Dmitry Ivanov <d.ivanov@postgrespro.ru> writes:
> Tom Lane wrote:
>> Uh, why would you see that? The planner would never generate an
>> IndexOnlyScan in the first place if the query required any columns
>> not available from the index.
> True, but as you can see, create_append_plan() produces its own targetlist:
> static Plan *
> create_append_plan(PlannerInfo *root, AppendPath *best_path)
> {
> Append *plan;
> List *tlist = build_path_tlist(root, &best_path->path);
> ...
> If we replace Append with some custom node, the plan will instantly become
> invalid (it won't be be able to build a projection from 'custom_scan_tlist'
> to 'targetlist'). However, this doesn't mean that it's unable to produce
> the same result.
You haven't really convinced me that anything is wrong there. The append
plan's tlist isn't going to contain unwanted variables either.
Reading between the lines, I think the problem may be that you're not
being careful about how you set up custom_scan_tlist. But since the
core code has zero involvement in that decision, it's hard to see why
it would be a core code bug.
regards, tom lane