Re: BUG #15594: Unstable tests in contrib/test_decoding/output_iso/

On 16/01/2019 20:09, Andres Freund wrote:
> On 2019-01-16 14:56:34 +0000, PG Bug reporting form wrote:
>> The following bug has been logged on the website:
>>
>> Bug reference: 15594
>> Logged by: Alexander Lakhin
>> Email address: exclusion(at)gmail(dot)com
>> PostgreSQL version: 11.1
>> Operating system: Ubuntu 18.04
>> Description:
>>
>> Some tests in contrib/test_decoding/output_iso/ occasionally fail.
>> To reproduce it reliable you can do:
>> # sed 's/ mxact / $(shell printf "mxact %.0s" `seq 1000` ) /' -i
>> contrib/test_decoding/Makefile
>> # make check -C contrib/test_decoding/
>>
>> ...
>> test delayed_startup ... ok
>> test ondisk_startup ... ok
>> test concurrent_ddl_dml ... ok
>> test oldest_xmin ... ok
>> test snapshot_transfer ... ok
>> ============== shutting down postmaster ==============
>>
>> ==========================
>> 19 of 1005 tests failed.
>> ==========================
>
>> The differences that caused some tests to fail can be viewed in the
>> file ".../contrib/test_decoding/output_iso/regression.diffs". A copy of the
>> test summary that you see
>> above is saved in the file
>> ".../contrib/test_decoding/output_iso/regression.out".
>>
>> The diff file contains:
>> --- .../contrib/test_decoding/expected/delayed_startup.out 2018-02-08
>> 16:57:42.816744192 +0300
>> +++ .../contrib/test_decoding/output_iso/results/delayed_startup.out
>> 2019-01-16 17:46:33.138962021 +0300
>> @@ -12,6 +12,8 @@
>> step s2start: SELECT data FROM
>> pg_logical_slot_get_changes('isolation_slot', NULL, NULL, 'include-xids',
>> 'false');
>> data
>>
>> +BEGIN
>> +COMMIT
>> step s1b: BEGIN ISOLATION LEVEL SERIALIZABLE;
>> step s1w: INSERT INTO do_write DEFAULT VALUES;
>> step s1c: COMMIT;
>>
>> ======================================================================
>> ...
>>
>>
>> I observe this issue with the tests: mxact, delayed_startup, and
>> ondisk_startup.
>
> Interesting - I've not seen that on the buildfarm before, or in local tests
> with a lot of repetitions. I assume the problem vanishes if you add 'skip-empty-xacts', '1'
> to the options of pg_logical_slot_get_changes()?
>

I bet those empty transactions are from autovacuum. The probability of
hitting this could be exaggerated by the fact that the test is run
within the same instance 1000x rather than running the make check 1000x
so we produce plenty of catalog bloat for autovacuum to clean.

--
Petr Jelinek http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services