MLflow Production Cheat Sheet

MLflow for ML experiment tracking and model deployment — experiment tracking, model registry, project packaging, and serving models in production with MLflow.

Last Updated: May 1, 2025

Experiment Tracking

mlflow.set_experiment('my-experiment')

Create or set active experiment

mlflow.start_run()

Begin a tracked run — all subsequent logging goes here

mlflow.log_params({'lr': 0.001, 'batch_size': 32})

Log hyperparameters — key-value pairs

mlflow.log_metric('accuracy', 0.95, step=epoch)

Log metrics — can log per-step during training

mlflow.log_artifact('model.pkl')

Log files — models, plots, configs

mlflow.autolog()

Auto-log from PyTorch, TensorFlow, sklearn, XGBoost, LightGBM — one line magic

mlflow.end_run()

End the current run

Model Registry

Item	Description
`Register Model`	mlflow.register_model(model_uri, name) — promote from experiment to registry
`Versioning`	Each registration creates a new version — immutable, auditable
`Stages`	None → Staging → Production → Archived. Promote via UI or API.
`Stage Transitions`	mlflow.transition_model_version_stage(name, version, 'Production')
`Aliases`	'champion'/'challenger' — newer than stages, more flexible. Assign via UI/API.
`Model Signature`	Input/output schema — enforced at serving time. Add with infer_signature().

MLflow Projects

Item	Description
`MLproject File`	Define environment (conda/docker), entry points, parameters — reproducible runs
`mlflow run .`	Run current project — auto-creates conda env or Docker container
`mlflow run -P param=value`	Pass parameters — overrides MLproject defaults
`GitHub Integration`	mlflow run [email protected]:user/repo — run projects from Git
`Multi-step Workflows`	Compose projects: preprocess → train → evaluate → register

Serving Models

mlflow models serve -m runs:/run-id/model --port 5000

Serve model as REST API locally

mlflow models build-docker -m models:/name/version

Build Docker image for deployment

mlflow deployments create -t sagemaker -m models:/name/1

Deploy to cloud: SageMaker, Azure ML, Databricks

POST /invocations

REST endpoint — send DataFrame JSON, get predictions

--env-manager local

Skip conda — use existing environment (faster, simpler for Docker)

Pro Tip: Log EVERYTHING — parameters, metrics, artifacts, environment. Disk is cheap; the time you spend wondering 'what hyperparameters did I use?' is expensive.