guaibasaurus just failed in a way I'd not seen before:
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=guaibasaurus&dt=2018-03-14%2006%3A33%3A01
Inspection of the diffs makes it depressingly obvious what happened:
the concurrently-executing "alter_table" and "with" scripts both create
short-lived tables named "test". Given the right timing, those tests
stomp on each other. This seems to have been there a long time;
surprising we'd not noticed it before.
Moral: choose names less generic than "test" for globally-visible
objects. I wonder if there's some way we could check for such
conflicts automatically?
regards, tom lane