Question

A data science team runs a Vertex AI Pipeline that preprocesses data in BigQuery, trains a model, and evaluates it. Three months after a run, an auditor asks the team to prove exactly which dataset snapshot, container image, and hyperparameters produced a specific deployed model, and to reproduce that run. The team logged experiment metrics in Vertex AI Experiments but is now unsure how to trace the complete chain of artifacts and executions. What is the most effective way to satisfy the auditor's request going forward?

Accepted Answer

Query Vertex ML Metadata to trace the lineage graph of artifacts and executions produced by the pipeline, linking datasets, parameters, container images, and the resulting model.

Answer

Rely on Vertex AI Experiments metrics alone, since experiment runs automatically capture full artifact lineage and reproducibility metadata.

Answer

Enable Cloud Audit Logs on the Vertex AI API and reconstruct the model's inputs from the API request logs.

Answer

Store a copy of the training data and hyperparameters in a spreadsheet at the end of each pipeline run for manual auditing.