I just put forward this question, because some discussion in this list
mentioned this issue.
but I don't think record assemble can improve the performance, actually we
just consume the column-oriented record directly. because record assemble
need to parse and generate the necessary data for the result. and it cannot
avoid skip other columns it maybe not used in the query.
and in the Dremel paper, they also wrote in the Observation section: they
should try to consume column data directly.