On 2024-07-21 Su 4:08 PM, Alexander Lakhin wrote:
> Hello,
>
> 21.07.2024 20:34, Tom Lane wrote:
>> Andrew Dunstan <andrew@dunslane.net> writes:
>>> I noticed this when working on the PostgreSQL::Test::Session project I
>>> have in hand. All the tests pass except occasionally the xid_wraparound
>>> tests fail. It's not always the same test script that fails either. I
>>> tried everything but couldn't make the failure stop. So then I switched
>>> out my patch so it's running on plain master and set things running
>>> in a
>>> loop. Lo and behold it can be relied on to fail after only a few
>>> iterations.
>> I have been noticing xid_wraparound failures in the buildfarm too.
>> They seemed quite infrequent, but it wasn't till just now that
>> I realized that xid_wraparound is not run by default. (You have to
>> put "xid_wraparound" in PG_TEST_EXTRA to enable it.) AFAICS the
>> only buildfarm animals that have enabled it are dodo and perentie.
>> dodo is failing this test fairly often:
>>
>> https://buildfarm.postgresql.org/cgi-bin/show_history.pl?nm=dodo&br=HEAD
>
> I think this failure is counted at [1]. Please look at the linked message
> [2], where I described what makes the test fail.
>
> [1]
>
https://wiki.postgresql.org/wiki/Known_Buildfarm_Test_Failures#001_emergency_vacuum.pl_fails_to_wait_for_datfrozenxid_advancing
> [2]
> https://www.postgresql.org/message-id/5811175c-1a31-4869-032f-7af5e3e4506a@gmail.com
It's sad nothing has happened abut this for 2 months.
There's no point in having unreliable tests. What's not 100% clear to me
is whether this failure indicates a badly formulated test or the test is
correct and has identified an underlying bug.
Regarding the point in [2] about the test being run twice in buildfarm
clients, I think we should mark the module as NO_INSTALLCHECK in the
Makefile.
cheers
andrew
--
Andrew Dunstan
EDB: https://www.enterprisedb.com