MLflow Tracking

MLflow gives you experiment tracking, a model registry, and (since 2.14+) first-class LLM observability — all from one Python library + UI. Unlike DVC it does require a tracking backend (file / SQLite / server), but it gives you a real dashboard and multi-user collaboration in return.

This skill is opinionated about the three deployment modes that actually get used in practice, with a vendored production stack you can copy into any project. It defers to the official docs for everything else.

When to use

• User wants to track ML experiments (params, metrics, artifacts) with a UI
• User mentions mlflow.start_run, mlflow.log_metric, mlflow.set_tracking_uri, MLFLOW_TRACKING_URI, mlflow ui
• User wants framework autologging (sklearn / PyTorch / Lightning / XGBoost / LightGBM / Keras / TensorFlow / Transformers / spark)
• User wants LLM trace observability (OpenAI, Anthropic, LangChain, LlamaIndex, DSPy, AutoGen, CrewAI, etc.)
• User wants to spin up a self-hosted tracking server with PostgreSQL + MinIO (production)
• User wants a model registry with aliases (Champion / Challenger / Production)
• User asks "how do I compare runs", "where do my logged params go", "how do I serve a logged model"

When NOT to use