The machine learning flywheel is a crucial concept in operationalizing machine learning, which involves creating a feedback loop between decision-making, data collection, organization, and learning. The flywheel consists of four stages: Decide, Collect, Organize, and Learn, which mirror the human experience of making decisions, observing results, combining knowledge, and updating understanding. Successful teams intentionally build tools to manage the entire ML lifecycle, including ownership, agency, and clear dependencies. However, building a great machine learning flywheel is challenging due to the diverse number of tools on the market, lack of interoperability, and need for simple abstractions. Data flows are at the core of the ML flywheel, and managing them can be hard. Feature platforms like Tecton have had an impact by solving part of the top path but leaving out the Collect and Organize stages. To extend data management to the entire ML flywheel, teams should close the loop, establish a unified flywheel data model, and support use-case-specific architectures. A unified ML flywheel can achieve reliable, repeatable, and accurate decision APIs for products, supporting ML product teams using fewer resources and making iteration, building, and deployment faster.