Обсуждение: GiST or GIN, I feel like I am doing something wrong

Поиск
Список
Период
Сортировка

GiST or GIN, I feel like I am doing something wrong

От
Aaron
Дата:
We are testing full text searching on a small chunk of our data. We
have created an INDEX to make searching faster.  From the PostgreSQL
8.3 docs, we are running 8.3.7, it seems we should be running GIN
indexes.  The reason GIN on paper seems like the right INDEX:
* we have static data
* we have over 241071 unique words (lexemes)
* GIN index lookups are about three times faster and we are 99.9% searching

The problem is that we have been testing with both INDEX types and
GiST is killing GIN.  I believe it has to do with the size of our GiST
index.
SELECT * from relation_size where relation like '%full%';
                  relation                  |  size
--------------------------------------------+--------
 public.profile_images_fulltext_gin         | 437 MB
 public.profile_images_fulltext             | 161 MB
 public.profile_images_fulltext_gist        | 66 MB
 public.profile_images_fulltext_pif_key_key | 18 MB
(4 rows)

So my questions...
Why is the GiST index so large?
Would the large size likely effect performance?
Am I doing something fundamentally wrong?
Yes I was sure to ANALYZE public.profile_images_fulltext between all
my INDEX DROP and CREATE

More details:
owl=# \d profile_images_fulltext
             Table "public.profile_images_fulltext"
      Column      |            Type             |   Modifiers
------------------+-----------------------------+---------------
 pif_key          | bigint                      | not null
 content          | tsvector                    |
 datetime_created | timestamp without time zone | default now()
 raw              | text                        |

owl=# SELECT count (pif_key) from public.profile_images_fulltext;
 count
--------
 630699
(1 row)

owl=# SELECT count(word) FROM ts_stat('SELECT content FROM
profile_images_fulltext');
 count
--------
 241071
(1 row)

CREATE INDEX profile_images_fulltext_gin ON profile_images_fulltext
USING gin(content);
CREATE INDEX profile_images_fulltext_gist ON profile_images_fulltext
USING gist(content);



Any and all thoughts would be greatly appreciated,
Aaron Thul
http://www.chasingnuts.com
Life is complex: it has real and imaginary components.

Re: GiST or GIN, I feel like I am doing something wrong

От
Tom Lane
Дата:
Aaron <aaron@chasingnuts.com> writes:
> CREATE INDEX profile_images_fulltext_gin ON profile_images_fulltext
> USING gin(content);
> CREATE INDEX profile_images_fulltext_gist ON profile_images_fulltext
> USING gist(content);

What did you have maintenance_work_mem set to while you did this?
GIST doesn't care, but GIN likes to have lots of workspace while
building an index.  I'm not entirely sure if small workspace only
affects build time, or if it could result in a bloated/inefficient
index ... but if the latter is true it might explain your results.

            regards, tom lane

Re: GiST or GIN, I feel like I am doing something wrong

От
Aaron
Дата:
Tom,

Our maintenance_work_mem is 1024MB so there should have been plenty of
memory for INDEX creation.  I happened to be watching top when we
created the GiN INDEX and the process used about 500MB of non-shared
memory.

Aaron Thul
http://www.chasingnuts.com



On Wed, Jun 17, 2009 at 11:30 AM, Tom Lane<tgl@sss.pgh.pa.us> wrote:
> Aaron <aaron@chasingnuts.com> writes:
>> CREATE INDEX profile_images_fulltext_gin ON profile_images_fulltext
>> USING gin(content);
>> CREATE INDEX profile_images_fulltext_gist ON profile_images_fulltext
>> USING gist(content);
>
> What did you have maintenance_work_mem set to while you did this?
> GIST doesn't care, but GIN likes to have lots of workspace while
> building an index.  I'm not entirely sure if small workspace only
> affects build time, or if it could result in a bloated/inefficient
> index ... but if the latter is true it might explain your results.
>
>                        regards, tom lane
>

Re: GiST or GIN, I feel like I am doing something wrong

От
Oleg Bartunov
Дата:
Aaron,

did you actually check performance of search in both cases ?
GiST index can be small but very inefficient, since top-level
signatures can be degenerated, so we just remove them. It's easy to
see that looking in explain analyze - see difference between number of
rows found by index and actual number after recheck. And remember, recheck
needs access to the heap, which is slow and can kill performance.
GIN is big, but it should be fine for your setup. Also, see explain analyze.

Oleg
On Wed, 17 Jun 2009, Aaron wrote:

> We are testing full text searching on a small chunk of our data. We
> have created an INDEX to make searching faster.  From the PostgreSQL
> 8.3 docs, we are running 8.3.7, it seems we should be running GIN
> indexes.  The reason GIN on paper seems like the right INDEX:
> * we have static data
> * we have over 241071 unique words (lexemes)
> * GIN index lookups are about three times faster and we are 99.9% searching
>
> The problem is that we have been testing with both INDEX types and
> GiST is killing GIN.  I believe it has to do with the size of our GiST
> index.
> SELECT * from relation_size where relation like '%full%';
>                  relation                  |  size
> --------------------------------------------+--------
> public.profile_images_fulltext_gin         | 437 MB
> public.profile_images_fulltext             | 161 MB
> public.profile_images_fulltext_gist        | 66 MB
> public.profile_images_fulltext_pif_key_key | 18 MB
> (4 rows)
>
> So my questions...
> Why is the GiST index so large?
> Would the large size likely effect performance?
> Am I doing something fundamentally wrong?
> Yes I was sure to ANALYZE public.profile_images_fulltext between all
> my INDEX DROP and CREATE
>
> More details:
> owl=# \d profile_images_fulltext
>             Table "public.profile_images_fulltext"
>      Column      |            Type             |   Modifiers
> ------------------+-----------------------------+---------------
> pif_key          | bigint                      | not null
> content          | tsvector                    |
> datetime_created | timestamp without time zone | default now()
> raw              | text                        |
>
> owl=# SELECT count (pif_key) from public.profile_images_fulltext;
> count
> --------
> 630699
> (1 row)
>
> owl=# SELECT count(word) FROM ts_stat('SELECT content FROM
> profile_images_fulltext');
> count
> --------
> 241071
> (1 row)
>
> CREATE INDEX profile_images_fulltext_gin ON profile_images_fulltext
> USING gin(content);
> CREATE INDEX profile_images_fulltext_gist ON profile_images_fulltext
> USING gist(content);
>
>
>
> Any and all thoughts would be greatly appreciated,
> Aaron Thul
> http://www.chasingnuts.com
> Life is complex: it has real and imaginary components.
>
>

     Regards,
         Oleg
_____________________________________________________________
Oleg Bartunov, Research Scientist, Head of AstroNet (www.astronet.ru),
Sternberg Astronomical Institute, Moscow University, Russia
Internet: oleg@sai.msu.su, http://www.sai.msu.su/~megera/
phone: +007(495)939-16-83, +007(495)939-23-83