Buildfarm TAP testing is useless as currently implemented

I challenge anybody to figure out what happened here:
http://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=hornet&dt=2015-07-27%2010%3A25%3A17
or here:
http://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=hamster&dt=2015-07-04%2016%3A00%3A23
or here:
http://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=crake&dt=2015-07-07%2016%3A35%3A06

With no visibility of pg_ctl's output, and no copy of the postmaster log,
there is no chance of debugging intermittent failures like this one.
This isn't entirely the buildfarm's fault --- AFAICS, prove-based testing
has inadequate error reporting by design. If "not ok" isn't enough
information for you, tough beans. (It might help if the farm script
captured the postmaster log after a failure, but that would do nothing
for prove's unwillingness to pass through client-side messages.)

I think we should disable TAP testing in the buildfarm until there is
some credible form of error reporting for it. I've grown tired of
looking into buildfarm failure reports only to meet a dead end.
Aside from the wasted investigation time, which admittedly isn't huge,
there's an opportunity cost in that subsequent test steps didn't get run.

regards, tom lane