Re: Re: Abbreviated keys for Datum tuplesort

On 25.1.2015 12:15, Andrew Gierth wrote:
>
> So given some suitable test data, such as
>
> create table stuff as select random()::text as randtext
> from generate_series(1,1000000); -- or however many rows
>
> you can do
>
> select percentile_disc(0) within group (order by randtext) from stuff;
>
> or
>
> select count(distinct randtext) from stuff;
>
> The performance improvements I saw were pretty much exactly as
> expected from the improvement in the ORDER BY and CREATE INDEX cases.

I've spent a fair amount of testing this today, and when using the
simple percentile_disc example mentioned above, I see this pattern:

master patched speedup
---------------------------------------------------------
generate_series(1,1000000) 4.2 0.7 6
generate_series(1,2000000) 9.2 9.8 0.93
generate_series(1,3000000) 14.5 15.3 0.95

so for a small dataset the speedup is very nice, but for larger sets
there's ~5% slowdown. Is this expected?

--
Tomas Vondra http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services