Postgresql

在 f_unaccent() 上使用 GIN 索引的查詢似乎很慢?

  • April 10, 2018

我在一個包含產品的表中有近 20M 行,包括他們的名字。

我希望通過名稱快速搜尋全文,所以我創建了這個索引:

CREATE INDEX uprice_item_occurrence_unaccent_name_trgm_idx ON price_item_occurrence USING gin (f_unaccent(name) gin_trgm_ops);;

我正在跳躍以下查詢將花費更少的時間(比如說)500ms

select * from price_item_occurrence as oo
where f_unaccent(oo.name) % f_unaccent('iphone');

但這需要將近2s:

postgres=# explain analyze select * from price_item_occurrence as oo where f_unaccent(oo.name) % f_unaccent('iphone');
                                                                            QUERY PLAN                                                                              
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on price_item_occurrence oo  (cost=1956.42..63674.14 rows=16570 width=287) (actual time=247.918..1880.759 rows=94 loops=1)
  Recheck Cond: (f_unaccent((name)::text) % 'iphone'::text)
  Rows Removed by Index Recheck: 87838
  Heap Blocks: exact=76663
  ->  Bitmap Index Scan on uprice_item_occurrence_unaccent_name_trgm_idx  (cost=0.00..1952.28 rows=16570 width=0) (actual time=195.418..195.418 rows=88962 loops=1)
        Index Cond: (f_unaccent((name)::text) % 'iphone'::text)
Planning time: 0.444 ms
Execution time: 1880.833 ms

數據庫可能很忙,但我不確定。

我試著玩select set_limit(0.9);(增加),它有一點幫助,但沒有多大幫助。

我正在使用 Postgres 10,可以更改 Postgres 配置,我願意接受建議。

我嘗試了它ilike並改進了一些:

postgres=# explain analyze select * from price_item_occurrence as oo where f_unaccent(oo.name) ilike ('%' || f_unaccent('iphone') || '%');
                                                                            QUERY PLAN                                                                             
--------------------------------------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on price_item_occurrence oo  (cost=3135.08..416823.45 rows=166075 width=286) (actual time=50.258..670.085 rows=65917 loops=1)
  Recheck Cond: (f_unaccent((name)::text) ~~* '%iphone%'::text)
  Rows Removed by Index Recheck: 10
  Heap Blocks: exact=59750
  ->  Bitmap Index Scan on uprice_item_occurrence_unaccent_name_trgm_idx  (cost=0.00..3093.56 rows=166075 width=0) (actual time=37.385..37.385 rows=67700 loops=1)
        Index Cond: (f_unaccent((name)::text) ~~* '%iphone%'::text)
Planning time: 0.545 ms
Execution time: 675.776 ms
(8 rows)

大約快 2 倍。

我試過limit 10

postgres=# explain analyze select * from price_item_occurrence as oo where f_unaccent(oo.name) % f_unaccent('iphone') limit 10;
                                                                               QUERY PLAN                                                                                
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit  (cost=373.27..410.51 rows=10 width=287) (actual time=268.718..589.131 rows=10 loops=1)
  ->  Bitmap Heap Scan on price_item_occurrence oo  (cost=373.27..62493.45 rows=16680 width=287) (actual time=268.715..589.123 rows=10 loops=1)
        Recheck Cond: (f_unaccent((name)::text) % 'iphone'::text)
        Rows Removed by Index Recheck: 18917
        Heap Blocks: exact=17100
        ->  Bitmap Index Scan on uprice_item_occurrence_unaccent_name_trgm_idx  (cost=0.00..369.10 rows=16680 width=0) (actual time=165.958..165.958 rows=69268 loops=1)
              Index Cond: (f_unaccent((name)::text) % 'iphone'::text)
Planning time: 0.397 ms
Execution time: 589.187 ms
(9 rows)

這也更快,也許幾乎足夠好

我已經看到GIN索引對於這些查詢的執行速度通常比 GiST 快得多。試試這個索引:

CREATE INDEX price_item_occurrence_name_trgm_gin idx ON price_item_occurrence
USING GIN (f_unaccent(name) gin_trgm_ops);

有關的:

所有關於性能優化的基本建議都適用。對於初學者來說,您的表需要足夠快的VACUUMed 和d。ANALYZE

引用自:https://dba.stackexchange.com/questions/203429