On 3/11/22 12:34, Amit Kapila wrote:
> On Tue, Mar 8, 2022 at 11:59 PM Tomas Vondra
> <tomas.vondra@enterprisedb.com> wrote:
>>
>> On 3/7/22 22:25, Tomas Vondra wrote:
>>>>
>>>> Interesting. I can think of one reason that might cause this - we log
>>>> the first sequence increment after a checkpoint. So if a checkpoint
>>>> happens in an unfortunate place, there'll be an extra WAL record. On
>>>> slow / busy machines that's quite possible, I guess.
>>>>
>>>
>>> I've tweaked the checkpoint_interval to make checkpoints more aggressive
>>> (set it to 1s), and it seems my hunch was correct - it produces failures
>>> exactly like this one. The best fix probably is to just disable decoding
>>> of sequences in those tests that are not aimed at testing sequence decoding.
>>>
>>
>> I've pushed a fix for this, adding "include-sequences=0" to a couple
>> test_decoding tests, which were failing with concurrent checkpoints.
>>
>> Unfortunately, I realized we have a similar issue in the "sequences"
>> tests too :-( Imagine you do a series of sequence increments, e.g.
>>
>> SELECT nextval('s') FROM generate_sequences(1,100);
>>
>> If there's a concurrent checkpoint, this may add an extra WAL record,
>> affecting the decoded output (and also the data stored in the sequence
>> relation itself). Not sure what to do about this ...
>>
>
> I am also not sure what to do for it but maybe if in some way we can
> increase checkpoint timeout or other parameters for these tests then
> it would reduce the chances of such failures. The other idea could be
> to perform checkpoint before the start of tests to reduce the
> possibility of another checkpoint.
>
Yeah, I had the same ideas, but I'm not sure I like any of them. I doubt
we want to make checkpoints extremely rare, and even if we do that it'll
still fail on slow machines (e.g. with valgrind, clobber cache etc.).
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company