Open Source Summit + Embedded Linux Conference North America 2026

The transition of Machine Learning (ML) models from experimental notebooks to reliable production environments often reveals a significant disconnect between Data Scientists and Infrastructure/Operations teams. While traditional DevOps has mastered code delivery, the unique "state" of ML—comprising both code and massive datasets—requires a specialized evolution: MLOps. This session provides a practical roadmap for building a simple yet highly scalable CI/CD pipeline using a purely open-source stack.

We begin by addressing the critical challenge of Model Reproducibility. Standard version control systems like Git excel at managing algorithms but fail when handling the 500MB weights or multi-gigabyte training sets typical of modern ML. Our proposed architecture integrates DVC (Data Version Control) to version-control data alongside source code, ensuring that every deployment is fully traceable and repeatable.