MIME-Version: 1.0
From: =?UTF-8?B?5Y2X5ouT5byl?= <minamitakuya@lifull.com>
Date: Tue, 21 Apr 2026 14:35:35 +0900
Message-ID: 
 <CAP6NMUKYgRJ3Z5ThhtXpYYkRBffT=fAStwQ3_1Cab0tREZ=5NQ@mail.gmail.com>
Subject: Warn on missing replica identity in CREATE/ALTER PUBLICATION
To: pgsql-hackers@lists.postgresql.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: 
 <https://www.postgresql.org/message-id/CAP6NMUKYgRJ3Z5ThhtXpYYkRBffT%3DfAStwQ3_1Cab0tREZ%3D5NQ%40mail.gmail.com>
Precedence: bulk

Hi hackers,

CREATE PUBLICATION silently succeeds even when target tables lack a
usable replica identity, while the publication publishes UPDATE and/or
DELETE. The error only surfaces later at replication time:

  ERROR: cannot delete from table "foo" because it does not have a
  replica identity and publishes deletes

This gap has caused real production incidents =E2=80=94 in one case, a CDC
pipeline using FOR TABLES IN SCHEMA included a table without a primary
key, and replication stalled for hours before the cause was found.

I'd like to propose emitting a WARNING at publication creation/alter
time when this mismatch exists. The check would cover all paths:

- CREATE PUBLICATION ... FOR TABLE / FOR TABLES IN SCHEMA / FOR ALL TABLES
- ALTER PUBLICATION ... ADD/SET TABLE / ADD/SET TABLES IN SCHEMA
- ALTER PUBLICATION ... SET (publish =3D 'update, delete')

The approach I'm considering is a publication-level check that runs
after the final publication state is known, scanning the effective set
of published tables via GetIncludedPublicationRelations() /
GetAllSchemaPublicationRelations() / GetAllPublicationRelations() and
checking each table's replica identity.

I have a working prototype for the FOR TABLE / ADD TABLE paths. A few
open questions before I post a full patch:

1. For FOR ALL TABLES, the check would scan pg_class. Acceptable for
   a DDL operation, or too expensive?

2. Should we cap the number of warnings when many tables are affected?

3. Should this be controllable via a GUC, or is a simple WARNING
   sufficient?

Thoughts welcome.

--
Best regards,