●Change8

v3.8.0

📦 mlflowView on GitHub →
✨ 23 features🐛 23 fixes⚡ 2 deprecations🔧 17 symbols

Summary

MLflow 3.8.0 introduces major new capabilities such as prompt model configuration, in‑progress trace display, DeepEval/RAGAS judges integration, and two new conversational scorers, while also adding numerous tracking, tracing, and evaluation enhancements and fixing a wide range of bugs.

Migration Steps

  1. If you manage your own MLflow server, disable UI telemetry by setting `MLFLOW_DISABLE_TELEMETRY=true` or `DO_NOT_TRACK=true`.
  2. If you use a managed MLflow service, you can opt‑out of UI telemetry via the Settings tab in the UI.
  3. Update any Docker Compose commands to replace `--default-artifact-root` with `--artifacts-destination`.
  4. Review code that uses the deprecated Unity Catalog function integration in AI Gateway and migrate to the new approach.
  5. Replace any usage of OpenAI‑flavor model logging with the supported alternatives.

✨ New Features

  • Prompt Model Configuration: prompts can now include model configuration for reproducible LLM workflows.
  • In-Progress Trace Display: Traces UI now shows spans from in-progress traces with auto-polling for real‑time debugging.
  • DeepEval and RAGAS Judges Integration: new get_judge API provides access to 20+ evaluation metrics such as answer relevancy, faithfulness, and hallucination detection.
  • Conversational Safety Scorer: built‑in scorer for evaluating safety of multi‑turn conversations (hate speech, harassment, violence, etc.).
  • Conversational Tool Call Efficiency Scorer: built‑in scorer for evaluating tool‑call efficiency in multi‑turn agent interactions.
  • Add default passphrase support for tracking.
  • Add Pydantic AI Stream support in tracing.
  • Add --max-results option to `mlflow experiments search`.
  • Enhance encryption security for tracking.
  • Fix and simplify Gateway store interfaces.
  • Add inference_params support for LLM Judges.
  • Support batch span export to UC Table.
  • Add endpoint tags.
  • Add MLFLOW_GENAI_EVAL_MAX_SCORER_WORKERS env var to limit concurrent scorer execution.
  • Enable search_datasets in Databricks‑managed MLflow.
  • Render text prompt previews in markdown.
  • Add linked prompts filter for trace search tab.
  • Automatically wrap async functions when passed to `predict_fn`.
  • Conversational Role Adherence builtin judge.
  • Add backend DB tables for Endpoints and related entity definitions, abstract store interface, SQL Store backend, protos and entities interfaces, REST store implementation, credential cache, and provider/model/configuration handling.
  • Add show/hide visibility control for Evaluation runs chart view.
  • Add `mlflow experiments get` command.
  • Simplify secrets and masked secrets with map types in Gateway.

🐛 Bug Fixes

  • Fix GraphQL SearchRuns filter using invalid attribute key in trace comparison.
  • Fix artifact download performance regression.
  • Fix SQLAlchemy alias conflict in `_search_runs` for dataset filters.
  • Add auth support for GraphQL routes.
  • Fix SQL injection vulnerability in UC function execution.
  • Fix MultiIndex column search crash in dataset schema table.
  • Make datasource failures fail gracefully.
  • Fix litellm autolog for versions >= 1.78.
  • Fix SQLAlchemy engine connection pool leak in model registry and job stores.
  • Fix Traces UI filtering on assessments with multiple values.
  • Fix error initialization in Feedback.
  • Switch container build to subprocess for SageMaker.
  • Fix scorers issue on Strands traces.
  • Stop initializing backend stores in artifacts‑only mode.
  • Parallelize multi‑turn session evaluation.
  • Add safe attribute capture for pydantic_ai.
  • Fix UC to UC copying regression.
  • Fix artifact path traversal vector.
  • Fix auth controls on system metrics UI.
  • Add context loading for ChatModel.
  • Fix trace decorators usage for LangGraph async callers.
  • Update docker compose to use `--artifacts-destination` instead of `--default-artifact-root`.
  • Reduce clint error message verbosity by consolidating README instructions.

Affected Symbols

get_judgemlflow.genai.evaluatemlflow experiments getmlflow experiments searchmlflow.tracking._search_runsmlflow.tracking._search_datasetsmlflow.tracingmlflow.tracing.pydantic_aimlflow.tracing.litellm autologmlflow.tracing.langgraph async decoratorsmlflow.scoringmlflow.endpointsmlflow.gatewaymlflow.models.ChatModelmlflow.models.context loadingmlflow.trackingmlflow.tracking.graphql routes

⚡ Deprecations

  • Deprecate Unity Catalog function integration in AI Gateway (see Docs).
  • Deprecate model logging of OpenAI flavor (see Docs).