> On Jun 2, 2026, at 08:41, jian he <[email protected]> wrote:
>
> On Fri, May 29, 2026 at 6:58 AM Zsolt Parragi <[email protected]>
> wrote:
>>
>> Looks good, I have one nitpick question:
>>
>> + if (isnull)
>> + elog(WARNING, "null conbin for relation
>> \"%s\"",
>> + RelationGetRelationName(rel));
>>
>> Shouldn't this be an error instead? Most existing checks treat it as
>> an error, and that seems better to me as this shouldn't happen
>> normally?
>>
> OK.
>
> I changed it to:
>> + if (isnull)
>> + elog(ERROR, "null conbin for relation
>> \"%s\"",
>> + RelationGetRelationName(rel));
>
> I also slightly rephrased the commit message.
>
>
>
> --
> jian
> https://www.enterprisedb.com/
Sorry for not noticing this thread earlier. I also found the same problem while
testing the feature and posted a patch in [1]. Thanks to Jian for pointing out
this thread.
After reading v6 from this thread, I think my implementation is slightly
different:
* v6 adds a new helper, RememberWholeRowDependentForRebuilding(), which reads
pg_constraint and pg_index to find objects to rebuild.
* My implementation in [1] reuses RememberAllDependentForRebuilding(), which
reads pg_depend, because whole-row dependencies have refobjsubid = 0. My v1
missed the partial-index case, so I added a new helper in v2 to read pg_index.
The main difference is that my patch reuses the existing dependency-based
rebuild machinery where possible, and only scans pg_index directly for the
partial-index case where pg_depend does not record the whole-row Var
dependency. It is also smaller: this thread’s v6 adds 185 lines to tablecmds.c,
while my v2 adds 123 lines.
Would you mind taking a look at my implementation in [1]?
BTW, I have also added this problem to the open items list, as I think it
should be fixed before releasing v19.
[1]
https://www.postgresql.org/message-id/71242FEA-15A3-4073-9FA1-561CA744ACBD%40gmail.com
Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/
