At the Large Hadron Collider our datasets will reach the 6-7 PB range near the end of the 2020's. Our current bespoke solutions which integrate the science and data movement will no longer work. We need to shift from the old paradigm to one that separates the science from data location, movement, and processing. Declarative programming languages provide a way to do this - they allow the scientist to specify what they want to do with the data and allows a back-end system to actually build a work plan to execute the work. In this talk I discuss the general idea, some of the approaches we are taking, and some of the stumbling blocks we are currently working with.