What is a model registry and how does it enable CD?

WHAT THIS TESTS Whether you know how trained models are tracked, governed, and promoted to production in a reproducible, auditable way, rather than being copied around as loose files.

A GOOD ANSWER COVERS A model registry is a centralized catalog that versions every trained model and stores the artifact alongside rich metadata: training metrics, hyperparameters, dataset and code lineage, the input and output signature, and the environment needed to run it. Each registered model has multiple versions, and versions are assigned lifecycle stages such as none, staging, production, and archived. This gives you reproducibility, auditability, approval workflows, and one-step rollback by repointing the production stage to an earlier version. Integration with continuous deployment is event driven: when a version is promoted, for example from staging to production, the registry emits an event or fires a webhook. The CD system consumes that signal, pulls the artifact and its environment spec, runs validation and integration tests, builds a container, and deploys or updates the inference service. The registry becomes the single source of truth that decouples training pipelines from serving infrastructure.

COMMON WRONG ANSWERS Calling it just a place to dump model files in S3 misses versioning, lineage, stages, and approvals. Another weak answer is coupling deployment directly to the training job rather than to a registry stage transition, which breaks rollback and governance.

LIKELY FOLLOW-UPS How do you implement an approval gate before production promotion? How does the registry support rollback? How do you guarantee the serving environment matches the training environment recorded in the registry?

ONE CONCRETE EXAMPLE A team trains a recommender, logs version 7 to the registry with its metrics and a conda spec, and after offline evaluation a reviewer promotes it to production. A webhook triggers the CD pipeline, which builds an image from the recorded spec, runs a smoke test, and rolls it out. When latency regresses, they repoint production to version 6 in seconds.