Whoa, cool feature I hadn't heard about. Good to know!

Another approach to this problem I've used in the past is Postgres views, which basically let you create a third table that's either materialized (constructed in advance) or generated on the fly and has columns based on transforms of the first two tables. So in this case you could create a Post model backed by a view table that pulls in the names and other metadata it needs from the text and video tables and merges them into columns that function as a third table in the eyes of the Post model. But there's no actual duplication / denormalization happening, which keeps things clean..