BUG #19430: Autovacuums stopped working possible due to problem with vacuuming shared catalog pg_authid

On Thu, 2026-03-12 at 09:24 +0000, PG Bug reporting form wrote:

PostgreSQL version: 14.15
Operating system: Ubuntu 22.04

On a PostgreSQL 14.15 server with about 100 databases running on Ubuntu
22.04 I observed following problem:

1) According to the DB logs (log_autovacuum_min_duration = 1s) and
monitoring graphs autovacuums almost completely stopped working across all
DBs at once
2) This is also confirmed by checking pg_stat_user_tables.last_autovacuum in
all databases, there were no autovacuums anywhere for several days
3) Counters in pg_stat_user_tables updated normally, no anomalies there
4) pg_stat_all_tables.last_autovacuum shows one anomaly with pg_authid
catalog: only that catalog was autovacuumed recently (there are not many
dead rows there, looks normal). There might be other anomalies in other
databases as I only checked one but I doubt it.
5) I found nothing interesting in the logs around the time when problem
started (I checked for ERROR/WARNING/FATAL messages and possible DDL queries
but sadly log_statement = none there, log_min_duration_statement = 1s)
6) Autovacuum launcher process appears to be running normally, periodically
wakes up and consume 1-3% CPU of single core, pg_stat_activity shows the
wait event as AutoVacuumMain
7) Just in case, I checked dmesg entries and the time offset graph in
monitoring (to verify that time wasn't moving backwards for some reason) -
no anomalies there
8) Running vacuum on one of the bloated tables (which cause performance
problems) worked without issues
9) There is no long running transactions or lagging standbys with
hot_standby_feedback on
10) pg_authid catalog has `age(relfrozenxid)` around 290M but
autovacuum_freeze_max_age is 200M and anti-wraparound autovacuum should've
been triggered
11) After running `vacuum verbose pg_authid;` nothing changed (it wasn't an
aggressive vacuum) but running `vacuum (freeze, verbose) pg_authid;` fixed
problem and autovacuums started working properly. Output of these commands
are at the end of the report.

I assume that problem might have started when age(relfrozenxid) for
pg_authid approached autovacuum_freeze_max_age. A rough estimate based on
the current rate of age(relfrozenxid) confirms that we might indeed did 90M
transactions during the autovacuum malfunction.

An interesting observation: the query `ALTER ROLE xxx WITH ;` from some
permission-granting business logic triggered several autovacuums on ~half of
databases at same second for some time but didn't fix the problem
completely. Perhaps some cache invalidation did that?

Unfortunately, 14.15 is more than 1 year old. An update to 14.22 was
discussed recently but haven't been made yet. There were several fixes
related to vacuum/autovacuum but they didn't seem directly related to the
issue but I might be wrong.

Output of vacuum commands:

vacuum verbose pg_authid;
INFO: vacuuming "pg_catalog.pg_authid"
INFO: table "pg_authid": found 0 removable, 7 nonremovable row versions in
1 out of 57 pages
DETAIL: 0 dead row versions cannot be removed yet, oldest xmin: 3828262030
Skipped 0 pages due to buffer pins, 55 frozen pages.
CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s.
INFO: vacuuming "pg_toast.pg_toast_1260"
INFO: table "pg_toast_1260": found 0 removable, 0 nonremovable row versions
in 0 out of 0 pages
DETAIL: 0 dead row versions cannot be removed yet, oldest xmin: 3828262030
Skipped 0 pages due to buffer pins, 0 frozen pages.
CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s.
VACUUM

vacuum (freeze, verbose) pg_authid;
INFO: aggressively vacuuming "pg_catalog.pg_authid"
INFO: table "pg_authid": found 0 removable, 196 nonremovable row versions
in 4 out of 57 pages
DETAIL: 0 dead row versions cannot be removed yet, oldest xmin: 3828278683
Skipped 0 pages due to buffer pins, 53 frozen pages.
CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s.
INFO: aggressively vacuuming "pg_toast.pg_toast_1260"
INFO: table "pg_toast_1260": found 0 removable, 0 nonremovable row versions
in 0 out of 0 pages
DETAIL: 0 dead row versions cannot be removed yet, oldest xmin: 3828278683
Skipped 0 pages due to buffer pins, 0 frozen pages.
CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s.
VACUUM

This change in 14.21 looks relevant:

- Fix issues around in-place catalog updates (Noah Misch)

Send a nontransactional invalidation message for an in-place update, since
such an update will survive transaction rollback. Also ensure that the
update is WAL-logged before other sessions can see it. These fixes
primarily prevent scenarios in which relations' frozen-XID attributes
become inconsistent, possibly allowing premature CLOG truncation and
subsequent “could not access status of transaction” errors.

Yours,
Laurenz Albe