Most of my discussions with data scientists have revolved around exactly this point (usually to serious pushback): that they need to be end-to-end to deliver value in an organization.

One of the points in this list by Paul Osman that I recently stumbled on is that things like “staging”are indeed mostly useless however much people want to keep using them.

A long treatment of how to scale work as organizations grow with some helpful guidelines: “Keep the work parallel, the groups small, and the resources local.”