Thread: Random subscription 021_twophase test failure on kestrel

Hi,

The 021_twophase test has failed on Kestrel at [1] with the following error:
#   Failed test 'should be no prepared transactions on subscriber'
#   at /home/bf/bf-build/kestrel/HEAD/pgsql/src/test/subscription/t/021_twophase.pl
line 438.
#          got: '1'
#     expected: '0'
# Looks like you failed 1 test of 30.

This failure is caused by a prepared transaction that was not properly
committed due to replication lag on one of the subscriptions. The test
involves two subscriptions: tap_sub and tap_sub_copy. After committing
the prepared transaction 'mygid', the test only waits for tap_sub_copy
to catch up:
node_publisher->wait_for_catchup($appname_copy);

However, tap_sub is dropped before ensuring it has replayed the commit
of 'mygid' prepared transaction, leading to a leftover prepared
transaction on the subscriber:
$node_subscriber->safe_psql('postgres', "DROP SUBSCRIPTION tap_sub");

When the test later checks for the number of prepared transactions, it
fails because tap_sub had not finished applying the commit:
#   at line 438
#          got: '1'
#     expected: '0'

This issue can be consistently reproduced by injecting a delay (e.g.,
3 seconds) in tap_sub's walsender while decoding the commit of
'mygid'. A patch to demonstrate this behavior is provided at
021_two_phase_test_failure_reproduce.patch. The test can be fixed by
explicitly waiting for both subscriptions to catch up before dropping
either. A patch implementing this fix is attached.

Thanks Amit for the offline discussion and sharing your thoughts on the same.

[1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kestrel&dt=2025-05-22%2021%3A19%3A22

Regards,
Vignesh

On Sat, May 24, 2025 at 6:07 AM Michael Paquier <michael@paquier.xyz> wrote:
>
>
> Yes, agreed that your suggested fix looks sensible with an extra check
> for pg_prepared_xacts on the subscriber side that can be useful for
> debugging.
>

+1.

--
With Regards,
Amit Kapila.