Blog Post Submission: Reproducible Model Dependencies with uv and MLflow

[debu-sinha]

Blog Post Submission

Post Type

• Deep Dive
• How-To
• Features

Topics

• Advanced
• Core

Title

Reproducible Model Dependencies with uv and MLflow

Abstract

This post covers MLflow's integration with uv, the fast Python package manager from Astral. When you work in a uv-managed project, MLflow automatically detects uv.lock and pyproject.toml, runs uv export to capture the complete pinned dependency graph, and saves the lockfile as a model artifact for exact environment restoration via uv sync.

The post covers:

1. The reproducibility problem - why pip freeze fails for ML model dependencies (missing transitive deps, environment bleed, no platform markers)
2. Zero-config auto-detection - MLflow detects uv projects automatically, no parameters needed
3. Monorepo support - uv_project_path for projects where the lockfile is in a different directory
4. Dependency groups and extras - uv_groups and uv_extras for selective dependency export
5. How it works - architecture diagram showing the log-time and restore-time flows with graceful pip fallback
6. Environment variables - MLFLOW_UV_AUTO_DETECT and MLFLOW_LOG_UV_FILES for fine-grained control

Target Length

~1500 words

Related Artifacts

• Core uv support: Add UV package manager support for automatic dependency inference mlflow#20344 (+1,780 lines, merged)
• Dependency groups/extras: Add uv_groups and uv_extras params for uv dependency group control mlflow#20935 (merged)
• Documentation and examples: Add uv dependency management guide to model dependencies docs mlflow#21528 (open)
• uv docs: uv documentation

Provenance

• Implementation: @debu-sinha
• Code review: @WeichenXu123, @harupy
• Blog requested by: @WeichenXu123 on PR #21528

Consent Acknowledgment

• I have obtained consent from all individuals/organizations mentioned by name