MLflow
Open SourceDeliver High-Quality AI, Fast.
Scores
About
MLflow is an open-source platform for the machine learning lifecycle, created by Databricks and released in 2018. It provides four core components that work independently or together:
Tracking — log parameters, metrics, code versions, and artifacts for every training run through a lightweight API (mlflow.log_param, mlflow.log_metric, mlflow.autolog). A built-in UI lets you compare runs, visualise metrics over time, and reproduce any experiment.
Model Registry — a centralised model store with versioning, stage transitions (Staging → Production → Archived), and annotations. Integrates with CI/CD pipelines via webhooks to trigger downstream actions when a model is promoted.
Models — a standard format for packaging ML models with their dependencies, enabling deployment to REST API servers, AWS SageMaker, Azure ML, Google Cloud Vertex AI, Apache Spark for batch inference, and more.
Projects — a convention for packaging reproducible ML code (conda env, Docker container, or system Python) so any run can be re-executed on any machine.
MLflow 3.x extended the platform to cover generative AI: LLM tracing and observability, prompt versioning, an AI Gateway for unified LLM provider access, and systematic evaluation with 50+ built-in metrics. At 60+ million monthly PyPI downloads, it is the most widely adopted ML experiment tracking tool in the ecosystem.
Self-hosting is straightforward — a single-command tracking server backed by PostgreSQL + S3 covers most team setups. Databricks offers Managed MLflow with enterprise governance and Unity Catalog integration for organisations that prefer a fully managed option.
Key Features
- Experiment tracking: log parameters, metrics, and artifacts for every training run with autologging support
- Model Registry: versioned model store with stage transitions, lineage, and webhook notifications
- Standard model packaging format — deploy to SageMaker, Azure ML, Vertex AI, REST API, or Spark
- MLflow Projects for reproducible training runs across any environment
- LLM tracing and observability for generative AI applications (MLflow 3.x)
- AI Gateway: unified API across LLM providers with cost management and budget alerts
- Plugin system for custom backends, storage, and evaluation judges
- Self-hostable on any infrastructure (SQLite/PostgreSQL + local/S3/GCS/Azure Blob storage)
Pros
- Integrates with a few lines of code — `mlflow.autolog()` captures most frameworks automatically
- Framework and cloud agnostic — works with PyTorch, TensorFlow, scikit-learn, XGBoost, and any Python code
- 60M+ monthly PyPI downloads; largest open-source experiment tracking community
- Flexible self-hosting: start with SQLite on localhost, scale to PostgreSQL + cloud object storage
- Model Registry + webhooks enable clean CI/CD gates between experiment and production
Cons
- Tracking server setup for teams (auth, storage backends, TLS) requires non-trivial infrastructure work
- No native pipeline orchestration — must integrate with Airflow, Prefect, or Dagster for scheduling
- UI becomes slow when comparing hundreds of runs with many logged metrics or large artifacts
- Security is user-configured; a default tracking server has no authentication out of the box
Pricing
Open SourcePossible Stacks
PyTorch ML Training
ProjectTrain and track machine learning models with PyTorch. MLflow logs experiments, tracks hyperparameters, and manages model versions — giving you a reproducible ML workflow from notebook to production.
TensorFlow ML Training
ProjectBuild and train ML models using TensorFlow and Keras. MLflow tracks experiments and manages the model lifecycle, from prototyping in notebooks to production deployment.
Databricks Lakehouse Pipeline
ProjectUnified lakehouse architecture: Databricks runs Spark workloads on Delta Lake — combining the scale of a data lake with ACID transactions of a warehouse. Airflow orchestrates ingestion and transformation jobs; dbt handles SQL-based model layers; MLflow tracks experiments and manages model versions alongside the data pipeline.
Related Tools
Works well with (7)
Integrates with (5)
Learning Resources
No resources yet — check back soon.