未使用Postgres的gin_trgm_ops索引

时间:2019-06-06 21:43:09

标签: postgresql performance database-performance query-performance pg-trgm

我正在尝试使用pg_trgm扩展名speed up在Postgres中进行一些文本匹配:

CREATE TABLE test3 (id bigint, key text, value text);

insert into test3 values (1, 'first 1', 'second 3');
insert into test3 values (2, 'first 1', 'second 2');
insert into test3 values (2, 'first 2', 'second 3');
insert into test3 values (3, 'first 1', 'second 2');
insert into test3 values (3, 'first 1', 'second 3');
insert into test3 values (4, 'first 2', 'second 3');
insert into test3 values (4, 'first 2', 'second 3');
insert into test3 values (4, 'first 1', 'second 2');
insert into test3 values (4, 'first 1', 'second 2');

-- repeat the above 1,000,000x times, to have more rows for benchmarking
insert into test3(id, key, value) select id, key, value from test3 cross join generate_series(1, 1000000);

现在我用ILIKE查询此表:

select count(*) from test3 where key = 'first 1' and value ilike '%nd 3%';
Time: 918.265 ms

要查看索引编制是否可以加快速度,我在pg_trgmkey列上都添加了value

CREATE extension if not exists pg_trgm;
CREATE INDEX test3_key_trgm_idx ON test3 USING gin (key gin_trgm_ops);
CREATE INDEX test3_value_trgm_idx ON test3 USING gin (value gin_trgm_ops);

但是查询仍然需要相同的时间,EXPLAIN ANALYZE显示索引根本没有被使用:

explain analyze select count(*) from test3 where key = 'first 1' and value ilike '%nd 3%';
                                                                 QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------
 Finalize Aggregate  (cost=126905.14..126905.15 rows=1 width=8) (actual time=1017.666..1017.667 rows=1 loops=1)
   ->  Gather  (cost=126904.93..126905.14 rows=2 width=8) (actual time=1017.505..1018.778 rows=3 loops=1)
         Workers Planned: 2
         Workers Launched: 2
         ->  Partial Aggregate  (cost=125904.93..125904.94 rows=1 width=8) (actual time=1010.862..1010.862 rows=1 loops=3)
               ->  Parallel Seq Scan on test3  (cost=0.00..122427.06 rows=1391148 width=0) (actual time=0.041..973.550 rows=666667 loops=3)
                     Filter: ((value ~~* '%nd 3%'::text) AND (key = 'first 1'::text))
                     Rows Removed by Filter: 2333336
 Planning Time: 0.266 ms
 Execution Time: 1018.814 ms

Time: 1049.413 ms (00:01.049)

请注意顺序扫描。有什么作用?

1 个答案:

答案 0 :(得分:3)

没关系,我发现了问题。

查询计划者比我的玩具测试集更聪明;看到大多数行与查询匹配,就进行了顺序扫描。

如果我尝试使用ilike '%nd 0%',则没有行匹配并且EXPLAIN ANALYZE报告Bitmap Index Scan on test3_value_trgm_idx正确。

因此,以这种方式规范化原始JSONB是可行的。但是,我还将尝试寻找和比较另一种方法,即在TEXT上使用正则表达式,以避免不得不创建和维护另一个表。