Re: GIN improvements part2: fast scan

Hi Oleg,

On 3 Únor 2014, 7:53, Oleg Bartunov wrote:
> Tomasa, it'd be nice if you use real data in your testing.

I'm using a mailing list archive (the benchmark is essentially a simple
search engine on top of the archive, implemented using built-in
full-text). So I think this is a quite 'real' dataset, not something
synthetic.

The queries are generated, of course, but I strived to make them as real
as possible.

Sure, this is not a hstore-centric benchmark, but the thread is about GIN
indexes, so I think it's fair.

> One very good application of gin fast-scan is dramatic performance
> improvement  of hstore/jsonb @> operator, see slides 57, 58
> http://www.sai.msu.su/~megera/postgres/talks/hstore-dublin-2013.pdf.
> I'd like not to lost this benefit :)

Yes, that's something we could/should test, probably. Sadly I don't have a
dataset or a complete real-world test case at hand. Any ideas?

I certainly agree that it'd be very sad to lose the performance gain for
hstore/json. OTOH my fear is that to achieve that gain, we'll noticeably
slow down other important use cases (e.g. full-text search), which is one
of the reasons why I was doing the tests.

regards
Tomas