Hello
I think the old cluster issue still exists eveen with v22/23, as
pointed out by claude:
Set up a PG15 subscriber with an enabled subscription, with its table
in state i at clean shutdown, 50 extra databases (to make the check
phase longer), and pre_auth_delay=1 on the publisher to simulate a
slow connection.
And then pg_upgrade from v23:
Checking subscription state ok
...
-- pg_upgrade_dump_globals.sql:
SELECT pg_catalog.binary_upgrade_create_replication_origin('2'::pg_catalog.oid,
'pg_16391_16384'::pg_catalog.text, '0/40466F8'::pg_catalog.pg_lsn);
-- new cluster, after DROP SUBSCRIPTION s:
2 | pg_16391_16384
With the default timing the opposite happens: the tablesync starts
before the check, and pg_upgrade fails with 'The table sync state "d"
is not allowed', even though the table was in state i when the cluster
was stopped. The TAP test doesn't see either case because it sets
max_logical_replication_workers = 0 on the old subscriber.