I wrote:
> When I checked the behavior of 5d1ff6bd559ea8df, I must have only
> tried it for unshared catalogs. Those are set up by
> RelationCacheInitializePhase3, which is post-authentication, so the
> message comes out and causes regression test failures as expected.
> This is kind of annoying :-(. As noted in elog.c, it doesn't seem
> like a terribly good idea to send WARNING messages while the client
> is still in authentication mode; we can't be very sure that clients
> will react desirably. So we can't fix it by mucking with that.
> One answer is to promote the case to an ERROR. We could (probably) keep
> a bad initfile from becoming a permanent lockout condition by unlinking
> the initfile before reporting ERROR, but this way still seems like a
> reliability hazard that could be worse than the original problem.
After sleeping on it, the best compromise I can think of is to add an
"Assert(false)" after the WARNING report for the shared-catalogs case.
This will make the failure un-missable in any development build, while
not breaking production builds' ability to recover from corner cases
we might not've foreseen.
Of course, if you run an assert-enabled build in production, you might
possibly lose. But that's never been recommended practice.
regards, tom lane