MIME-Version: 1.0
References: 
 <CA+3i_M8eT6j8_cBHkYykV-SXCxbmAxpVSKptjDVq+MFtpT-Paw@mail.gmail.com>
In-Reply-To: 
 <CA+3i_M8eT6j8_cBHkYykV-SXCxbmAxpVSKptjDVq+MFtpT-Paw@mail.gmail.com>
From: shveta malik <shveta.malik@gmail.com>
Date: Wed, 22 Apr 2026 09:57:10 +0530
Message-ID: 
 <CAJpy0uD3Tre0jEBKZoVsxgdHQGTVTXW5hGmiOxB7TG=eWAABGw@mail.gmail.com>
Subject: Re: Fix pg_upgrade to detect invalid logical replication slots on
 PG19
To: Lakshmi N <lakshmin.jhs@gmail.com>,
 Sawada Masahiko <sawada.mshk@gmail.com>
Cc: "pgsql-hackers@lists.postgresql.org" <pgsql-hackers@lists.postgresql.org>,
	shveta malik <shveta.malik@gmail.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: 
 <https://www.postgresql.org/message-id/CAJpy0uD3Tre0jEBKZoVsxgdHQGTVTXW5hGmiOxB7TG%3DeWAABGw%40mail.gmail.com>
Precedence: bulk

On Mon, Apr 20, 2026 at 2:28=E2=80=AFPM Lakshmi N <lakshmin.jhs@gmail.com> =
wrote:
>
> Hi Hackers,
>
> The PG19-optimized slot catchup query uses a CTE that filters on
> invalidation_reason IS NULL, then cross-joins it with the main slot
> query.  When ALL logical slots in a database are invalid, the CTE
> returns zero rows, and the cross join produces an empty result set.
> This causes pg_upgrade to silently skip those slots entirely --
> neither detecting them as invalid (which should block the upgrade)
> nor attempting to migrate them.
>
> The pre-PG19 query path does not have this problem because it queries
> pg_replication_slots directly without a cross join. This may not impact
> upgrade to PG19 but will change the behavior for PG20 upgrade.
>
> Fix by changing the cross join to a LEFT JOIN,
> so that invalid slots still appear in the result set with NULL
> caught_up values.
>

I agree with the problem here.

Another way to solve this would be using a scalar subquery(see [1]),
but that would reduce readability. Thus, I prefer a LEFT OUTER JOIN on
TRUE here. There should also be no performance impact, since the
right-hand side query will always return at most one row due to the
LIMIT 1 clause. So IMO, the proposed solution is good. Copying
Sawada-san, as he was the author of the original patch.


[1]:
SELECT slot_name, plugin, two_phase, failover,
CASE
  WHEN invalidation_reason IS NOT NULL THEN FALSE
  ELSE (
    (SELECT last_pending_wal FROM check_caught_up) IS NULL
    OR confirmed_flush_lsn > (SELECT last_pending_wal FROM check_caught_up)
  )
END as caught_up,
invalidation_reason IS NOT NULL as invalid
FROM pg_catalog.pg_replication_slots
WHERE slot_type =3D 'logical'
  AND database =3D current_database()
  AND temporary IS FALSE;

thanks
Shveta