On 2018/10/19 16:47, Michael Paquier wrote:
> On Fri, Oct 19, 2018 at 01:05:52PM +0900, Amit Langote wrote:
>> As I said above, the price of removing relhassubclass might be a bit
>> steep. So, the other alternative I mentioned before is to set
>> relhassubclass correctly even for indexes if only for pg_partition_tree to
>> be able to use find_inheritance_children unchanged.
>
> Playing devil's advocate a bit more... Another alternative here would
> be to remove the fast path using relhassubclass from
> find_inheritance_children and instead have its callers check for it :)
Yeah, we could make it the responsibility of the callers of
find_all_inheritors and find_inheritance_children to check relhassubclass
as an optimization and remove any reference to relhassubclass from
pg_inherits.c. Although we can write such a patch, it seems like it'd be
bigger than the patch to ensure the correct value of relhassubclass for
indexes, which I just posted on the other thread [1].
> Anyway, it seems that you are right here. Just setting relhassubclass
> for partitioned indexes feels more natural with what's on HEAD now.
> Even if I'd like to see all those hypothetical columns in pg_class go
> away, that cannot happen without a close lookup at the performance
> impact.
Okay, I updated the patch on this thread.
Since the updated patch depends on the correct value of relhassubclass
being set for indexes, this patch should be applied on top of the other
patch. I've attached here both.
Another change I made is something Robert and Alvaro seem to agree about
-- to use regclass instead of oid type as input/output columns.
Thanks,
Amit
[1]
https://www.postgresql.org/message-id/85d50b48-1b59-ae6c-e984-dd0b2926be3c%40lab.ntt.co.jp
