I wrote:
> Yeah, this is sufficient reason why we must use the more invasive
> patch on those branches. What I'm wondering now is if there's a
> way to break even-older branches based on failure to handle dropped
> columns here.
After tracing through the example in v11, I see why those branches
are not broken: when ExecBRUpdateTriggers decides to return the
trigger-returned tuple, it sticks it into a target slot like this:
/*
* Return the modified tuple using the es_trig_tuple_slot. We assume
* the tuple was allocated in per-tuple memory context, and therefore
* will go away by itself. The tuple table slot should not try to
* clear it.
*/
TupleTableSlot *newslot = estate->es_trig_tuple_slot;
TupleDesc tupdesc = RelationGetDescr(relinfo->ri_RelationDesc);
if (newslot->tts_tupleDescriptor != tupdesc)
ExecSetSlotDescriptor(newslot, tupdesc);
ExecStoreTuple(newtuple, newslot, InvalidBuffer, false);
So the slot that ExecConstraints et al will be working with contains
the relation's actual tuple descriptor, not the approximate descr
obtained by looking at the plan tlist.
This logic is entirely gone in v12, which confirms my instinct that
there was something about Andres' slot-manipulation changes that
broke this scenario. In v12 we end up using the junkfilter's output
slot, which does not have a sufficiently accurate tupdesc to deal with
an on-disk tuple rather than one constructed by the executor.
So I'll go see about back-patching 20d3fe900.
regards, tom lane