Re: BUG #18644: ALTER PUBLICATION ... SET (publish_via_partition_root) wrong/undocumented behavior.

On Fri, Oct 25, 2024 at 6:12 AM Peter Smith <smithpb2250@gmail.com> wrote:
>
> Hi Kuroda-san,
>
> This is only a caution for ALTER of 'publish_via_partition_root', so
> that should be immediately clear up-front in the first few words of
> the text.
>
> Also, the current explanation appears to give more details about how
> and why it happens than I felt was necessary. e.g. this is for user
> documentation, not for a code comment. Basically, I'm wondering if the
> whole thing could be "dumbed down" a little bit, keeping only the
> information essential so the users can protect themselves. Maybe
> something a bit more like below?
>
> SUGGESTION
>      <caution>
>       <para>
>        Altering the <literal>publish_via_partition_root</literal> parameter
>        can lead to data loss or duplication at the subscriber because it
>        changes the identity and schema of the published tables.
>       </para>

This appears precise but lacks the key information that the problem
can happen when a partitioned root table is specified as a replication
target. So, how about one of the following:

* Altering the <literal>publish_via_partition_root</literal> parameter
when the partition root table is specified as the replication target
can lead to data loss or duplication at the subscriber because it
changes the identity and schema of the published tables.

* Altering the <literal>publish_via_partition_root</literal> parameter
can lead to data loss or duplication at the subscriber because it
changes the identity and schema of the published tables. Note this
happens only when the partition root table is specified as the
replication target.

>       <para>
>        This problem can be avoided by refraining from modifying
> partition leaf tables
>        after the <command>ALTER PUBLICATION ... SET</command> until the
>        <link linkend="sql-altersubscription"><command>ALTER
> SUBSCRIPTION ... REFRESH PUBLICATION</command></link>
>       is executed, and by only refreshing using the <literal>copy_data
> = off</literal> option.
>       </para>
>      </caution>
>

We can keep this part as you proposed.

--
With Regards,
Amit Kapila.