v3.2.0rc0
📦 mlflowView on GitHub →
✨ 15 features🐛 4 fixes🔧 7 symbols
Summary
MLflow 3.2.0 introduces extensive GenAI tracing enhancements, new feedback tracking, UI improvements, Polars dataset support, and several feature additions and bug fixes, all without breaking changes.
✨ New Features
- Added TypeScript SDK support for MLflow Tracing, enabling GenAI tracing in TypeScript environments.
- Provided automatic tracing support for Semantic Kernel, simplifying trace capture for SK workflows.
- Introduced native feedback tracking for human feedbacks, ground truths, and LLM judges on traces.
- Redesigned MLflow UI experiment home view and added pagination on the model page for better usability.
- Enhanced Trace UI with image rendering for OpenAI, Langchain, and Anthropic chat messages and added a summary view for spans.
- Added PII masking capability via a custom span post-processor in tracing.
- Added support for Polars datasets, expanding compatibility with high‑performance DataFrame libraries.
- Started anonymized usage data collection with opt‑out capability.
- Introduced MLFLOW_DISABLE_SCHEMA_DETAILS environment variable to toggle detailed schema errors.
- Added support for chat‑style prompts with structured output using a prompt object.
- Added support for responses.parse calls in the OpenAI autologger.
- Renamed evaluation metric guideline_adherence to guidelines.
- Added tag filter to the experiments page in the UI.
- Enabled editing of experiment tags directly in the UI.
- Extended spark_udf to support the 'uv' environment manager.
🐛 Bug Fixes
- Fixed proper chat message reconstruction from OpenAI streaming responses.
- Converted the trace column in search_traces() response to a JSON string.
- Resolved crashes in mlflow.evaluate when calling _get_binary_classifier_metrics.
- Various small bug fixes and documentation updates across the project.