Metaflow is a human-friendly framework for building, managing and deploying real-life machine learning and AI systems, originally created at Netflix. It wraps the unglamorous infrastructure of data science โ compute, data versioning, orchestration and deployment โ in clean Python, and has about 10,300 stars on GitHub with active development in 2026.
You write a workflow as a Python flow, run it on your laptop, and Metaflow can scale the same code unchanged to cloud batch compute, Kubernetes or workflow schedulers, with every run’s code, data and models automatically versioned and inspectable. Built-in features cover the ML lifecycle: namespaced experiments so teams never collide, resume from failed steps, a client API to query past results, and Cards that render live visualizations and reports for each run. It integrates with AWS, Azure and GCP services underneath while keeping the developer experience identical everywhere, and it now supports generative AI and LLM workloads alongside classical modeling. Metaflow is maintained by the team at Outerbounds together with the open-source community. Metaflow’s philosophy is that infrastructure should stay invisible until you need it: a decorator opts a step into more memory, a GPU or a thousand parallel executions, and the framework handles provisioning, data movement and failure recovery behind the scenes. Netflix developed it to support hundreds of production models, and it now underpins machine learning at many other organizations, with commercial support available from Outerbounds for teams that want a managed experience on the same open framework.
Metaflow is free and open source under the Apache 2.0 license. You pay only for the cloud infrastructure your flows run on.
Metaflow is for data scientists and ML engineers who want Netflix-grade workflow infrastructure without giving up a notebook-friendly, Python-first way of working.