Csaba Nagy <nagy@ecircle-ag.com> writes:
> So an implementation which optimistically builds the new index
> concurrently while holding no lock, and then hopes for the 3rd
> transaction to be able to get the exclusive lock and be able to swap the
> new index in the place of the old index, and error out if it can't - it
> is perfectly acceptable.
It would maybe be acceptable if there were a way to clean up the mess
after a failure, but there wouldn't be ...
regards, tom lane