Scale will still be a problem
The more typecasting you can do as part of ETL the better: pgloader is a good example for this. But as soon as you want to start indexing for queries you're going to want to make tradeoffs.
This approach sounds very much like being able to run prepared queries, at will and via batches (renamed as "rollups" by those clever chaps in the marketing department), in near real time. This will certainly drive up the RAM / server requirements.