The columnar implementations are cool and all, but they don't really do what you'd expect columnar storage to do—i.e. make each column independently retrievable / make each column effectively its own "single data column plus shared CTID" table for the dataset. Instead, when you fetch one column of a row from e.g. a Citus "USING columnar" table, that still bakes down to the query engine fetching the entire row-tuple for that row (and then discarding most of it), which in turn means that the columnar storage engine must read in + decompress the equivalent compressed blocks for every column participating in that row.

So, unlike in a columnar database, you still have to worry about splitting up your narrow+hot "fact" columns from your wide+cold "dimension" columns, and joining them back together later only when you need them, in order to not be polluting your page cache with the wide+cold column data.