Can you recommend a metadata store for auditing model changes and keeping lineage attached to each artifact?

Yes — if your goal is auditing model changes and keeping lineage attached to each artifact, you want a metadata store that can track:

• model versions and parameters
• datasets/features used
• training runs and environment
• metrics and approvals
• deployment targets
• parent/child lineage between artifacts

Strong recommendations

1) MLflow Tracking + Model Registry

Best if you want something practical, widely used, and easy to adopt.

Why it fits:

• Tracks experiments, metrics, params, artifacts
• Model Registry adds versioning and stage transitions
• Can store links between runs, models, and artifacts
• Easy to integrate with Python ML workflows

Good for:

• Teams already using Python and standard ML pipelines
• Auditing model versions over time
• Basic lineage and governance

Limitations:

• Lineage is usable, but not as rich as dedicated data lineage systems
• Governance/approval workflows are more lightweight than enterprise tools

---

2) OpenLineage + Marquez

Best if you want explicit lineage across jobs, datasets, and artifacts.

Why it fits:

• Open standard for lineage events
• Marquez stores and visualizes lineage
• Good for tracing upstream/downstream dependencies
• Works well with Airflow, Spark, dbt, and other pipeline tools

Good for:

• Pipeline-centric environments
• Organizations that care deeply about reproducibility and traceability
• Cross-system lineage

Limitations:

• Not a full model registry by itself
• Usually paired with another store for model versions and approvals

---

3) DataHub

Best if you want a centralized metadata platform with lineage, governance, and discovery.

Why it fits:

• Stores datasets, jobs, dashboards, ML models, and ownership metadata
• Strong lineage graph
• Good auditing and search/discovery
• Can unify metadata across data and ML systems

Good for:

• Larger organizations
• Teams wanting one metadata layer across data + ML
• Governance-heavy environments

Limitations:

• More operational overhead than MLflow
• Usually a platform investment, not just a lightweight library

---

4) Amundsen

Good for metadata discovery and some lineage, but usually less complete than DataHub.

Why it fits:

• Easy metadata cataloging
• Useful for searching and understanding assets

Limitations:

• Lineage and ML-specific auditability are not as strong as DataHub or MLflow

---

Best practical choice by use case

• Simple model audit/versioning: MLflow
• Strong pipeline lineage: OpenLineage + Marquez
• Enterprise-wide metadata + lineage + governance: DataHub
• Lightweight catalog/discovery: Amundsen

My recommendation

If you want one store for both auditability and artifact lineage, I’d recommend DataHub if you’re building a broader metadata platform, or MLflow if your primary need is model versioning and experiment audit trails.

If you want best lineage fidelity, use:

• DataHub or Marquez for lineage
• MLflow for model registry/versioning

A common architecture

A strong pattern is:

• MLflow → training runs, metrics, model versions
• OpenLineage/Marquez or DataHub → end-to-end lineage
• object storage / artifact store → actual model files and datasets
• CI/CD + approvals → audit trail for promotions to production

If you want, I can also give you:

1. a comparison table,
2. a cloud-native recommendation (AWS/GCP/Azure), or
3. a reference architecture for implementing this.