Something you need to bear in mind is that distributed joins are very expensive; you have a better time designing your schema such that related data can be plac

Something you need to bear in mind is that distributed joins are very expensive; you have a better time designing your schema such that related data can be placed logically close together, whether it's arrays / maps inside rows (for one to many), or very wide rows (denormalizing what might be a star schema). (I know, in a column store having related data in another column isn't actually close together; but it can be stepped through at the same time, it doesn't need a join to be correlated, it's correlated naturally.)