Should it be used as a database in the way that Postgres would be used as a database? It seems pretty unfit for that role, with its past durability issues and that writes have to be tailored to how the data will be consumed, due to lack of ad-hoc queries.

As something layered on top of the actual source of truth it may be reasonable, writing a materialized version of a costly query that's read very often, sure, but that's an optimization and it'd be competing with Redis for what matters there.