want reproducibility (knowing exactly what data/code produced a model), or require governance such as approvals, audit trails, and rollback.
- How much does a Model Registry cost?
- Costs are usually driven by the underlying services: storage for model artifacts and metadata, API calls/operations, and any integrated CI/CD or pipeline runs. Some platforms include registry features as part of their ML service at no separate line item, but you still pay for associated storage, training, and deployment resources.
Category: ai-ml
Difficulty: advanced
Related Terms
See Also