Andrew Dunstan <andrew@dunslane.net> writes:
> If this were a significant risk wouldn't we have seen many such failures
> before now?
Hard to tell. It's possibly architecture-dependent, for one thing
(MAXALIGN will affect space availability). Since this happened in a
parallel regression run, it could also be a matter of timing relative to
the concurrent tests. I've often thought that we are not getting as much
mileage out of the parallel-testing facility as we could, because it's
really not exercising variations in timing all that much. It'd be
interesting to throw in a small random delay at the start of each member
of a concurrent set of tests.
regards, tom lane