This is a complete, hands-on MLflow cheat sheet: install, configure, operate, automate and secure it - with copy-paste commands grouped by task. Bookmark it for day-to-day work, incidents and interviews alike.
MLflow tracks ML experiments (params, metrics, artifacts), packages models, and provides a model registry with stages - bringing DevOps-style versioning and reproducibility to machine learning.
Why you reach for it
- Reproducible experiments with tracked params and metrics
- A model registry with staging/production promotion
- Framework-agnostic packaging and serving
Installation
pip install mlflow
mlflow server --host 0.0.0.0 --port 5000 --backend-store-uri sqlite:///mlflow.dbConfiguration reference
Track a run and log a model.
import mlflow
mlflow.set_tracking_uri('http://localhost:5000')
with mlflow.start_run():
mlflow.log_param('lr', 0.01)
mlflow.log_metric('accuracy', 0.94)
mlflow.sklearn.log_model(model, 'model', registered_model_name='fraud')Core commands you use daily
mlflow models serve -m 'models:/fraud/Production' -p 5001
curl -X POST http://localhost:5001/invocations -H 'Content-Type: application/json' -d '{"inputs": [[1,2,3]]}'Server & tracking
mlflow server --host 0.0.0.0 --port 5000 --backend-store-uri postgresql://... --artifacts-destination s3://mlflow
export MLFLOW_TRACKING_URI=http://localhost:5000
mlflow experiments create -n fraudModels & serving
mlflow models serve -m 'models:/fraud/Production' -p 5001
mlflow models build-docker -m models:/fraud/Production -n fraud:latest
curl -X POST http://localhost:5001/invocations -H 'Content-Type: application/json' -d '{"inputs": [[1,2,3]]}'Use it in CI/CD
Log evaluation metrics from a training pipeline.
- run: |
python train.py # logs params/metrics to MLflow
mlflow models build-docker -m models:/fraud/Production -n fraud:latestBest-practice checklist
- Use a database backend + object storage for artifacts
- Gate registry promotion on evaluation metrics
- Version data and code alongside the model run
Common pitfalls to avoid
- Using the local file store in production instead of a real backend
- No artifact store, so large models are lost
- Promoting models to Production without an evaluation gate
Security quick wins
- Put the tracking server behind auth and TLS
- Restrict who can promote models to Production
Works well with
- Kubeflow - The ML toolkit for Kubernetes - pipelines, training operators and model serving on your cluster.
- Python for DevOps - The glue language of DevOps - automation scripts, CLIs, API clients and operators.
- Kubernetes - Declarative orchestration for containers at scale - self-healing, rolling updates and service discovery.
A cheat sheet is muscle memory on paper - keep it close until the commands live in your fingers.
Want senior help putting MLflow to work - or a team bootcamp on it? CodeWizard runs Israel's largest DevOps community and embeds named senior engineers with a transparent hours log. Book a free consult or explore the Academy.