Remind me again why I would want to pull consistency guarantees out of an actual datastore (which is designed to provide those guarantees) and force-fit that problem into a fancy caching layer?

The Kleppman critique about efficiency vs correctness is exactly right. Redis for efficiency is great as a work-saving optimization. Redis for correctness (assuming it is possible) asks a remote system to enforce behavior at a far remove from the place where correctness is evident, which is system state, usually data in a datastore.