I use PG as a place to put OLAP cubes generated from primary-source event-sourcing data. In that use-case, PG really doesn't need any kind of durability guarantees, because it's not the canonical store of the data it's holding (i.e. the same data could just be repopulated into a new PG instance in the event of a hardware failure.) PG's default tuning is very sub-optimal for this use-case. :)

(Actually, question for the audience, anyone with experience in heavily-OLAP workloads: what kind of DBMS would you load an OLAP cube into, if it were still too large after projection/dimension-reduction to fit into memory, and you needed to run CPU-intensive, non-embarrassingly-parallel queries on it?)