On Thursday, March 16, 2023 12:30 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
>
> On Wed, Mar 15, 2023 at 2:00 PM houzj.fnst@fujitsu.com
> <houzj.fnst@fujitsu.com> wrote:
> >
> > I noticed that there are some duplicated codes in pgoutput_change()
> function
> > which can be simplified, and here is an attempt to do that.
> >
>
> For REORDER_BUFFER_CHANGE_DELETE, when the old tuple is missing, after
> this patch, we will still send BEGIN and do OutputPluginWrite, etc.
> Also, it will try to perform row_filter when none of old_slot or
> new_slot is set. I don't know for which particular case we have s
> handling missing old tuples for deletes but that might require changes
> in your proposed patch.
I researched this a bit. I think the old tuple will be null only if the
modified table doesn't have PK or RI when the DELETE happens (referred to
the heap_delete()), but in that case the DELETE won't be allowed to be
replicated(e.g. the DELETE will either error out or be filtered by table level
filter in pgoutput_change).
I also checked this for system table and in that case it is null but
reorderbuffer doesn't forward it. For user_catalog_table, similarily, the
DELETE should be filtered by table filter in pgoutput_change as well.
So, I think we can remove this check and log.
And here is the new version patch which removes that for now.
Best Regards,
Hou zj
