The model registry

Fully Automated MLOps

Arturo Opsetmoen Amador

Senior Consultant - Machine Learning

The ML model lifecycle

An image showing the ML model lifecycle. 1: model training, 2: model evaluation and validation, 3: model deployment and monitoring, 4: model decommission.

Throwing a model over the fence

An image illustrating the expression throw the model over the fence. After a training pipeline is done, the model is sent via email to the Ops team, creating a clear silo.

A first step towards automated MLOps

Workflow figure. After training a model, this is delivered to the Ops team by registering the model into the model registry.

What is the model registry?

A figure showing the model registry: a component where the stage, prod, and archive stages in the models lifecycle are managed.

What is the model registry? - Experimentation

A figure that shows orchestrated experiments in the development environment delivering trained ML models to the model registry via the experiment tracking system. Metadata is stored in the metadata store.

What is the model registry? - Registering a model

Continuation of the previous figure. Here, a model is sent from Development to Staging via the model registry. CI/CD is implemented across the workflow.

What is the model registry? - Updated deployment

The figure is continued showing a model that has made it through staging and prod and is now used by downstream services.

What is the model registry? - Model decommission

Continuation and finalization of the figure, the old model previously deployed to the downstream services is archived in the model registry after a new, improved model has been deployed to the downstream services.

Let's practice!

Fully Automated MLOps