RE: Incorrect messages emitted from pgoutput when using column lists

On Friday, November 25, 2022 6:09 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Fri, Nov 25, 2022 at 8:16 AM houzj(dot)fnst(at)fujitsu(dot)com
> <houzj(dot)fnst(at)fujitsu(dot)com> wrote:
> >
> > I think the reason is that we didn't filter the column when sending
> > the old tuple in pgoutput. We thought that the old tuple won't include
> > columns that not in RI, but it seems it will still be null values for
> > such columns in the old tuple.
> >
>
> Yes, that is correct. We do fill null values for non-replica identity columns in the
> old tuple. See ExtractReplicaIdentity.
>
> > So, I think we'd better filter the column for old tuple as well.
> >
>
> Your fix looks correct to me though I haven't tested it yet.
>
> Can we think of writing a test case using
> pg_logical_slot_peek_binary_changes() similar to what we have in
> 020_messages.pl?

Yeah, I think it works.

Besides, I find that we don't filter the column for DELETE as well because
DELETE change only have old tuple. This looks like a similar problem as UPDATE,
and I suppose we need to fix them all. Attach the new version patch which added
the testcase as suggested and fix both DELETE and UPDATE cases.

Best regards,
Hou zj