●Change8

v3.11.0rc0

Breaking Changes
📦 mlflowView on GitHub →
⚠ 1 breaking✨ 7 features🔧 6 symbols

Summary

MLflow 3.11.0rc0 introduces major features like AI-powered issue detection, Gateway budget controls, and a Trace Graph View. This release also enhances security with pickle-free model serialization and adds native OpenTelemetry GenAI convention support.

⚠️ Breaking Changes

  • The MLflow TypeScript SDK packages have been renamed to use npm organization scoping. Update your `package.json` dependencies and import statements: `mlflow-tracing` must be changed to `@mlflow/core`, `mlflow-openai` to `@mlflow/openai`, `mlflow-anthropic` to `@mlflow/anthropic`, and `mlflow-gemini` to `@mlflow/gemini`. All packages are now at version `0.2.0`.

Migration Steps

  1. If using the TypeScript SDK, update `package.json` dependencies and import statements: `mlflow-tracing` -> `@mlflow/core`, `mlflow-openai` -> `@mlflow/openai`, `mlflow-anthropic` -> `@mlflow/anthropic`, `mlflow-gemini` -> `@mlflow/gemini`.
  2. For users concerned with security, review documentation on migrating existing models to pickle-free formats for production deployments.

✨ New Features

  • Automatic Issue Identification feature added to the traces table, allowing users to analyze selected traces for quality issues (correctness, safety, performance) via an AI-powered "Detect Issues" button.
  • Gateway Budget Alerts & Limits introduced, allowing configuration of spending limits by time window (daily, weekly, monthly), setting up alerts, and automatically blocking requests to prevent runaway costs.
  • Interactive Trace Graph View added for visualizing complex trace hierarchies and debugging multi-level trace structures.
  • Native support for OpenTelemetry GenAI Semantic Conventions when exporting traces via OTLP (requires `MLFLOW_ENABLE_OTEL_GENAI_SEMCONV` enabled).
  • Opencode CLI integration added for tracking and analyzing code execution flows directly from the development workflow.
  • Automatic dependency inference now supports the UV package manager, capturing exact, locked dependencies from lockfiles when logging models.
  • Support for pickle-free model serialization using `torch.export` and `skops` formats, with enhanced security controls when `MLFLOW_ALLOW_PICKLE_DESERIALIZATION=False`.

Affected Symbols

mlflow-tracingmlflow-openaimlflow-anthropicmlflow-geminitorch.exportskops