Re: Fix for gistchoose

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Thu, Aug 30, 2012 at 4:15 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> Yeah, the idea of replacing sum_grow with a boolean just occurred to me
>> too. As is, I think the code is making some less-than-portable
>> assumptions about what will happen if sum_grow overflows; which can
>> definitely happen, seeing that gistpenalty and its callees intentionally
>> return infinity in some cases. I'd rather it didn't attempt to add
>> column penalties together, and I think there's a case to be made that
>> not doing so is a back-patchable bug fix.

> Keep in mind that the worst case outcome is the index quality is worse
> than it otherwise would have been, so it's not like
> OMG-PostgreSQ-eats-your-data.

Agreed, but we've seen plenty of complaining about bloated gist indexes,
and this might be the cause.

>> I'll take a shot at improving this some more.

> Okey dokey.

Attached is a revised version of gistchoose(); I've not yet transposed
the changes into gistRelocateBuildBuffersOnSplit(). It looks a lot
more readable to me now. Objections?

regards, tom lane