MLflow 3.17.0 added fine-grained access control for runs and traces and linked evaluation metrics to managed datasets.

MLflow shipped three releases in October 2026. MLflow 3.17.0 on October 7 added fine-grained access control for runs and traces, linked evaluation metrics directly to managed datasets, and added support for Databricks Unity Catalog model service resources. Self-hosted SQL tracking servers needed a coordinated schema migration with `mlflow db upgrade` and all writers stopped, because rolling deployments were not supported [14].

MLflow 3.17.1 followed on October 13. It introduced the `mlflow.genai.databricks.review_queues` module for managing review queues on Databricks and added Decision API support to `make_judge()`. It also fixed Claude Code OpenTelemetry trace rendering in the UI and made Databricks evaluation feedback record the latest scorer version [11]. MLflow 3.17.2 on October 21 recorded dataset, scorer, and agent lineage on runs created by `mlflow.genai.evaluate`. It fixed `MlflowSparkStudy` so executors reseed the sampler and forward batch settings, and it stopped `MlflowStorage` from marking Optuna trials finished before their values were saved [4].

Tracing of agents was the main theme in articles. Claude products began emitting OpenTelemetry traces natively, and guides showed how to export them to open-source MLflow as the trace backend, including tips for tracking token usage and comparing prompts [9][7]. MLflow tracing and evaluation also featured in Databricks material on building agents with Agent Bricks and on governed agents with Unity AI Gateway [5][16].

Other items showed MLflow inside wider workflows. TinyFish used MLflow for evaluation in a path from post-training on AI Runtime to production serving [13], and Databricks AI Research described native MLflow integration in its AI Runtime experiments [8]. A guide on ai_extract described versioning prompts and schemas in MLflow Prompt Registry, with aliases for deployment and rollback [1]. Databricks Free Edition teaching material listed MLflow among the tools students could use [2]. A community post described Zepto scaling customer support on Databricks and MLflow [6].