Back to radar
Production AI Radar
Dataset versioning (DVC)
Git-linked dataset and artifact versions so every model can name its training data.
TrialMLOpsNew
- Why this ring
- Without data versions, registry provenance is fiction. Trial when more than one person trains models.
- Production risk if ignored
- Cannot reproduce or explain a production model when auditors ask which data trained it.
- EU AI Act relevance
- Supports data governance and technical documentation for high-risk systems.
- Typical effort
- weeks
- Low FinOps impact
Use cases
- Reproducible training
- Audit data lineage
- Multi-person ML teams
Adoption steps
- Remote storage for data
- Link DVC revs to MLflow runs
- Block promote without data hash
- Document retention
Related tools
In your assessment
Data lineage maturity + storage remote review