Practical MLflow for Generative AI on Databricks

Foreword

We created MLflow after noticing a familiar pattern across machine learning teams: building the model was only part of the work, and often not even the hardest part. The real difficulty involved managing everything around the models: tracking experiments, packaging code, reproducing results, collaborating across teams, deploying reliably, and understanding what was actually running in production.

That was the problem MLflow was designed to address. The goal was to create an open, practical way to manage the lifecycle around machine learning work across different tools, frameworks, and environments. We wanted to make the messy parts of development more manageable, more reproducible, and easier to share across teams.

Since then, the breadth of adoption has been rewarding. MLflow grew because practitioners recognized the problem. They knew what it felt like to work with scattered workflows, inconsistent environments, and irreproducible results. Over time, MLflow grew into a large, open source community with contributors from many more places than we imagined.

If anything, the rise of generative AI has made that original problem even more prominent.

A modern AI application is not just a model call; it is a compound system. It includes prompts, retrieval pipelines, tools, agent skills, feedback loops, serving infrastructure, evaluation, and governance. The quality of the application depends strongly on how all these components work together. These components give teams more ...