[vtuulos]

one of the biggest benefits of HLL is that if you are only interested in set cardinalities, you don't have to store the original items at all, just a small HLL data structure which can be used to compute set unions, and with some caveats, intersections.

It would be really great if Redshift supported HLL as a proper data type, and not just an optimization for COUNT(). A proper data type would allow us to store pre-aggregated data instead of billions of unnecessarily granular rows.