Re: BUG #17558: 15beta2: Endless loop with UNIQUE NULLS NOT DISTINCT and INSERT ... ON CONFLICT

Michael Paquier <michael(at)paquier(dot)xyz> writes:
> On Tue, Jul 26, 2022 at 11:17:45AM +0800, Richard Guo wrote:
>> Yeah, I can see the same problem. _bt_check_unique() can catch violation
>> of unique index constraint successfully, so insert without 'ON CONFLICT'
>> has no problem. But ExecCheckIndexConstraints() fails to tell the NULL
>> tuple violates unique constraint, and the logic below is quite
>> suspicious to me.

> This is still an open item assigned to Peter, and beta3 is planned for
> the 11th of August. Would a resolution by the beginning of next week
> be doable?

Since nothing seems to be happening here, I looked into it. I think
Richard has correctly fingered the buggy code: we need to teach
check_exclusion_or_unique_constraint() that a NULL isn't necessarily
an instant pass. As attached.

Since the presented test case caused an uninterruptible loop, I also added
a CHECK_FOR_INTERRUPTS() that could stop it. (I was initially pretty
surprised that we could get through an index scan without hitting any CFI,
but on reflection it's less surprising, because no index scan actually
happens: check_exclusion_or_unique_constraint() short-circuits things
before very much code has been reached.) I'm kind of inclined to
back-patch the nodeModifyTable.c change all the way, just for safety.

regards, tom lane