v3.1.0

Breaking Changes

📅 Jun 10, 2025📦 mlflowView on GitHub →

⚠ 4 breaking✨ 25 features🐛 9 fixes🔧 18 symbols

Summary

MLflow 3 introduces a model‑centric GenAI architecture with new LoggedModel entity, prompt optimization, enhanced tracing, and many integrations, while also delivering several breaking changes that necessitate migration steps.

⚠️ Breaking Changes

Prompt registry APIs have been moved under the `mlflow.genai.prompts` namespace; update imports and calls to use `mlflow.genai.prompts` instead of the previous location.
When the tracking URI is set to Databricks and no registry URI is provided, the default registry URI now points to `databricks-uc`; explicitly set `registry_uri` if you relied on the previous default.
`mlflow.evaluate` no longer logs the SHAP explainer; remove any code that expects SHAP artifacts from evaluation runs.
`mlflow.search_trace()` now returns a DataFrame in V3 schema format; adjust column names and data handling to the new schema.

Migration Steps

Update all imports of prompt registry APIs to use the new `mlflow.genai.prompts` namespace.
If you relied on the implicit default registry URI when using Databricks, explicitly set `registry_uri` to the desired value.
Remove any expectations of SHAP explainer artifacts from `mlflow.evaluate` outputs.
Adapt code that consumes `mlflow.search_trace()` results to the new V3 DataFrame schema (e.g., column name changes).
If you lock model dependencies, ensure `uv` is installed and use the new `uv`‑based locking option when calling `mlflow.log_model`.
Review any custom prompt judge integrations to use the updated `mlflow[databricks]` extra.
Verify that any custom tracing logic handles the new request/response preview fields.

✨ New Features

Introduced `LoggedModel` as a first‑class entity for GenAI model tracking.
Comprehensive lineage tracking between models, runs, traces, prompts, and evaluation metrics.
Production‑grade GenAI evaluation with reusable quality scorers for accuracy, cost, and latency.
Human‑in‑the‑loop feedback tracking for model predictions.
Prompt optimization capabilities in the Prompt Registry via `mlflow.genai.optimize_prompt`.
Added custom prompt judges to the `mlflow[databricks]` extra.
Package model environment automatically when registering a model.
New `MlflowSparkStudy` class for Spark‑based study tracking.
`spark_udf` now supports DBConnect and DBR 15.4 / dedicated clusters.
Model dependency locking when logging a model using `uv`.
Client and fluent APIs for `LogLoggedModelParams`.
Support for `predict_stream` in the DSPy flavor and for `ResponsesAgent.predict_stream`.
Tracing now supports custom request/response preview.
Integration for AutoGen > 0.4.
Token tracking support for OpenAI.
Automatic recording of notebook and git metadata in trace metadata.
`search_prompts` function added to list all registered prompts.
Compression support for pyfunc logged models.
Gemini model support added to AI Gateway.
PydanticAI autologging for tracing.
Databricks auth profile can be set via `DATABRICKS_CONFIG_PROFILE` environment variable.
Tracing support for `smolagents`.
Video artifact type support in the UI.
`load_prompt` now accepts an `allow_missing` parameter.
Warning emitted when `mlflow.get_artifact_uri()` is called outside an active run.

🐛 Bug Fixes

Added Databricks App resource for GenAI.
Supported JSON‑string inputs/expectations column in Spark DataFrames.
Avoided generating traces from scorers during evaluation.
Allowed multi‑input modules in DSPy.
Improved error handling for missing tracking URI during `mlflow gc`.
Trace search now avoids spawning threads when `include_spans=False`.
Fixed `global_guideline_adherence` metric calculation.
Logged `Resources` from `SystemAuthPolicy` in `CreateModelVersion`.
Updated `ResponsesAgent` interface to align with new streaming API.

Affected Symbols

mlflow.genai.prompts mlflow.evaluate mlflow.search_trace mlflow.genai.optimize_prompt mlflow[databricks] (custom prompt judges)mlflow.tracking.MlflowSparkStudy mlflow.spark_udf mlflow.logging.lock_uv mlflow.tracing.LogLoggedModelParams mlflow.models.DSPyFlavor mlflow.models.ResponsesAgent.predict_stream mlflow.tracking.search_prompts mlflow.load_prompt mlflow.gateway.GeminiSupport mlflow.tracing.PydanticAIAutologging mlflow.tracing.smolagents mlflow.artifacts.video mlflow.get_artifact_uri