PostgreSQL 索引虚拟列 - 表达式索引 - JOIN提速

背景

CASE： 使用虚拟索引，响应时间从2.3秒下降到0.3毫秒

业务系统在设计时，为了减少数据冗余，提升可读性，通常需要将不同的数据放到不同的表。

在查询时，通过多表JOIN来补齐需要查询或在过滤的内容。

比如这样的例子：

有两张表，分别有1千万和100万数据，当用户查询时，需要补齐那100万表中的某个字段进行过滤。

create table a (id int, bid int, c1 int, c2 int, c3 int);  
  
CREATE TABLE b (id int primary key, path text);  
  
insert into a select id, random()*1000000 , random()*10000000, random()*10000000 , random()*10000000 from generate_series(1,10000000) t(id);  
  
insert into b select id, md5(random()::text) from generate_series(1,1000000) t(id);  
  
create index idx_b_1 on b(path text_pattern_ops);  
  
-- 查询  
select a.* from a left join b on (a.bid=b.id and b.path like 'abc%');  

那么它的性能如何呢？

postgres=# explain select a.* from a left join b on (a.bid=b.id) where b.path like 'abcde%';  
                                     QUERY PLAN                                       
------------------------------------------------------------------------------------  
 Hash Join  (cost=9.70..289954.61 rows=1000 width=20)  
   Hash Cond: (a.bid = b.id)  
   ->  Seq Scan on a  (cost=0.00..163695.00 rows=10000000 width=20)  
   ->  Hash  (cost=8.45..8.45 rows=100 width=4)  
         ->  Index Scan using idx_b_1 on b  (cost=0.42..8.45 rows=100 width=4)  
               Index Cond: ((path ~>=~ 'abcde'::text) AND (path ~<~ 'abcdf'::text))  
               Filter: (path ~~ 'abcde%'::text)  
(7 rows)  
  
Time: 0.777 ms  
postgres=# select a.* from a left join b on (a.bid=b.id) where b.path like 'abcde%';  
   id    |  bid   |   c1    |   c2    |   c3      
---------+--------+---------+---------+---------  
 2423577 | 633740 |  846719 | 1720744 |  416608  
 2433286 | 633740 | 9797626 | 6737349 | 5669893  
 3851817 | 633740 | 8764393 | 3779499 | 2830950  
 4889541 | 633740 | 3892055 | 9470525 |  611262  
 5004634 | 633740 | 5420943 | 2448245 | 5719976  
 5372019 | 633740 | 5402891 | 3441462 | 8194368  
 6051251 | 633740 | 8691218 | 7184625 | 5940346  
 6344344 | 633740 | 5869018 | 9352883 |  636112  
 9751456 | 633740 | 3797867 | 1934900 | 2511398  
(9 rows)  
  
Time: 2348.506 ms (00:02.349)  

条件越多，性能会越差。

这样的查询，并发一高，性能影响会比较大。

当b表是静态的时(没有DML)，可以用虚拟列索引来实现优化。

表达式索引 - 虚拟列索引

假设B表不会发生DML，是一个静态表。

1、创建一个获取path的函数

create or replace function get_path(int) returns text as $$  
  select path from b where id=$1;  
$$ language sql strict immutable;  

这个函数用于从B表获取path，假设B表静态（不会有增删改），那么这个函数就是immutable的，无论什么时候输入一个ID，返回的都是同一个path。

2、在a表直接创建表达式索引（虚拟列索引）

create index idx_a_1 on a(get_path(bid) text_pattern_ops);  

3、修改SQL语句如下

select * from a where get_path(bid) like 'abc%';  

4、响应时间从2.3秒下降到0.3毫秒。

postgres=# select * from a where get_path(bid) like 'abcde%';  
   id    |  bid   |   c1    |   c2    |   c3      
---------+--------+---------+---------+---------  
 2423577 | 633740 |  846719 | 1720744 |  416608  
 2433286 | 633740 | 9797626 | 6737349 | 5669893  
 3851817 | 633740 | 8764393 | 3779499 | 2830950  
 4889541 | 633740 | 3892055 | 9470525 |  611262  
 5004634 | 633740 | 5420943 | 2448245 | 5719976  
 5372019 | 633740 | 5402891 | 3441462 | 8194368  
 6051251 | 633740 | 8691218 | 7184625 | 5940346  
 6344344 | 633740 | 5869018 | 9352883 |  636112  
 9751456 | 633740 | 3797867 | 1934900 | 2511398  
(9 rows)  
  
postgres=# select * from b where path like 'abcde%';  
   id   |               path                 
--------+----------------------------------  
 633740 | abcde980c8568a9a6a140885d92fcebe  
(1 row)  
  
  
postgres=# explain (analyze,verbose,timing,costs,buffers) select * from a where get_path(bid) like 'abcde%';  
                                                         QUERY PLAN                                                           
----------------------------------------------------------------------------------------------------------------------------  
 Index Scan using idx_a_1 on public.a  (cost=0.56..158697.56 rows=50000 width=20) (actual time=0.151..0.276 rows=9 loops=1)  
   Output: id, bid, c1, c2, c3  
   Index Cond: ((get_path(a.bid) ~>=~ 'abcde'::text) AND (get_path(a.bid) ~<~ 'abcdf'::text))  
   Filter: (get_path(a.bid) ~~ 'abcde%'::text)  
   Buffers: shared hit=50  
 Planning time: 0.092 ms  
 Execution time: 0.300 ms  
(7 rows)  

即使没有虚拟索引，也可以PostgreSQL的使用并行计算

32个并行，耗时454毫秒，依旧不如使用虚拟列索引的效果。

postgres=# set parallel_tuple_cost =0;  
SET  
postgres=# set parallel_setup_cost =0;  
SET  
postgres=# set max_parallel_workers_per_gather =32;  
SET  
postgres=# alter table a set (parallel_workers =32);  
ALTER TABLE  
  
postgres=# explain select a.* from a left join b on (a.bid=b.id) where b.path like 'abcde%';  
                                   QUERY PLAN                                     
--------------------------------------------------------------------------------  
 Gather  (cost=20835.25..91600.56 rows=1000 width=20)  
   Workers Planned: 32  
   ->  Hash Join  (cost=20835.25..91600.56 rows=31 width=20)  
         Hash Cond: (a.bid = b.id)  
         ->  Parallel Seq Scan on a  (cost=0.00..66820.00 rows=312500 width=20)  
         ->  Hash  (cost=20834.00..20834.00 rows=100 width=4)  
               ->  Seq Scan on b  (cost=0.00..20834.00 rows=100 width=4)  
                     Filter: (path ~~ 'abcde%'::text)  
(8 rows)  
  
Time: 0.685 ms  
  
postgres=# select a.* from a left join b on (a.bid=b.id) where b.path like 'abcde%';  
   id    |  bid   |   c1    |   c2    |   c3      
---------+--------+---------+---------+---------  
 5004634 | 633740 | 5420943 | 2448245 | 5719976  
 9751456 | 633740 | 3797867 | 1934900 | 2511398  
 3851817 | 633740 | 8764393 | 3779499 | 2830950  
 4889541 | 633740 | 3892055 | 9470525 |  611262  
 6344344 | 633740 | 5869018 | 9352883 |  636112  
 2433286 | 633740 | 9797626 | 6737349 | 5669893  
 2423577 | 633740 |  846719 | 1720744 |  416608  
 6051251 | 633740 | 8691218 | 7184625 | 5940346  
 5372019 | 633740 | 5402891 | 3441462 | 8194368  
(9 rows)  
  
Time: 454.405 ms  

digoal’s 大量PostgreSQL文章入口