Re: BUG #16329: Valgrind detects an invalid read when building a gist index with buffering

Andrey Borodin <x4mmm(at)yandex-team(dot)ru> writes:
>> 13 окт. 2020 г., в 03:16, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> написал(а):
>> 2. The test exposed the gistRelocateBuildBuffersOnSplit bug only about
>> one time in ten for me. I suppose this is due to the random split
>> choices gistchoose makes for equally-good tuples, so it's not terribly
>> surprising; but it is problematic for a test that we're hoping to use
>> to provide reliable code coverage.
>>
>> I'm not really sure what we could do about #2. Perhaps, instead of
>> relying on random(), we could make gistchoose() use our own PRNG and
>> then invent a debugging function that forces the seed to a known value.
>> (GEQO already does something similar, although I'd not go as far as
>> exposing the seed as a GUC. Resetting it via some quick-hack C
>> function in regress.c would be enough IMO.) Or perhaps gist.sql could
>> be adjusted so that the test data is less full of equally-good tuples.

> I think we should use not entropy-based tie breaker in GiST. We can extract some randomness from tuples using hash.
> I'd be much happier if GiST behaviour was deterministic.

If we started using our own PRNG, we could very easily make the "random"
sequence be the same in every GiST build --- again, see GEQO for prior
art. I'm a little suspicious of trying to pull some entropy out of the
tuples themselves: that seems like it'd be tremendously prone to cross-
platform and cross-version behavioral differences.

regards, tom lane