MLflow Model Registry
Introduction
The MLflow Model Registry is a centralized store for managing model versions and their
lifecycle stages (Staging, Production, Archived). It's where you go from "I trained a
model" to "this model is approved for production."
Why This Matters
Training produces models. But which model should be serving production traffic?
Who approved it? What version is currently live? Can we roll back if needed?
The model registry answers these questions:
- Models are registered with versions
- Each version can be in a stage: None, Staging, Production, Archived
- Transitions are tracked (who moved what to production and when)
- Deployment systems pull from the registry
In a mature MLOps setup:
- Data scientists register promising models
- CI/CD runs validation tests
- Approved models are promoted to Production
- Serving infrastructure pulls the Production model automatically
Tasks
Hint: mlflow.register_model('runs:/<run_id>/model', 'MyModel')
Hint: mlflow.pyfunc.load_model('models:/MyModel/Production')