Columnar: vacuum freeze does not advance relfrozenxid

[jeff-davis]

With postgresql.conf:

wal_level=logical
shared_preload_libraries='citus'
autovacuum_freeze_max_age = 100000
vacuum_freeze_min_age = 50000
vacuum_freeze_table_age = 75000
synchronous_commit = off
fsync = off

And doing:

CREATE EXTENSION citus;
CREATE TABLE test_row(i int);
INSERT INTO test_row VALUES (1);
CREATE TABLE test_col(i int) USING columnar;
INSERT INTO test_col VALUES (1);

Then doing:

(for i in `seq 1 200000`; do echo "UPDATE test_row SET i = i + 1;"; done) | ./bin/psql -q -X postgres

I see:

=# select datname, relname, age(datfrozenxid) datage, age(relfrozenxid) as relage from pg_database, pg_class where datname='postgres' and (relname='test_row' or relname='test_col');
 datname  | relname  | datage | relage 
----------+----------+--------+--------
 postgres | test_row | 200018 | 200010
 postgres | test_col | 200018 | 200008
(2 rows)

After I wait for autovacuum to kick in, I see:

=# select datname, relname, age(datfrozenxid) datage, age(relfrozenxid) as relage from pg_database, pg_class where datname='postgres' and (relname='test_row' or relname='test_col');
 datname  | relname  | datage | relage 
----------+----------+--------+--------
 postgres | test_row | 200010 |  50002
 postgres | test_col | 200010 | 200010
(2 rows)

And further manual VACUUM or VACUUM FREEZE do not fix it.

[jeff-davis]

This is a fairly serious issue. I am sure we tested this before, but it's hard to create acceptance tests for this, so we must have seen a regression. Perhaps we can add a TAP test.

[amalek215]

I am also experiencing this issue
Postgres 14.1
citus 10.2-4

I've had to increase autovacuum_freeze_max_age to prevent autovacuum: VACUUM .... (to prevent wraparound) from constantly running.
I have 10s of thousands of tables using Citus columnar store and now that they are aging (created 5 months ago) those autovacuums began to always run.