Daniel Farina <daniel@heroku.com> writes:
> This contains some edits to comments that referred to the obsolete and
> bogus TupleDesc scanning. No mechanical alterations.
Applied with some substantial revisions. I didn't like where you'd put
the apply/restore calls, for one thing --- we need to wait to do the
applies until we have the PGresult in hand, else we might be applying
stale values of the remote's GUCs. Also, adding a call that could throw
errors right before materializeResult() won't do, because that would
result in leaking the PGresult on error. The struct for state seemed a
bit of a mess too, given that you couldn't always initialize it in one
place. (In hindsight I could have left that alone given where I ended
up putting the calls, but it didn't seem to be providing any useful
isolation.)
regards, tom lane