It has always seemed to me that the solution to the duplication/normalization dilemma is an approach that gets "normalized" data (not necessarily normalized in the sense that it is scattered among many tables, but in the sense that there is a single source of truth for each data point) and uses it to write materialized views, automatically and efficiently recomputed and stored for fast querying.

CouchDB does this to a certain extent, but its JSON-document approach is not appropriate enough.