Обсуждение: Index usage btree+gist ?

Hi List !

I have a table with a lot of rows (~3.000.000 I believe), and two
indexes.
The first one is a BTree index on a column (lets call it
btreecolumn) which contains only 8 different integer values (from
0 to 8).
The second one is a Gist index on a geometry column (gistcolumn)
in PostGIS format.
I run a query on this table that looks like :
SELECT gistcolumn FROM mytable
WHERE btreecolumn=0
AND (SELECT AGeometry FROM anothertable) && gistcolumn;

EXPLAIN on this query tells me :

Index Scan using gistcolumn_gist on table  (cost=13.52..188.20
rows=1 width=136)"
    Index Cond: ($0 && gistcolumn)"
    Filter: ((btreecolumn = 0) AND ($0 && gistcolumn))"
    InitPlan"
      ->  Aggregate  (cost=13.51..13.52 rows=1 width=32)"
            ->  Seq Scan on anothertable  (cost=0.00..13.50 rows=1
width=32)"
                  Filter: ((somecolumn)::text = 'value'::text)"


So if I understand this correctly, only the Gist index is used
here ? I thought that first using the Btree index to filter some
data, then the Gist index to refine the result would have been
more efficient ?

Am I correct, or am I misinterpreting the EXPLAIN result ?
If not, what is wrong with my index or my query ?

Thanks for your help !

Regards
--
Arnaud

Arnaud Lesauvage <thewild@freesurf.fr> writes:
> I have a table with a lot of rows (~3.000.000 I believe), and two
> indexes.
> The first one is a BTree index on a column (lets call it
> btreecolumn) which contains only 8 different integer values (from
> 0 to 8).

An index as unselective as that is almost useless.  It's not surprising
that the planner doesn't think it's worth the trouble to use it.

            regards, tom lane

Tom Lane a écrit :
> Arnaud Lesauvage <thewild@freesurf.fr> writes:
>> I have a table with a lot of rows (~3.000.000 I believe), and two
>> indexes.
>> The first one is a BTree index on a column (lets call it
>> btreecolumn) which contains only 8 different integer values (from
>> 0 to 8).
>
> An index as unselective as that is almost useless.  It's not surprising
> that the planner doesn't think it's worth the trouble to use it.

OK, thanks for pointing this out.
I was advised the PostGis list to use a multicolumn index on both
the integer column and the geometry column.
Another suggestion was to cluster the table on the gist-geometry
index.
What do you think about that ?

Regards
--
Arnaud

Arnaud Lesauvage <thewild@freesurf.fr> writes:
> I was advised the PostGis list to use a multicolumn index on both
> the integer column and the geometry column.

You could try that (put the geometry column first!).  I'm not sure how
effective additional columns in a gist index really are, but it's worth
experimenting with.

            regards, tom lane

Tom Lane a écrit :
> Arnaud Lesauvage <thewild@freesurf.fr> writes:
>> I was advised the PostGis list to use a multicolumn index on both
>> the integer column and the geometry column.
>
> You could try that (put the geometry column first!).  I'm not sure how
> effective additional columns in a gist index really are, but it's worth
> experimenting with.


OK, I'll try this first then !

Thanks Tom !

Regards
--
Arnaud