Re: pg primary key bug?

Tom Lane wrote:

>pginfo <pginfo(at)t1(dot)unisoftbg(dot)com> writes:
>
>
>>You idea was that we have "vacuum full" + update or select for update in
>>the same time.
>>I think it is not the case, because we start vacuum full at 1:00 AM and
>>no one is working in this time.
>>
>>
>
>Hmm. AFAICT the duplicate row copies could only be produced by vacuum
>full, so that's certainly part of the issue. But if vacuum full in
>isolation were broken, we'd surely know it; so there must be some other
>contributing factor involved that your setup is exercising but other
>people are (mostly) not doing.
>
>I agree with the plan to use plain vacuum for awhile and see if that
>makes the problem go away. I think it would have to, but maybe I'm
>all wet about that.
>
>
Ok, we can still using vacuum full on some installs (with risk to make
problems to customes).
I will to ask if it will be possible to start some querys (I do not know
the query) exactly before running vacuum full and to save the results in
some log file. If it is possible, we will be able to post the results to
the list in case of ne problem and to have some start point for
reproducing the problem. My idea is some one more familiar with pg to
send this querys (if it exists) and we will install it in vacuum scripts.

>In the meantime I would suggest seeing if you can distill your
>application down into a test case that other people can run to reproduce
>the problem. It doesn't matter if the test doesn't make the bug happen
>very often, but we have to see the problem happening before we have much
>hope of fixing it.
>
>
>
>>Will vacuum full generate this problem if we have locked table in this
>>time? (It is possible to have locked table in theory)
>>
>>
>
>No, that's hardly likely. vacuum full deals with locks all the time.
>
>
>
>>Can you describe more detailed the idea of problem with "vacuum full" +
>>"update" and can some one make patch if this problem exists in theory
>>(if I understand you right)?
>>
>>
>
>I have no idea what the actual failure mechanism might be.
>
> regards, tom lane
>
>
regards,
ivan.

>
>
>