postgres_fdw vs data formatting GUCs (was Re: [v9.3] writable foreign tables)

I wrote:
> There's a lot left to do here of course. One thing I was wondering
> about was why we don't allow DEFAULTs to be attached to foreign-table
> columns. There was no use in it before, but it seems sensible enough
> now.

Hmm ... the buildfarm just rubbed my nose in a more immediate issue,
which is that postgres_fdw is vulnerable to problems if the remote
server is using different GUC settings than it is for things like
timezone and datestyle. The failure seen here:
http://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=rover_firefly&dt=2013-03-10%2018%3A30%3A00
is basically just cosmetic, but it's not hard to imagine non-cosmetic
problems coming up. For instance, suppose our instance is running in
DMY datestyle and transmits an ambiguous date to a remote running in
MDY datestyle.

We could consider sending our settings to the remote at connection
establishment, but that doesn't seem terribly bulletproof --- what if
someone does a local SET later? What seems safer is to set the remote
to ISO style always, but then we have to figure out how to get the local
timestamptz_out to emit that style without touching our local GUC.
Ugh.

(One more reason why GUCs that affect application-visible semantics are
dangerous.)

regards, tom lane