PRODUCTION SCENARIO
A retailer's data science team has trained 14 versions of a demand forecasting model over two years. Nobody can say which version serves production traffic or how its evaluation compared with the one before it.
Which lifecycle stage is missing, and which capability covers it?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Model management, using Model Registry to track model versions
Once training is finished the work moves to model management, and Model Registry is the central repository for it: it tracks model versions, records how each one evaluated, and deploys models for serving. That is the exact gap here. Versioning the training data through managed datasets tells the team what went in, not which model came out and went live. Better training tooling produces a fifteenth version without answering the question about the first fourteen. Landing source data in Cloud Storage and BigQuery is the ingestion stage, upstream of the problem entirely - and a bucket of model files is a folder, not a registry: no lineage, no evaluation history, no deployment record.