MIME-Version: 1.0
References: 
 <CA+UBfaktds57dw2M8BEv_kS-=ixph3w+3MxKixtaDQMi_k7Ybg@mail.gmail.com>
 <CAJpy0uDmBEHJAk7i5afYaqGA8V8j1XdfAT_=WYwZ=AhHZu6Z1A@mail.gmail.com>
In-Reply-To: 
 <CAJpy0uDmBEHJAk7i5afYaqGA8V8j1XdfAT_=WYwZ=AhHZu6Z1A@mail.gmail.com>
From: Masahiko Sawada <sawada.mshk@gmail.com>
Date: Thu, 21 May 2026 23:10:20 -0700
Message-ID: 
 <CAD21AoBSUi58=aL0uRPO57d8Qjrd_cakGeuguDEy+fPRKRLh2Q@mail.gmail.com>
Subject: Re: effective_wal_level is not decreasing after using REPACK
 (CONCURRENTLY)
To: shveta malik <shveta.malik@gmail.com>
Cc: Imran Zaheer <imran.zhir@gmail.com>,
 pgsql-hackers <pgsql-hackers@postgresql.org>,
	shveta malik <shvetamalik@gmail.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: 
 <https://www.postgresql.org/message-id/CAD21AoBSUi58%3DaL0uRPO57d8Qjrd_cakGeuguDEy%2BfPRKRLh2Q%40mail.gmail.com>
Precedence: bulk

On Thu, May 21, 2026 at 9:19=E2=80=AFPM shveta malik <shveta.malik@gmail.co=
m> wrote:
>
> On Thu, May 21, 2026 at 10:02=E2=80=AFPM Imran Zaheer <imran.zhir@gmail.c=
om> wrote:
> >
> > Hi
> >
> > The recent support for dynamic toggling of logical decoding (67c2097)
> > disables logical
> > decoding if no logical slots are present. But the repack command doesn'=
t seem to
> > coordinate with this toggling. The effective_wal_level is not
> > decreasing after using repack concurrently.
> >
> > postgres=3D# show effective_wal_level;
> >  effective_wal_level
> > ---------------------
> >  replica
> > (1 row)
> >
> > postgres=3D# create table foo(a int primary key);
> > CREATE TABLE
> > postgres=3D# REPACK (CONCURRENTLY) foo;
> > 2026-05-21 20:46:25.423 PKT [1591896] LOG:  logical decoding is
> > enabled upon creating a new logical replication slot
> > 2026-05-21 20:46:25.634 PKT [1591896] LOG:  logical decoding found
> > consistent point at 0/018F36D0
> > 2026-05-21 20:46:25.634 PKT [1591896] DETAIL:  There are no running
> > transactions.
> > REPACK
> > postgres=3D# select slot_name from pg_replication_slots;
> >  slot_name
> > -----------
> > (0 rows)
> >
> > postgres=3D# show effective_wal_level;
> >  effective_wal_level
> > ---------------------
> >  logical
> > (1 row)
> >
> >
> > The server has to be restarted in order to decrease the
> > effective_wal_level. REPACK CONCURRENTLY uses a temporary slot that is
> > dropped at the time of cleanup,  but logical decoding is not disabled.
> >
> > This may be related to both commits, 28d534e and 67c2097
> >
> > The attached patch adds the `RequestDisableLogicalDecoding` call to
> > `repack_cleanup_logical_decoding` after the replication slot is
> > dropped so the checkpointer will take care of it..
> >

Good catch!

>
> Thanks for reporting the issue. I agree with both the problem
> statement and the proposed fix.
>
> The fix LGTM. The only point I=E2=80=99d like to discuss is whether it wo=
uld
> make more sense for RequestDisableLogicalDecoding() to be called
> directly from ReplicationSlotDropAcquired().
>
> Currently, ReplicationSlotRelease(), ReplicationSlotDrop(), and now
> repack_cleanup_logical_decoding() all invoke
> RequestDisableLogicalDecoding() immediately after
> ReplicationSlotDropAcquired(). Given this pattern, it may be cleaner
> and less error-prone to make RequestDisableLogicalDecoding() part of
> ReplicationSlotDropAcquired() itself, which could also help avoid
> similar bugs in the future.
>
> That said, one concern is that ReplicationSlotsDropDBSlots() could end
> up issuing too many disable requests if there are many logical slots
> in the target database, so I=E2=80=99m not entirely sure whether this is =
the
> right direction. Thoughts?

Good point. I think we can have ReplicationSlotDropAcquired() have a
flag to skip sending a deactivation request. That way,
ReplicationSlotsDropDBSlots() can check the logical slot presence
after processing all slots and other callers can request the
deactivation after dropping the slot. It would help simplify the code
somewhat. It's conventional that when dropping a slot we acquire the
slot first and call RepicationSlotDropAcquired() to reliably drop a
slot (ReplicationSlotCleanup() is an exception). Therefore, I think
that having a flag to ReplicationSlotDropAcquired() could help future
developers to make sure to disable logical decoding at the slot drop.

Regards,

--=20
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com