Netflix open‑sources Metaflow at scale
Netflix’s Metaflow MLOps framework has been made fully open‑source and is described as powering 3,000+ AI/ML projects with local‑to‑cloud transitions, auto‑versioning and GPU orchestration. The post highlights Metaflow’s production pedigree at large organizations and its focus on scaling experiment and pipeline management (x.com).
Metaflow is now being pitched as fully open source, with Netflix saying the machine learning framework runs more than 3,000 internal artificial intelligence and machine learning projects. (github.com) Metaflow is software for managing the steps of an artificial intelligence project, from a laptop experiment to a cloud job, while tracking code, data, and results so teams can rerun work later. Netflix’s documentation says it handles local development, cloud execution, deployment, and artifact versioning in one Python-based workflow system. (docs.metaflow.org, metaflow.org) Netflix first open-sourced Metaflow in December 2019 with Amazon Web Services, describing it then as a framework to help data scientists use cloud scale without rebuilding their tools around infrastructure. The current GitHub repository says the project now supports “hundreds of millions” of compute jobs at Netflix and manages “tens of petabytes” of models and artifacts. (aws.amazon.com, github.com) The change is less about a brand-new code release than about how much of Netflix’s production setup is now described as available in the open. A Netflix extension package on Python Package Index had previously said internal development continued around infrastructure-specific extensions after the 2019 open-source release. (pypi.org, github.com) That matters in a market crowded with tools that promise to move artificial intelligence projects from prototype to production. Metaflow’s pitch is that researchers write ordinary Python, then send heavy jobs to cloud compute or graphics processors without rewriting the whole pipeline for another system. (metaflow.org, docs.metaflow.org) Netflix has spent the past two years publicly tying Metaflow to more of its internal machine learning stack. In 2024, the company said Maestro, its workflow orchestrator, powers nearly every machine learning and artificial intelligence system at Netflix and serves as a backbone for Metaflow itself. (netflixtechblog.com) The company has also framed Metaflow as infrastructure for a wide range of systems, not just model training. A 2024 Netflix engineering post said the Machine Learning Platform team uses the broader Metaflow ecosystem to support recommendation, content understanding, and other production systems across the company. (netflixtechblog.com) Outside Netflix, Metaflow’s own project pages list users including Amazon, DoorDash, Dyson, Goldman Sachs, and Ramp. The Kubeflow community, in a February 2026 integration post, described Metaflow as an actively developed open-source project maintained by multiple organizations, including Netflix and Outerbounds. (github.com, blog.kubeflow.org) The practical test will be whether outside teams can reproduce more of Netflix’s “production pedigree” without relying on Netflix-only glue code. Netflix’s own numbers suggest the framework is no longer a side project: it is already operating at a scale measured in thousands of projects, petabytes of data, and hundreds of millions of jobs. (github.com, pypi.org)
