Describe a Model Registry and how it differs from versioned storage

WHAT THIS TESTS: This question probes whether you understand the operational boundary between raw artifact storage and model lifecycle governance. A senior candidate should show that a Model Registry is a metadata control plane that enables reproducibility, auditability, and safe promotion, while versioned object storage is just a persistence layer. The interviewer wants to hear that you have shipped models through environments and felt the pain of using buckets alone.

A GOOD ANSWER COVERS: First, define the registry as a system of record for model metadata, not just files. Second, explain that it tracks lineage by linking a model version to the exact training run, code commit, dataset hash, and hyperparameters. Third, describe stage management, meaning a model moves through logical states like staging, production, or archived with controlled transitions and access rules. Fourth, mention artifact metadata such as signatures, schemas, performance metrics, and approval annotations. Fifth, contrast this with versioned S3 or GCS, which only preserve file versions and offer no semantic understanding of which version is serving production traffic or which training job produced it.

COMMON WRONG ANSWERS: A red flag is saying a registry is just a database pointer to S3 paths with no explanation of lifecycle semantics. Another mistake is claiming that S3 object versioning plus bucket tags equals a registry; tags lack atomic stage transitions and queryable lineage graphs. Candidates also err by focusing only on file deduplication or backup rather than governance, or by ignoring the human-in-the-loop approval steps that registries enforce before a model reaches production.

LIKELY FOLLOW-UPS: The interviewer may ask how you would build a minimal registry if you only had a database and object storage, which tests whether you can separate the metadata layer from the blob layer. They might ask how to handle model rollback during an incident, or how to synchronize registry state with a serving platform so that production traffic actually points to the registry-approved artifact. Another follow-up is how to manage permissions so that data scientists can register models but only platform engineers can promote them to production.

ONE CONCRETE EXAMPLE: Imagine a computer vision team training a new object detector. In a versioned GCS bucket, you might have model_7.pt and model_8.pt with generation IDs, but nothing tells you that model_8 was trained on a corrupted augmentation pipeline. A Model Registry would show that model_8 is linked to run_id abc123, dataset_v3, and a mAP of 0.92, while also marking model_7 as the current production candidate with an approval timestamp and a deployment artifact checksum. When an on-call engineer needs to roll back, the registry provides the exact previous production version and its full context in seconds rather than hours of digging through bucket logs.