Digoal.zhou

Digoal.zhou

公益是一辈子的事, I'm digoal, Just do it.

PostgreSQL 一复合查询SQL优化例子 - (多个exists , 范围检索 , IN检索 , 模糊检索 组合)

4 minute read

背景

当一个SQL包含复杂的多个exists , 范围检索 , IN检索 , 模糊检索 , 组合查询时，可能由于索引使用不当导致查询性能较慢。

主要的问题在于，索引使用不当，可能导致几个问题：

1、索引本身扫描的耗时过多

2、位图扫描引入的recheck过多

3、subplan 引入的 filter过多

一个现实的例子，可以看到耗时集中在recheck和filter上面，每个索引扫描返回的记录数都很多，但是组合起来是0条符合条件的记录。

问题就出在索引不正确上，导致了问题。

->  Subquery Scan on "*SELECT* 2"  (cost=273453.65..432483146.70 rows=223 width=349) (actual time=25932.371..25932.371 rows=0 loops=1)  
      Output: ...................................  
      Buffers: shared hit=920071 read=269255  
      I/O Timings: read=1552.767  
      ->  Bitmap Heap Scan on zjxftypt.tab1010201 t_1  (cost=273453.65..432483144.47 rows=223 width=349) (actual time=25932.370..25932.370 rows=0 loops=1)  
            Output: t_1.storeid, t_1.xfjbh, t_1.wtsd, t_1.rs, t_1.digoal123x, t_1.dz, t_1.blfsjd, t_1.qx, t_1.gk, t_1.xfrq, t_1.djsj, t_1.djdw, t_1.xfjclzt, t_1.digoal123, t_1.xfxs  
            -- 位图扫描的条件重新过滤 , 过滤太多了  
	    Recheck Cond: ((t_1.xfrq < (to_date('2018-06-11'::character varying, 'yyyy-mm-dd'::character varying) + 1)) AND (t_1.xfrq >= to_date('2014-02-12'::character varying, 'yyyy-mm-dd'::character varying)) AND (t_1.digoal123 = 1::numeric))  
            Rows Removed by Index Recheck: 1214155  
              
	    -- 过滤exists的JOIN条件值是否满足 ，过滤太多了  
	    Filter: (((t_1.digoal123x)::text ~~ '%阿里巴巴%'::text) AND ((alternatives: SubPlan 4 or hashed SubPlan 5) OR(alternatives: SubPlan 6 or hashed SubPlan 7)))  
            Rows Removed by Filter: 5215804  
            Buffers: shared hit=920071 read=269255  
            I/O Timings: read=1552.767  
            -- 条件1,2位图扫描  
	    ->  BitmapAnd  (cost=273453.65..273453.65 rows=4909643 width=0) (actual time=2510.718..2510.718 rows=0 loops=1)  
                  Buffers: shared hit=27036 read=16539  
                  I/O Timings: read=101.425  
                  -- 自身条件1 符合条件的记录太多了  
		  ->  Bitmap Index Scan on index_tab1010201_xfrq  (cost=0.00..126565.99 rows=4943755 width=0) (actual time=1085.429..1085.429 rows=5268071 loops=1)  
                        Index Cond: ((t_1.xfrq < (to_date('2018-06-11'::character varying, 'yyyy-mm-dd'::character varying) + 1)) AND (t_1.xfrq >= to_date('2014-02-12'::character varying, 'yyyy-mm-dd'::character varying)))  
                        Buffers: shared hit=3288 read=16539  
                        I/O Timings: read=101.425  
                  -- 自身条件2 符合条件的记录太多了  
		  ->  Bitmap Index Scan on index_tab1010201_digoal123  (cost=0.00..146887.30 rows=6599316 width=0) (actual time=1355.825..1355.825 rows=6845646 loops=1)  
                        Index Cond: (t_1.digoal123 = 1::numeric)  
                        Buffers: shared hit=23748  
                  ..............sub plans

优化举例

1、复现问题，创建测试表

create table test(id int, c1 text, c2 date, c3 text);

SQL如下

select * from test   
where   
(  
  exists (select 1 from pg_class where oid::int = test.id)   
  or   
  exists (select 1 from pg_attribute where attrelid::int=test.id)   
)   
and c1 in ('1','2','3')   
and c2 between current_date-1 and current_date    
and c3 ~ 'abcdef';

2、写入测试护甲1000万条

insert into test select id, (random()*10)::int::text, current_date, md5(random()::text) from generate_series(1,10000000) t(id);

3、创建索引，使之可以在索引层面过滤掉所有数据

create extension pg_trgm;  
create extension btree_gin;  
create index idx_test_1 on test using gin (c1, c2, c3 gin_trgm_ops);

如果是复现问题，应该是这两个索引

create index idx1 on test (c1);  
  
create index idx2 on test (c2);

4、查看执行计划

postgres=# explain (analyze,verbose,timing,costs,buffers)   
select * from test   
where   
(  
  exists (select 1 from pg_class where oid::int = test.id)   
  or   
  exists (select 1 from pg_attribute where attrelid::int=test.id)   
)   
and c1 in ('1','2','3')   
and c2 between current_date-1 and current_date    
and c3 ~ 'abcdef';   
                                                                                         QUERY PLAN                                                                                           
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------  
 Bitmap Heap Scan on public.test  (cost=156.43..8593.79 rows=228 width=43) (actual time=837.151..837.151 rows=0 loops=1)  
   Output: test.id, test.c1, test.c2, test.c3  
   -- 位图扫描重新RECHECK过滤  
   Recheck Cond: ((test.c1 = ANY ('{1,2,3}'::text[])) AND (test.c2 >= (CURRENT_DATE - 1)) AND (test.c2 <= CURRENT_DATE) AND (test.c3 ~ 'abcdef'::text))  
   Rows Removed by Index Recheck: 1  
     
   -- exists子句的条件检查，过滤   
   Filter: ((alternatives: SubPlan 1 or hashed SubPlan 2) OR (alternatives: SubPlan 3 or hashed SubPlan 4))  
   Rows Removed by Filter: 7  
   Heap Blocks: exact=8  
   Buffers: shared hit=11658 read=23  
     
   -- 所有条件压到GIN复合索引里面  
   -- GIN多个条件时，会自动内部位图扫描  
   ->  Bitmap Index Scan on idx_test_1  (cost=0.00..156.37 rows=304 width=0) (actual time=834.418..834.418 rows=8 loops=1)  
         Index Cond: ((test.c1 = ANY ('{1,2,3}'::text[])) AND (test.c2 >= (CURRENT_DATE - 1)) AND (test.c2 <= CURRENT_DATE) AND (test.c3 ~ 'abcdef'::text))  
         Buffers: shared hit=11582 read=23  
   SubPlan 1  
     ->  Seq Scan on pg_catalog.pg_class  (cost=0.00..15.84 rows=1 width=0) (never executed)  
           Filter: ((pg_class.oid)::integer = test.id)  
   SubPlan 2  
     ->  Seq Scan on pg_catalog.pg_class pg_class_1  (cost=0.00..14.87 rows=387 width=4) (actual time=0.014..0.155 rows=388 loops=1)  
           Output: (pg_class_1.oid)::integer  
           Buffers: shared hit=11  
   SubPlan 3  
     ->  Index Only Scan using pg_attribute_relid_attnum_index on pg_catalog.pg_attribute  (cost=0.28..84.39 rows=8 width=0) (never executed)  
           Filter: ((pg_attribute.attrelid)::integer = test.id)  
           Heap Fetches: 0  
   SubPlan 4  
     ->  Index Only Scan using pg_attribute_relid_attnum_index on pg_catalog.pg_attribute pg_attribute_1  (cost=0.28..77.13 rows=2904 width=4) (actual time=0.029..1.081 rows=2941 loops=1)  
           Output: (pg_attribute_1.attrelid)::integer  
           Heap Fetches: 459  
           Buffers: shared hit=57  
 Planning time: 1.070 ms  
 Execution time: 839.834 ms  
(29 rows)

看起来还不错，但是仔细深究实际上并没有优化太多，还可以有更好的优化。

5、深入优化，需要理解GIN复合索引内部的执行机制(位图扫描)。

因为满足C3条件的记录本身就很少，所以完全不需要使用GIN内部的位图扫描。

postgres=# select count(*) from test where c3 ~ 'abcdef';  
 count   
-------  
    23  
(1 row)

修改为如下索引

postgres=# drop index idx_test_1 ;  
DROP INDEX  
  
postgres=# create index idx_test_1 on test using gin (c3 gin_trgm_ops) ;  
CREATE INDEX

6、耗时编程24毫秒

postgres=# explain (analyze,verbose,timing,costs,buffers)   
select * from test   
where   
(  
  exists (select 1 from pg_class where oid::int = test.id)   
  or   
  exists (select 1 from pg_attribute where attrelid::int=test.id)  
)   
and c1 in ('1','2','3')  
and c2 between current_date-1 and current_date   
and c3 ~ 'abcdef';  
                                                                                                       QUERY PLAN                                                                                                         
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------  
 Bitmap Heap Scan on public.test  (cost=53.76..27798.16 rows=228 width=43) (actual time=24.287..24.287 rows=0 loops=1)  
   Output: test.id, test.c1, test.c2, test.c3  
   Recheck Cond: (test.c3 ~ 'abcdef'::text)  
   Rows Removed by Index Recheck: 6  
   Filter: ((test.c1 = ANY ('{1,2,3}'::text[])) AND (test.c2 <= CURRENT_DATE) AND (test.c2 >= (CURRENT_DATE - 1)) AND ((alternatives: SubPlan 1 or hashed SubPlan 2) OR (alternatives: SubPlan 3 or hashed SubPlan 4)))  
   Rows Removed by Filter: 23  
   Heap Blocks: exact=29  
   Buffers: shared hit=226  
   ->  Bitmap Index Scan on idx_test_1  (cost=0.00..53.70 rows=1000 width=0) (actual time=21.517..21.517 rows=29 loops=1)  
         Index Cond: (test.c3 ~ 'abcdef'::text)  
         Buffers: shared hit=128  
   SubPlan 1  
     ->  Seq Scan on pg_catalog.pg_class  (cost=0.00..15.84 rows=1 width=0) (never executed)  
           Filter: ((pg_class.oid)::integer = test.id)  
   SubPlan 2  
     ->  Seq Scan on pg_catalog.pg_class pg_class_1  (cost=0.00..14.87 rows=387 width=4) (actual time=0.011..0.156 rows=387 loops=1)  
           Output: (pg_class_1.oid)::integer  
           Buffers: shared hit=11  
   SubPlan 3  
     ->  Index Only Scan using pg_attribute_relid_attnum_index on pg_catalog.pg_attribute  (cost=0.28..84.39 rows=8 width=0) (never executed)  
           Filter: ((pg_attribute.attrelid)::integer = test.id)  
           Heap Fetches: 0  
   SubPlan 4  
     ->  Index Only Scan using pg_attribute_relid_attnum_index on pg_catalog.pg_attribute pg_attribute_1  (cost=0.28..77.13 rows=2904 width=4) (actual time=0.028..1.099 rows=2938 loops=1)  
           Output: (pg_attribute_1.attrelid)::integer  
           Heap Fetches: 456  
           Buffers: shared hit=58  
 Planning time: 0.801 ms  
 Execution time: 24.403 ms  
(29 rows)  
  
Time: 26.052 ms

小结

本文的SQL比较复杂，优化的思路和其他SQL差不多，只是本例可以理解BITMAP SCAN以及GIN索引的内部BITMAP SCAN在对较大数据进行合并时，可能引入的开销。

切入点依旧是explain，找耗时段，找背后的原因，解决。

1、什么时候使用GIN复合？

当任意一个条件，选择性不好时，使用复合。

什么时候使用GIN非复合？

2、当有有一个条件，选择性很好时，把它单独拿出来，作为一个独立索引。比如本例的c3模糊查询字段，过滤性好，应该单独拿出来。

其实就是说，选择性不好的列，不要放到索引里面，即使要放，也应该等PG出了分区索引后，将这种列作为分区索引的分区键。(多颗树)，或者使用partial index。

《PostgreSQL 黑科技 - 空间聚集存储, 内窥GIN, GiST, SP-GiST索引》

《宝剑赠英雄 - 任意组合字段等效查询, 探探PostgreSQL多列展开式B树 (GIN)》

《PostgreSQL GIN索引实现原理》

《PostgreSQL GIN multi-key search 优化》

《从难缠的模糊查询聊开 - PostgreSQL独门绝招之一 GIN , GiST , SP-GiST , RUM 索引原理与技术背景》

《PostgreSQL GIN索引limit慢的原因分析》

《beyond b-tree (gin\gist索引讲解PDF)》

Flag Counter

digoal’s 大量PostgreSQL文章入口

You May Also Enjoy

PostgreSQL(PPAS 兼容Oracle) 从零开始入门手册 - 珍藏版

17 minute read

背景 云数据库PPAS版，是阿里云与EnterpriseDB公司(简称EDB)合作基于PostgreSQL高度兼容Oracle语法的数据库服务，为用户提供易于操作的迁移工具，兼容范围涵盖：PL/SQL、数据类型、高级函数、表分区等。 用户可以直接在阿里云购买PPAS进行使用。 如果在购买PPAS前，想试用一下...

PostgreSQL pipelinedb 流计算插件 - IoT应用 - 实时轨迹聚合

1 minute read

背景 IoT场景，车联网场景，共享单车场景，人的行为位点等，终端实时上报的是孤立的位点，我们需要将其补齐成轨迹。 例如共享单车，下单，开锁，生成订单，骑行，关闭订单，关锁。这个过程有一个唯一的订单号，每次上报的位点会包含时间，订单号，位置。 根据订单号，将点聚合为轨迹。 使用pipelinedb插件，可以实...

PostgreSQL datediff 日期间隔（单位转换）兼容SQL用法

2 minute read

背景 使用datediff，对时间或日期相减，得到的间隔，转换为目标单位(日、月、季度、年、小时、秒。。。等)的数值。 DATEDIFF ( datepart, {date|timestamp}, {date|timestamp} ) 周 select datediff(week,'2009-01-0...