Metaflow
Open-source Python framework from Netflix for building, versioning, scaling and deploying ML, AI and data science workflows
Overview
Metaflow is "a framework for real-life ML, AI, and data science." You write a workflow as a plain Python class of steps, and Metaflow versions every variable and artifact automatically, manages dependencies from laptop to cloud, scales single steps to GPUs or large machines, and deploys a flow to a production scheduler with one command, with event triggers between flows. It runs on AWS, Azure, Google Cloud or your own Kubernetes. Netflix built it for its data scientists and open-sourced it in December 2019; it is Apache-2.0 licensed and the latest release is 2.19.41 (November 6, 2026). Outerbounds, the company founded by Metaflow's creators that sells a managed Metaflow platform, announced on April 29, 2026 that it is joining Anaconda, which says Metaflow stays open source.
The Verdict
Who Should Use Metaflow?
Best For
- Data scientists who want production workflows in plain Python, without learning an orchestrator
- Teams on AWS, Azure, GCP or Kubernetes that want cloud scaling per step
- Projects that need every run and artifact versioned for free
Not Ideal For
- Rich experiment dashboards and charts (pair it with a tracker such as MLflow or W&B)
- Teams that need flexible failure handling inside a flow
What's Great
- Plain Python flows feel easier for data scientists than Airflow or Kubeflow
- Automatic versioning of code, data and artifacts for every run
- One-command deploy from prototype to a production scheduler
- Battle-tested at Netflix; users include 23andMe, CNN, REA Group and Realtor.com
Watch Out For
- Opinionated flow model: a failed step stops the flow
- No native scheduler: production runs rely on an external orchestrator
- Managed offering is in flux after the Anaconda deal; no public managed price
Pricing
View all features & details
Key Features
- Workflows as Python classes with steps and branches
- Automatic versioning and storage of artifacts
- Per-step compute (GPUs, cores, memory) in the cloud
- Dependency management from laptop to cloud
- Event triggers between flows
- Data warehouse access
Where It Runs
- AWS, Azure, Google Cloud
- Any Kubernetes cluster
- Local development on a laptop
Real-World Usage
Named Users
- Netflix (origin)
- 23andMe, CNN, REA Group, Realtor.com
- "Hundreds" of companies, per the official site
Project Health
- Releases every week or two through late 2026
- Outerbounds joining Anaconda (announced April 29, 2026); Metaflow stays open source
How It Compares
| Feature | Metaflow | Kubeflow | ZenML | MLflow |
|---|---|---|---|---|
| Main job | Python ML workflows with versioning | Kubernetes-native ML platform | Pipelines and agent ops | Tracking, registry, GenAI tracing |
| Infrastructure | AWS, Azure, GCP, Kubernetes | Kubernetes only | Any stack | Any (self-host or managed) |
| Reviewer view of learning curve | Low for data scientists | Steep | Steep past tutorials | Moderate (server ops) |
| License | Apache-2.0 | Apache-2.0 | Apache-2.0 | Apache-2.0 |
Sources: metaflow.org · GitHub · I Programmer (2019) · Valohai
Is this your tool? Add a badge