Re: BUG #18692: Segmentation fault when extending a varchar column with a gist index with custom signal length

Alexander Korotkov <aekorotkov@gmail.com> writes:
> On Fri, Nov 8, 2024 at 4:05 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I think this is nonsense.  What about toasted datums, or even
>> just short-header ones?

> You are correct.  I quickly skim trough the sources and didn't find a
> function which compares detoasted contents of Datums excluding
> headers.

Right, there isn't one AFAIK.  For context, datumIsEqual() is mainly
meant to support cases such as comparison of Const nodes in equal(),
where

  a) false match is unacceptable, but false no-match will at worst
     cause some performance loss;
  b) the two inputs have probably come through similar paths (eg,
     both are fresh from the parser, or both were on-disk), so that
     any toasting effects will be the same.
Also
  c) for many of the callers, it'd be unsafe to perform disk
     access to allow accurate comparison of out-of-line values.

Problem (a) would be all right for this case, and (c) too,
but not so much (b).

You can certainly imagine writing a more aggressive routine
that can cope with detoasting, but I don't think we've yet
found a case where that seemed worth the trouble.  Partial
solutions such as "cope with short-header but not compressed
values" have even narrower possible use-cases.

> So, yes, comparison using C-collation seems the most
> reasonable option.

WFM.  But I'm concerned that you couldn't build a test case
where the comparison fails.  Seems like we need one, in view
of this bug having escaped detection for so long.

            regards, tom lane