# Metaflow

> Open-source Python framework for building, scaling, and deploying real-life ML, AI, and data science workflows.

Metaflow is an open-source, human-centric Python framework originally developed at Netflix, which open-sourced it in 2019, and now supported by Outerbounds. It lets data scientists and ML/AI engineers write workflows in plain Python, develop them locally, scale them to cloud compute, and deploy them to production orchestrators.

## What It Is

Metaflow is a workflow and ML infrastructure framework covering the lifecycle from notebook prototyping to production. Users define flows made of steps in Python, using any Python libraries for models and business logic. Variables inside a flow are tracked and stored automatically, which supports experiment tracking, versioning, and debugging. It is installed with pip or conda from PyPI or conda-forge.

## From Prototype to Production

Flows can be developed and debugged locally, including from notebooks, then run on cloud resources such as GPUs, multiple cores, and large-memory instances. The documentation describes parallel foreach workloads, gang-scheduled distributed compute, dependency management (including uv support), and checkpointing for long-running tasks. Flows can be deployed to production orchestrators with a single command and can react to events or updated data through reactive orchestration. Results can be visualized with cards, including real-time dynamic cards.

## Bring Your Own Cloud

The Metaflow stack can be deployed on AWS (EKS and S3, or AWS Batch and Step Functions), Azure (AKS and Blob Storage), Google Cloud (GKE and Cloud Storage), or a custom Kubernetes cluster. A browser-based Metaflow Sandbox and a one-click local development stack are available for trying it out.

## Recent Releases

The homepage lists recent highlights including the spin command for incremental flow development (November 2025), recursive and conditional steps for agentic systems (August 2025), custom decorators, uv support, the @checkpoint decorator, configurable flows with Config objects, and programmatic run and deploy APIs. The repository lists version 2.19.39 as the latest release.

## Features
- Pythonic flow and step API
- Local development and notebook support
- Automatic experiment tracking and versioning
- Scale-out to cloud CPUs and GPUs
- Foreach parallelism and gang-scheduled distributed compute
- Dependency management including uv
- Checkpointing with @checkpoint
- One-command production deployment
- Reactive event-triggered orchestration
- Custom decorators and Config objects
- Real-time dynamic cards for visualization
- Recursive and conditional steps for agentic systems
- One-click local development stack

## Integrations
AWS, AWS Batch, AWS Step Functions, Amazon EKS, Amazon S3, AWS Trainium, Azure, Azure Kubernetes Service, Azure Blob Storage, Google Cloud, Google Kubernetes Engine, Google Cloud Storage, Kubernetes, uv, conda-forge, PyPI

## Platforms
CLI, DEVELOPER_SDK, API

## Pricing
Open Source

## Version
2.19.39

## Links
- Website: https://metaflow.org
- Documentation: https://docs.metaflow.org/
- Repository: https://github.com/Netflix/metaflow
- EveryDev.ai: https://www.everydev.ai/tools/metaflow