There's nothing intrinsic about not supporting joins, in a columnar store; it's just that you lose a huge amount of the linear scanning performance if you have to do joins for each value. Most columnar stores I've used (primarily Impala, SparkSQL and Clickhouse) all support joins, but they materialize one side of the join as an in-memory hash table, which limits the allowable size of the join, and is a cost multiplier for a distributed query. I believe per the docs that MemSQL can mix and match row-based with columnar more easily, but joins are always going to be really slow compared to the speed you can get from scanning the minimum number of columns to answer your question.