Question

A team of data scientists and ML engineers collaborates on a customer churn model. Over several months they retrain the model as new labeled data arrives. Compliance now requires that for any model version currently serving predictions, the team must be able to identify the exact training dataset snapshot that produced it and reproduce that dataset. The team already uses Vertex AI Pipelines to orchestrate training. What is the best approach to meet this requirement?

Accepted Answer

Use Vertex AI Datasets (managed datasets) to create a versioned dataset for each training run and rely on Vertex ML Metadata lineage tracking, which links the dataset artifact to the resulting model in Vertex AI Model Registry.

Answer

Store all training data in a single BigQuery table and add a comment on each model in Model Registry noting the retraining date so the data can be located manually.

Answer

Export each model to a Cloud Storage bucket alongside a CSV copy of the training data, using timestamped folder names to correlate models and data.

Answer

Enable Vertex AI Feature Store online serving so that features are always current, eliminating the need to track dataset versions for reproducibility.