On Fri, 23 Feb 2024 at 00:12, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> So after studying this for awhile, I see that the planner is emitting
> a PlanRowMark that presumes that the UNION ALL subquery will be
> scanned as though it's a base relation; but since we've converted it
> to an appendrel, the executor just ignores that rowmark, and the wrong
> things happen. I think the really right fix would be to teach the
> executor to honor such PlanRowMarks, by getting nodeAppend.c and
> nodeMergeAppend.c to perform EPQ row substitution.
Yes, I agree that's a much better solution, if it can be made to work,
though I have been really struggling to see how.
> the planner produces targetlists like
>
> Output: src_1.val, src_1.id, ROW(src_1.id, src_1.val)
>
> and as you can see the order of the columns doesn't match.
> I can see three ways we might attack that:
>
> 1. Persuade the planner to build output tlists that always match
> the row identity Var.
>
> 2. Change generation of the ROW() expression so that it lists only
> the values we're going to output, in the order we're going to
> output them.
>
> 3. Fix the executor to remap what it gets out of the ROW() into the
> order of the subquery tlists. This is probably do-able but I'm
> not certain; it may be that the executor hasn't enough info.
> We might need to teach the planner to produce a mapping projection
> and attach it to the Append node, which carries some ABI risk (but
> in the past we've gotten away with adding new fields to the ends
> of plan nodes in the back branches). Another objection is that
> adding cycles to execution rather than planning might be a poor
> tradeoff --- although if we only do the work when EPQ is invoked,
> maybe it'd be the best way.
>
Of those, option 3 feels like the best one, though I'm really not
sure. I played around with it and convinced myself that the executor
doesn't have the information it needs to make it work, but I think all
it needs is the Append node's original targetlist, as it is just
before it's rewritten by set_dummy_tlist_references(), which rewrites
the attribute numbers sequentially. In the original targetlist, all
the Vars have the right attribute numbers, so it can be used to build
the required projection (I think).
Attached is a very rough patch. It seemed better to build the
projection in the executor rather than the planner, since then the
extra work can be avoided, if EPQ is not invoked.
It seems to work (it passes the isolation tests, and I couldn't break
it in ad hoc testing), but it definitely needs tidying up, and it's
hard to be sure that it's not overlooking something.
Regards,
Dean