Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss

[GENERAL] Multicolumn index including tsvector.

1 view
Skip to first unread message

Kris Gale

unread,
Nov 23, 2009, 6:08:17 PM11/23/09
to
Hi everyone,

In the documentation and past messages to this list, it sounds like the btree_gist and btree_gin modules included in contrib/ in 8.4 should give me the functionality I'm looking for, but I don't seem to be getting the behavior I want.

I've made an example table representing something like a simplified version of a web discussion board.

     Table "public.example"
 Column  |   Type   | Modifiers
---------+----------+-----------
 body    | text     |
 vectors | tsvector |
 user_id | bigint   |

I've got btree_gin and btree_gist installed, so I can make a composite index on vectors and user_id (which is a bigint).

create index index_examples_gin on example using gist (user_id,vectors);
create index index_examples_gist on example using gist (user_id,vectors);

So what I'm expecting here is that it'll be able to use one of those composite indexes to satisfy both the user_id and the vectors constraints.  That doesn't seem to be the case, based on this query plan:

explain analyze select body from example where user_id=1 and vectors @@ to_tsquery('simple', 'when') limit 10;
                                                                   QUERY PLAN                                                                   
-------------------------------------------------------------------------------------------------------------------------------------------------
 Limit  (cost=1786.87..10510.47 rows=10 width=5) (actual time=203.155..244.987 rows=10 loops=1)
   ->  Bitmap Heap Scan on example  (cost=1786.87..38425.99 rows=42 width=5) (actual time=203.153..244.980 rows=10 loops=1)
         Recheck Cond: (vectors @@ '''when'''::tsquery)
         Filter: (user_id = 1)
         ->  Bitmap Index Scan on index_examples_gist  (cost=0.00..1786.86 rows=26535 width=0) (actual time=186.711..186.711 rows=27477 loops=1)
               Index Cond: (vectors @@ '''when'''::tsquery)
 Total runtime: 245.062 ms
(7 rows)

So it seems to be using the index only to satisfy the tsquery part of the where clause, and then applying the user_id filter to the rows it fetches.  Ideally, I'd want to see it using both columns as part of the index condition, and not using a filter at all.  Is what I'm trying to achieve possible?

Thanks in advance.

Kris

Tom Lane

unread,
Nov 23, 2009, 7:12:35 PM11/23/09
to
Kris Gale <kris...@gmail.com> writes:
> So what I'm expecting here is that it'll be able to use one of those
> composite indexes to satisfy both the user_id and the vectors constraints.
> That doesn't seem to be the case, based on this query plan:

> explain analyze select body from example where user_id=1 and vectors @@
> to_tsquery('simple', 'when') limit 10;

Try coercing the '1' to a bigint. I don't believe the btree_gist
opclasses have any support for cross-type operators.

regards, tom lane

--
Sent via pgsql-general mailing list (pgsql-...@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-general

Kris Gale

unread,
Nov 23, 2009, 7:15:40 PM11/23/09
to
> explain analyze select body from example where user_id=1 and vectors @@
> to_tsquery('simple', 'when') limit 10;

Try coercing the '1' to a bigint.  I don't believe the btree_gist
opclasses have any support for cross-type operators.

Perfect!  Thanks, Tom.  The query plan now shows it considering both columns in the index condition.

Kris

Alban Hertroys

unread,
Nov 24, 2009, 6:28:43 AM11/24/09
to
On 24 Nov 2009, at 24:08, Kris Gale wrote:

> Table "public.example"
> Column | Type | Modifiers
> ---------+----------+-----------
> body | text |
> vectors | tsvector |
> user_id | bigint |
>
> I've got btree_gin and btree_gist installed, so I can make a composite index on vectors and user_id (which is a bigint).


A bigint for userid? How many billions of users do you expect to get?

A bigint is 8 bytes or 64 bits. I think a normal int (32-bits) would be quite sufficient; it allows up to 2 billion users. It uses less storage space for both the data and the indices too.

And as Tom already said casting can be a problem with bigints, not only if certain operators aren't defined for comparison between int and bigint, but you'll also see a performance hit if table data you compare to needs to be upcasted to a bigint; joins come to mind.

Alban Hertroys

--
If you can't see the forest for the trees,
cut the trees and you'll see there is no forest.


!DSPAM:737,4b0bc36e11738279827033!

0 new messages