Re: Support logical replication of DDLs, take2

On Thu, Apr 30, 2026 at 7:03 AM Hannu Krosing <hannuk(at)google(dot)com> wrote:
>
> On Wed, Apr 29, 2026 at 2:10 PM Andres Freund <andres(at)anarazel(dot)de> wrote:
> >
> > Hi,
> >
> > On 2026-04-29 10:07:04 +0200, Hannu Krosing wrote:
> > > On Wed, Apr 29, 2026 at 5:39 AM Dilip Kumar <dilipbalaut(at)gmail(dot)com> wrote:
> > >
> > > > I am trying to understand your idea. If we are trying to deparse from
> > > > an actual system table using a snapshot, why don't we just use the
> > > > WAL? I mean, the WAL should contain the actual catalog modifications
> > > > it has made.
> > >
> > > We have the full data in the catalog and we would likely need catalog
> > > queries for any change, even when de-parsing the tree.
> > >
> > > And we should not add the extra load on the original DDL side, just as
> > > we don't for DML.
> >
> > That can't be a relevant cost compared to everything else.
>
> Probably not. But unless we somehow encode "everything" at that point
> we will make building different DDL decoders harder down the line.
>
> So why not just save the normally serialised parse tree at this point
> and let the decoders decide to do whatever they need.
>
> > > At most we could just serialize the statement tree into the WAL,
> > > though even that may be an overkill if we can get the change from
> > > existing records.
> > >
> > > - insert new row in pg_class --> extract the CREATE TABLE (or INDEX, or ...)
> > > - update row in pg_class or insert, update or delete a row in
> > > pg_attribute --> extract ALTER TABLE
> > > - except when it just updates relfilenod --> extract TRUNCATE
> > > - delete row in pg_class --> DROP TABLE
> > > - dml on pg_constraint --> ALTER TABLE
> > >
> > > ... etc
> >
> > That doesn't work in the general case, think of
> > ALTER TABLE ... ALTER COLUMN ... TYPE foo USING (...)
> >
> > There's a big difference between USING(foo::int8) and USING (pg_size_bytes(foo))
> > but it's nowhere visible in the WAL.
>
> It can't be a big difference if it is not visible in the WAL.

If we send the rewritten tuples made during ALTER TABLE execution via
logical replication, there would not be a big difference. However, if
we send only the re-constructed ALTER TABLE statement, there is. I
think that replicating ALTER TABLE should behave the latter because we
might not need table rewrites in more ALTER TABLE cases in newer
PostgreSQL versions.

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com