Re: PostgreSQL12 crash bug report

Andres Freund <andres@anarazel.de> writes:
> On 2019-09-05 12:59:11 -0700, Andres Freund wrote:
>> I tried for a while to develop one for mark/restore of IndexOnlyScans,
>> but I concluded that that code is basically dead right now. Every scan
>> node of a normal that gets modified or needs a rowmark implies having
>> ctid as part of the targetlist. And we neither allow ctid to be part of
>> index definitions, nor understand that we actually kinda know the ctid
>> from within the index scan (HOT would make using the tid hard).  So the
>> relevant code in nodeIndexOnly.c seems dead?

> I wonder if, on master, we should make ExecIndexOnlyMarkPos(),
> ExecIndexOnlyRestrPos() ERROR out in case they're hit for an EPQ
> relation, given that they ought to be unreachable.

I'd vote against.  The chain of reasoning that says they're unreachable
is long and involves mostly code that's nowhere near there, so when/if
somebody made a change that broke that reasoning, they'd not necessarily
notice that the ERROR has to be undone.

            regards, tom lane