From analysis to product
A data product has consumers, an interface and an SLA — a mart, an API, a scored table or an embedded model. Define ownership, refresh cadence, freshness and quality guarantees, and a deprecation path, exactly as you would for a service.
Consistency layers
A semantic layer or metric store defines each metric once so 'active user' means the same thing everywhere. A feature store shares feature definitions between training and serving, which is the standard fix for training–serving skew.
Governance and readiness
Classify sensitive data, apply row- and column-level access controls, mask personal data in non-production environments, retain only what policy allows, and keep lineage so an incident's blast radius is knowable.
- Tests and freshness monitors on every gold table.
- Documented ownership and on-call for pipelines.
- Backfill procedure and rollback documented before launch.