●Change8

v2.21.0

Breaking Changes
📦 mlflowView on GitHub →
⚠ 2 breaking✨ 13 features🐛 14 fixes🔧 13 symbols

Summary

MLflow 2.21.0 introduces a redesigned documentation site, a Prompt Registry, FastAPI‑based inference server, enhanced tracing, and numerous new integrations and bug fixes.

⚠️ Breaking Changes

  • Inference server migrated from Flask to FastAPI, breaking custom Flask‑based integrations. Update your client code to use the new FastAPI ASGI endpoints.
  • LangGraph integration had a breaking change; the library now expects updated API signatures. Adjust your LangGraph usage accordingly.

Migration Steps

  1. Switch any custom inference server code from Flask routes to FastAPI routes as described in the new inference server documentation.
  2. If you use LangGraph, update your code to match the new API signatures introduced in this release.
  3. Upgrade to MLflow 2.21.0 and verify that mlflow.doctor fallback works automatically; no code change required.
  4. Refresh imports for txtai integration if you previously imported it directly.
  5. Review usage of ExperimentViewRunsControlsActionsSelectTags to ensure proper loading‑state handling after tag operations.
  6. Avoid using tag values that contain ": " without escaping, or update your code to handle the new truncation‑prevention logic.

✨ New Features

  • Documentation redesign with a new MDX‑based website.
  • Prompt Registry for versioning, tracking, and reusing prompts in GenAI applications.
  • FastAPI Scoring Server replaces the Flask inference server, enabling ASGI‑based scalable inference.
  • Enhanced tracing now supports synchronous/asynchronous generators and auto‑tracing for Async OpenAI.
  • Support for OpenAI Agent SDK auto tracing.
  • Support for MySQL SSL connections with client certificates.
  • Support for ADLS artifact repository.
  • Import and documentation added for txtai integration.
  • User Auth Policy introduced for Pyfunc models.
  • Support for the new Google GenAI SDK.
  • Generating traces from DSPy built‑in compilation and evaluation.
  • mlflow.log_trace API added.
  • ChatAgent LangGraph and LangChain connectors added.

🐛 Bug Fixes

  • Fixed infinite recursion error with warning handler module in Models.
  • Fixed invalid type issue for ModelRegistry RestStore.
  • Fixed ExperimentViewRunsControlsActionsSelectTags not resetting loading state on set‑tag failure.
  • Fixed tag creation truncation when tag values contain ": ".
  • Fixed false alert from AMD GPU monitor.
  • mlflow.doctor now falls back to mlflow‑skinny when mlflow is not found.
  • Handled LangGraph breaking change.
  • Fixed DSPy tracing in serving.
  • Added limit to the length of experiment artifact locations.
  • Fixed build.py to restore specific files.
  • Fixed false alert for ChatModel type hint.
  • Updated Model Registry to use aes256 for S3 communication.
  • Fixed LiteLLM autologging.
  • Fixed ChatCompletionResponse for model serving with Pydantic 1.x.

Affected Symbols

mlflow.log_tracemlflow.doctormlflow.tracking.ExperimentViewRunsControlsActionsSelectTagsmlflow.models.ChatAgentmlflow.models.UserAuthPolicymlflow.models.ChatModelmlflow.models.ChatCompletionResponsemlflow.models.LangGraphmlflow.models.LangChainmlflow.serving.fastapimlflow.prompt_registrymlflow.tracingmlflow.tracking