Re: relhassubclass and partitioned indexes

Hi,

On 2018/10/22 11:09, Michael Paquier wrote:
> On Fri, Oct 19, 2018 at 06:46:15PM +0900, Amit Langote wrote:
>> Thanks. Attached a patch to set relhassubclass when an index partition is
>> added to a partitioned index.
>
> Thanks, committed after adding a test with ALTER TABLE ONLY, and
> checking upgrades as well as ATTACH partition for ALTER INDEX and ALTER
> TABLE. In all cases the same code paths are taken.

Thank you for committing with those changes. I didn't know about create
index on "only".

>> /* update pg_inherits, if needed */
>> if (OidIsValid(parentIndexRelid))
>> + {
>> StoreSingleInheritance(indexRelationId, parentIndexRelid, 1);
>>
>> + /* Also, set the parent's relhassubclass. */
>> + SetRelationHasSubclass(parentIndexRelid, true);
>> + }
>
> Calling SetRelationHasSubclass() updates pg_class for this parent
> relation. We would need CommandCounterIncrement() if the tuple gets
> updated, but that's not the case as far as I checked for all code paths
> where this gets called. This would be seen immediately by the way..

I assume you're talking about avoiding getting into a situation that
results in "ERROR: tuple concurrently updated".

Afaics, acquire_inherited_sample_rows() "does" perform CCI, because as the
comment says the parent's pg_class row may already have been updated in
that case:

/* CCI because we already updated the pg_class row in this command */
CommandCounterIncrement();
SetRelationHasSubclass(RelationGetRelid(onerel), false);

In all the other case, SetRelationHasSubclass() seems to be the first time
that the parent's pg_class row is updated.

Thanks,
Amit