This should have been part of the original commit, but was missed.
Pushing data between processes is expensive, so we definitely want
to project away unneeded columns here, just as we do for other nodes
like Sort and Hash that care about the volume of data.