With the release of MLflow 3.0, the platform has shifted from a run-centric approach to a model-centric architecture. To ensure Databricks MLOps Stacks remain the "gold standard" for production best practices, the templates should be updated to leverage these new capabilities.
Proposed Changes:
Value: Updating to 3.0 will provide users of MLOps Stacks with better lineage tracking, native evaluation support, and a more intuitive model-first development loop.
With the release of MLflow 3.0, the platform has shifted from a run-centric approach to a model-centric architecture. To ensure Databricks MLOps Stacks remain the "gold standard" for production best practices, the templates should be updated to leverage these new capabilities.
Proposed Changes:
Update ML Code Templates: Refactor training notebooks to utilize the new LoggedModel entity. Specifically, move away from artifact_path in favor of the name parameter in log_model() to enable easier model searching.
Modernize Evaluation: Replace any calls to the deprecated mlflow.evaluate with the specialized mlflow.models.evaluate or mlflow.genai.evaluate APIs.
Streamline UC Integration: Since MLflow 3.0 defaults to databricks-uc, update the databricks.yml and resource configurations to assume Unity Catalog as the primary registry, simplifying the three-level namespace logic.
Dependency Alignment: Update requirements.txt and cluster environment definitions in the stack to target mlflow>=3.0.0.
Remove Legacy Refs: Ensure no references to mlflow.recipes (now removed) remain in the codebase.
Value: Updating to 3.0 will provide users of MLOps Stacks with better lineage tracking, native evaluation support, and a more intuitive model-first development loop.