How would GDPR requirements influence experiment tracking and model management design?

WHAT THIS TESTS: This question tests whether you treat privacy regulations as first-class systems constraints rather than compliance paperwork. The interviewer cares if you understand that GDPR's right to erasure and explainability requirements ripple through the entire ML lifecycle, from data versioning to model provenance and deployment. They want to see that you can design for auditability, surgical data removal, and reproducibility without raw PII.

A GOOD ANSWER COVERS: A strong answer hits four design areas in order. First, immutable lineage: every experiment must log the exact dataset version, feature pipeline hash, and code commit so you can reconstruct what data influenced which model. Second, data isolation and mapping: maintain row-level provenance so a user deletion request can pinpoint exactly which training runs included that individual. Third, retraining mechanics: design pipelines that can regenerate a training set with specific records excluded and retrain only affected model versions, rather than full retraining from scratch. Fourth, explainability and access control: version SHAP or LIME artifacts alongside the model in the registry, and enforce RBAC plus immutable audit trails for every promotion or rollback.

COMMON WRONG ANSWERS: Red flags include suggesting manual spreadsheet tracking, saying you will just delete the user from the database without addressing model unlearning, claiming explainability is only a serving-layer UI concern, or proposing to store raw training data inside the model registry. Another weak pattern is assuming immutable logs conflict with GDPR deletion rights; a senior candidate should explain how metadata can be retained for audit while personal data is purged.

LIKELY FOLLOW-UPS: Expect the interviewer to ask how you would handle a deletion request if the model is already in production and retraining is expensive. They may also probe the difference between model explainability and data provenance, or ask how you would prove to an auditor that a specific user's data no longer influences a given model version.

ONE CONCRETE EXAMPLE: A senior candidate might describe a system using Delta Lake or DVC for dataset versioning with row-level hashes. When a deletion request arrives, a lineage service queries the experiment graph to find all model versions trained on shards containing that user. It triggers a retraining job using a new dataset version with those rows excluded, promotes the replacement through a canary deployment, and archives the old model artifacts while retaining metadata for regulatory audit. The experiment tracker stores the new lineage link and a tombstone record for the old model.

Source: NIST AI Risk Management Framework