DVC Combined with SageMaker MLflow for End-to-End Model Lineage Tracking

Introduction: Why Model Lineage Tracking Matters

In enterprise machine learning practices, model lineage tracking has always been a core requirement for ensuring model trustworthiness and auditability. As AI regulations tighten and model governance demands increase, teams need to clearly answer a critical question: What data, what code, and what parameters were used to train this model? Recently, the official AWS blog published a technical practice article detailing how to combine DVC (Data Version Control), Amazon SageMaker AI, and Amazon SageMaker AI MLflow Apps to build an end-to-end ML model lineage tracking system, drawing widespread attention across the industry.

Core Solution: Three Tools Working Together to Build a Lineage Chain

The core idea of this solution is to leverage the respective strengths of three tools to form a complete lineage tracking loop.

DVC (Data Version Control) handles data version management. As an open-source data version control tool, DVC manages dataset version changes the same way Git manages code, ensuring every data change is documented. It stores large file metadata in Git repositories while hosting the actual data in object storage services like Amazon S3, achieving lightweight data version tracking.

Amazon SageMaker AI provides the infrastructure for model training and deployment. As AWS's flagship machine learning platform, it supports full-pipeline management from data preparation and model training to deployment and inference, with built-in experiment tracking and model registry capabilities.

Amazon SageMaker AI MLflow Apps plays the role of experiment management and lineage visualization. As a widely adopted ML lifecycle management tool in the industry, MLflow's managed version on SageMaker allows teams to enjoy experiment tracking, model registration, and lineage recording capabilities without having to build and maintain their own MLflow server.

The three tools work together as follows: DVC records where the data comes from, SageMaker records how the model is trained, and MLflow Apps connects this information into a complete lineage graph, enabling full-chain traceability from raw data to the final model.

Deep Dive into Two Deployable Patterns

The article proposes two practical lineage tracking patterns, each suited to different business scenarios.

Pattern One: Dataset-Level Lineage Tracking

Dataset-level lineage tracking focuses on which version of a dataset was used to train which version of a model. In this pattern, DVC generates a unique hash identifier for each dataset version, and this identifier is recorded as metadata in MLflow experiment runs. When auditing a specific model, teams can quickly locate the exact dataset version used to train that model through MLflow and restore the data snapshot from that point in time using DVC.

This pattern is suitable for most standardized machine learning projects, has relatively low implementation costs, and can meet basic compliance and audit requirements.

Pattern Two: Record-Level Lineage Tracking

Record-level lineage tracking is more granular, tracking which specific data records participated in model training. This level of tracking is particularly important in fields with strict data provenance requirements, such as financial risk management and healthcare AI. For example, when a batch of data is found to have quality issues, teams can quickly identify affected models and assess whether retraining is necessary.

Record-level lineage tracking is more complex to implement, requiring unique identifiers to be assigned to each record during the data preprocessing stage and associating these identifiers with model versions during training. However, the traceability improvement it delivers is significant, especially when facing data privacy regulations such as the "right to be forgotten" under GDPR, where this capability is nearly indispensable.

Notably, AWS provides accompanying Jupyter Notebooks that users can run directly in their own AWS accounts to quickly validate and deploy both patterns.

Technical Analysis: Why This Combination

From a technology selection perspective, this solution is thoughtfully designed. DVC, as an open-source tool, has a large community ecosystem and naturally fits into Git workflows, reducing the team's learning curve. MLflow is also an important member of the open-source ecosystem, and its managed version on SageMaker eliminates operational overhead. SageMaker, as the underlying platform, provides elastic computing resources and security compliance guarantees.

Compared to proprietary solutions that rely entirely on a single platform, this "open-source tools plus cloud platform" combination gives teams greater flexibility. Teams can use DVC and MLflow for experiments in local development environments and seamlessly migrate to SageMaker for large-scale training, with lineage information remaining consistent throughout the process.

However, this multi-tool collaborative approach also introduces a degree of integration complexity. Teams need to ensure that DVC version identifiers are correctly passed to MLflow records, as any oversight in the chain could break the lineage link. Therefore, establishing standardized pipeline templates and automated validation mechanisms is particularly important.

Outlook: Model Governance Will Become a Core MLOps Capability

As AI regulatory frameworks are gradually implemented worldwide — from the EU AI Act to China's generative AI management regulations — model explainability and traceability are shifting from nice-to-have features to hard requirements. End-to-end lineage tracking is not just a technical best practice but will become foundational infrastructure for enterprise AI compliance.

It is foreseeable that future MLOps platforms will incorporate lineage tracking as a built-in core capability rather than an add-on feature requiring additional integration. The solution AWS showcased provides the industry with a pragmatic reference architecture and sends a clear signal: in the era of AI deployment at scale, knowing where a model comes from is just as important as making it perform better.

For teams currently building or optimizing their MLOps systems, now is the ideal time to evaluate and introduce lineage tracking capabilities.