Re: 'ERROR: attempted to update invisible tuple' from 'ALTER INDEX ... ATTACH PARTITION' on parent index

On Wed, Jul 12, 2023 at 11:38:05AM +0530, Dilip Kumar wrote:
> On Wed, Jul 12, 2023 at 11:12 AM Michael Paquier <michael@paquier.xyz> wrote:
>>
>> On Wed, Jul 12, 2023 at 09:38:41AM +0900, Michael Paquier wrote:
>> > While working recently on what has led to cfc43ae and fc55c7f, I
>> > really got the feeling that there could be some command sequences that
>> > lacked some CCIs (or CommandCounterIncrement calls) to make sure that
>> > the catalog updates are visible in any follow-up steps in the same
>> > transaction.
>>
>> Wait a minute.  The validation of a partitioned index uses a copy of
>> the pg_index tuple from the relcache, which be out of date:
>>        newtup = heap_copytuple(partedIdx->rd_indextuple);
>>        ((Form_pg_index) GETSTRUCT(newtup))->indisvalid = true;
> 
> But why the recache entry is outdated, does that mean that in the
> previous command, we missed registering the invalidation for the
> recache entry?

Yes, something's still a bit off here, even if switching a partitioned
index to become valid should use a fresh tuple copy from the syscache.

Taking the test case of upthread, from what I can see, the ALTER TABLE
.. REPLICA IDENTITY registers two relcache invalidations for pk_foo
(via RegisterRelcacheInvalidation), which is the relcache entry whose
stuff is messed up.  I would have expected a refresh of the cache of
pk_foo to happen when doing the ALTER INDEX .. ATTACH PARTITION, but
for some reason it does not happen when running the whole in a
transaction block.  I cannot put my finger on what's wrong for the
moment, but based on my current impressions the inval requests are
correctly registered when switching the replica identity, but nothing
about pk_foo is updated when attaching a partition to it in the last
step where the invisible update happens :/ 
--
Michael