On Thu, May 20, 2021 at 5:59 PM osumi.takamichi@fujitsu.com
<osumi.takamichi@fujitsu.com> wrote:
> On Tuesday, May 18, 2021 3:30 PM Amit Langote <amitlangote09@gmail.com> wrote:
> > While doing so, it occurred to me (maybe not for the first time) that we are
> > *unnecessarily* doing send_relation_and_attrs() for a relation if the changes
> > will be published using an ancestor's schema. In that case, sending only the
> > ancestor's schema suffices AFAICS. Changing the code that way doesn't
> > break any tests. I propose that we fix that too.
> I've analyzed this new change's validity.
> My conclusion for this is that we don't have
> any bad impact from this, which means your additional fix is acceptable.
> I think this addition blurs the purpose of the patch a bit, though.
Okay, I've extracted that change into 0002.
> With the removal of the send_relation_and_attrs() of the patch,
> we don't send one pair of LOGICAL_REP_MSG_TYPE('Y'),
> LOGICAL_REP_MSG_RELATION('R') message to the subscriber
> when we use ancestor. Therefore, we become
> not to register or update type and relation for maybe_send_schema()'s
> argument 'relation' with the patch, in the case to use ancestor's schema.
> However, both the pgoutput_change() and pgoutput_truncate()
> have conditions to check oids to send to the subscriber for any operations.
> Accordingly, the pair information for that argument 'relation'
> aren't used on the subscriber in that case and we are fine.
Thanks for checking that.
Here are updated/divided patches.
--
Amit Langote
EDB: http://www.enterprisedb.com