v3.17.1·MLflow 3.17.1
Summary
MLflow 3.17.1 introduces the mlflow.genai.databricks.review_queues module for managing review queues on Databricks and adds Decision API support to make_judge(). It also fixes Claude Code OpenTelemetry trace rendering in the UI and ensures Databricks evaluation feedback correctly records the latest scorer version.
Summary from the captured archive. Follow the original source for full context.
More from mlflow/mlflow
MLflow 3.17.2
MLflow 3.17.2 records dataset, scorer, and agent lineage on runs created by mlflow.genai.evaluate. It also fixes MlflowSparkStudy so executors reseed the sampler and forward batch settings, and stops MlflowStorage from marking Optuna trials finished before their values are saved.
MLflow 3.17.0
MLflow 3.17.0 introduces fine-grained access control for runs and traces, links evaluation metrics directly to managed datasets, and adds support for Databricks Unity Catalog model service resources. Self-hosted SQL tracking servers require a coordinated schema migration via mlflow db upgrade with all writers stopped, as rolling deployments are not supported.
MLflow 2.11.5
Users can now opt in to route Unity Catalog model registry artifact uploads and downloads through the Databricks SDK Files API. This update provides an alternative transfer mechanism for managing model artifacts within Databricks environments.
MLflow 3.16.1
MLflow 3.16.1 removes the default basic-auth admin password from basic_auth.ini as a security fix (GHSA-gq3w-7jj3-x7gr) and adds support for span links in Unity Catalog traces. The release also fixes trace location resolution in Databricks Model Serving, includes a timeout option for the @scorer decorator, and improves web UI access control for non-admin users.
MLflow 3.16.0
MLflow 3.16.0 makes the redesigned trace explorer the default interface, introducing natural-language custom trace views via the MLflow Assistant, session grouping for multi-turn conversations, and span links. The update also adds Unity Catalog model service support for built-in evaluation judges, per-user AI Gateway budget policies, and fail-closed authorization by default.