mlflow

Start MLflow tracking server

$ mlflow server --host [0.0.0.0] --port [5000]

$ mlflow ui

$ mlflow run [project-uri] -P [param=value]

$ mlflow run https://github.com/[user]/[repo]

$ mlflow experiments create -n [experiment-name]

$ mlflow experiments search

$ mlflow models serve -m [models:/model-name/version] -p [5001]

$ mlflow models build-docker -m [models:/model-name/version] -n [image-name]

mlflow command [options]

mlflow is the CLI for MLflow, an open-source platform for machine learning lifecycle management. It tracks experiments, packages code, and deploys models.
The tracking server stores experiment metadata, parameters, metrics, and artifacts. Use mlflow ui for local development or mlflow server for team deployment.
mlflow run executes MLflow Projects—directories or Git repos with MLproject files defining entry points, parameters, and environments. It ensures reproducibility.
Model serving with models serve creates REST endpoints for predictions. models build-docker packages models as containers. The Models component supports multiple ML frameworks.
Artifacts include datasets, models, and outputs. The tracking server stores references; actual files go to configured storage (local, S3, GCS, Azure Blob).

server

Start tracking server.

Start local tracking UI.

Run MLflow project.

Manage experiments.

Manage runs.

Model deployment.

Manage artifacts.

Run ML recipes.

Manage model deployments.

Diagnose MLflow installation.

--host address

Bind address. Default: 127.0.0.1.

Server port. Default: 5000.

Database URI for experiment/run data.

Default artifact storage location.

Number of gunicorn workers.

-P, --param key=value

Project parameters.

Entry point. Default: main.

Experiment name for the run.

Environment manager: local, conda, virtualenv.

Tracking server default (SQLite) not suitable for production; use PostgreSQL/MySQL. Large artifacts need object storage. Some features require specific Python packages. Model serving uses Flask development server.

MLflow was created at Databricks and open-sourced in June 2018. It became an LF AI & Data Foundation project in 2020. The platform grew from Databricks' internal tools for managing ML workflows. Version 2.0 (2022) added MLflow Recipes and improved model registry. MLflow is widely adopted for experiment tracking and model management.

dvc(1), wandb(1), kubeflow(1)