Robert Haas <robertmhaas@gmail.com> writes:
> If we're remapping the varattnos, I don't see how we can just pass
> whole-row references through. I mean, if the partition and the parent
> have different varattnos, then a whole-row attribute for one is a
> different thing from a whole-row attribute for the other; the
> HeapTuple you would need to build in each case is different, based on
> the column order for the relation you're worrying about.
There is longstanding code in the planner to handle this for traditional-
inheritance cases; IIRC what it does is build a ROW() expression that
emits the proper output rowtype regardless of the discrepancies.
Not sure why that's apparently not getting applied for partitioning.
> (Boy, our implementation of DROP COLUMN is painful! If we really got
> rid of columns when they were dropped we could've avoided this whole
> mess.)
I think the pain arises mostly from the decision to allow partitions
to not all have identical rowtype. I would have lobbied against that
choice if I'd been paying more attention at the start ... but I wasn't.
regards, tom lane