Data science and machine learning demand innovative visualization approaches to handle the ever-increasing size and complexity of data. Yet, this specialization has led to isolated systems that have limited purposes. These monolithic systems take time to build, are hard to maintain, and have limited research purpose long term. In this talk, I argue that the HCI and visualization research community could foster a more vibrant ecosystem of tools by designing specifications and APIs and building composable systems on top of these declarative interfaces. Composable visualization frameworks can empower data scientists, machine learning engineers, or other end users to quickly create ephemeral data interfaces for their current needs. Specifications are intellectually challenging, yet encourage vibrant ecosystems and more reproducible research.