Re: WIP patch: reducing overhead for repeat de-TOASTing

Alvaro Herrera <alvherre(at)commandprompt(dot)com> writes:
> Tom Lane wrote:
>> It would be simple enough to fix nodeSubplan.c to copy the data into
>> an upper-level Slot rather than a bare tuple. But this makes me wonder
>> how many other places are like this. In the past there wasn't that much
>> benefit to pulling data from a Slot instead of a bare tuple, so I'm
>> afraid we might have a number of similar gotchas we'd have to track
>> down.

> I compare this to adding CHECK_FOR_INTERRUPTS(): let's declare that
> every usage of bare tuples is a not-very-serious bug, and we can fix
> them one by one as we come across them.

Unfortunately we can't usefully have such a rule --- consider sorting
for example. We're not going to change over to using TupleTableSlots
as the items being sorted. What I foresee if we go down this path
is that there will be some places where we can fix toasting performance
problems by inserting a Slot, and some where we can't.

regards, tom lane