postgresql 中的 like 查询优化方案

当时数量量比较庞大的时候，做模糊查询效率很慢，为了优化查询效率，尝试如下方法做效率对比

一、对比情况说明：

1、数据量100w条数据

2、执行sql

二、对比结果

explain analyze select
 c_patent,
 c_applyissno,
 d_applyissdate,
 d_applydate,
 c_patenttype_dimn,
 c_newlawstatus,
 c_abstract 
from
 public.t_knowl_patent_zlxx_temp 
where
 c_applicant like '%本溪满族自治县连山关镇安平安养殖场%';

1、未建索时执行计划：

"gather (cost=1000.00..83803.53 rows=92 width=1278) (actual time=217.264..217.264 rows=0 loops=1)
 workers planned: 2
 workers launched: 2
 -> parallel seq scan on t_knowl_patent_zlxx (cost=0.00..82794.33 rows=38 width=1278) (actual time=212.355..212.355 rows=0 loops=3)
  filter: ((c_applicant)::text ~~ '%本溪满族自治县连山关镇安平安养殖场%'::text)
  rows removed by filter: 333333
planning time: 0.272 ms
execution time: 228.116 ms"

2、btree索引

建索引语句

create index idx_public_t_knowl_patent_zlxx_applicant on public.t_knowl_patent_zlxx(c_applicant varchar_pattern_ops);

执行计划

"gather (cost=1000.00..83803.53 rows=92 width=1278) (actual time=208.253..208.253 rows=0 loops=1)
 workers planned: 2
 workers launched: 2
 -> parallel seq scan on t_knowl_patent_zlxx (cost=0.00..82794.33 rows=38 width=1278) (actual time=203.573..203.573 rows=0 loops=3)
  filter: ((c_applicant)::text ~~ '%本溪满族自治县连山关镇安平安养殖场%'::text)
  rows removed by filter: 333333
planning time: 0.116 ms
execution time: 218.189 ms"

但是如果将查询sql稍微改动一下，把like查询中的前置%去掉是这样的

index scan using idx_public_t_knowl_patent_zlxx_applicant on t_knowl_patent_zlxx_temp (cost=0.55..8.57 rows=92 width=1278) (actual time=0.292..0.292 rows=0 loops=1)
 index cond: (((c_applicant)::text ~>=~ '本溪满族自治县连山关镇安平安养殖场'::text) and ((c_applicant)::text ~<~ '本溪满族自治县连山关镇安平安养殖圻'::text))
 filter: ((c_applicant)::text ~~ '本溪满族自治县连山关镇安平安养殖场%'::text)
planning time: 0.710 ms
execution time: 0.378 ms

3、gin索引

创建索引语句(postgresql要求在9.6版本及以上)

create extension pg_trgm;
create index idx_public_t_knowl_patent_zlxx_applicant on public.t_knowl_patent_zlxx using gin (c_applicant gin_trgm_ops);

执行计划

bitmap heap scan on t_knowl_patent_zlxx (cost=244.71..600.42 rows=91 width=1268) (actual time=0.649..0.649 rows=0 loops=1)
 recheck cond: ((c_applicant)::text ~~ '%本溪满族自治县连山关镇安平安养殖场%'::text)
 -> bitmap index scan on idx_public_t_knowl_patent_zlxx_applicant (cost=0.00..244.69 rows=91 width=0) (actual time=0.647..0.647 rows=0 loops=1)
  index cond: ((c_applicant)::text ~~ '%本溪满族自治县连山关镇安平安养殖场%'::text)
planning time: 0.673 ms
execution time: 0.740 ms

三、结论

btree索引可以让后置% "abc%"的模糊匹配走索引，gin + gp_trgm可以让前后置% "%abc%" 走索引。但是gin 索引也有弊端，以下情况可能导致无法命中：

搜索字段少于3个字符时，不会命中索引，这是gin自身机制导致。

当搜索字段过长时，比如email检索，可能也不会命中索引，造成原因暂时未知。

补充：postgresql like 查询效率提升实验

一、未做索引的查询效率

作为对比，先对未索引的查询做测试

explain analyze select * from gallery_map where author = '曹志耘';
             query plan             
-----------------------------------------------------------------------------------------------------------------
 seq scan on gallery_map (cost=0.00..7002.32 rows=1025 width=621) (actual time=0.011..39.753 rows=1031 loops=1)
 filter: ((author)::text = '曹志耘'::text)
 rows removed by filter: 71315
 planning time: 0.194 ms
 execution time: 39.879 ms
(5 rows)
 
time: 40.599 ms
explain analyze select * from gallery_map where author like '曹志耘';
             query plan             
-----------------------------------------------------------------------------------------------------------------
 seq scan on gallery_map (cost=0.00..7002.32 rows=1025 width=621) (actual time=0.017..41.513 rows=1031 loops=1)
 filter: ((author)::text ~~ '曹志耘'::text)
 rows removed by filter: 71315
 planning time: 0.188 ms
 execution time: 41.669 ms
(5 rows)
 
time: 42.457 ms
 
explain analyze select * from gallery_map where author like '曹志耘%';
             query plan             
-----------------------------------------------------------------------------------------------------------------
 seq scan on gallery_map (cost=0.00..7002.32 rows=1028 width=621) (actual time=0.017..41.492 rows=1031 loops=1)
 filter: ((author)::text ~~ '曹志耘%'::text)
 rows removed by filter: 71315
 planning time: 0.307 ms
 execution time: 41.633 ms
(5 rows)
 
time: 42.676 ms

很显然都会做全表扫描

二、创建btree索引

postgresql默认索引是btree

create index ix_gallery_map_author on gallery_map (author);
 
explain analyze select * from gallery_map where author = '曹志耘';  
                query plan                
-------------------------------------------------------------------------------------------------------------------------------------
 bitmap heap scan on gallery_map (cost=36.36..2715.37 rows=1025 width=621) (actual time=0.457..1.312 rows=1031 loops=1)
 recheck cond: ((author)::text = '曹志耘'::text)
 heap blocks: exact=438
 -> bitmap index scan on ix_gallery_map_author (cost=0.00..36.10 rows=1025 width=0) (actual time=0.358..0.358 rows=1031 loops=1)
   index cond: ((author)::text = '曹志耘'::text)
 planning time: 0.416 ms
 execution time: 1.422 ms
(7 rows)
 
time: 2.462 ms
 
explain analyze select * from gallery_map where author like '曹志耘';
                query plan                
-------------------------------------------------------------------------------------------------------------------------------------
 bitmap heap scan on gallery_map (cost=36.36..2715.37 rows=1025 width=621) (actual time=0.752..2.119 rows=1031 loops=1)
 filter: ((author)::text ~~ '曹志耘'::text)
 heap blocks: exact=438
 -> bitmap index scan on ix_gallery_map_author (cost=0.00..36.10 rows=1025 width=0) (actual time=0.560..0.560 rows=1031 loops=1)
   index cond: ((author)::text = '曹志耘'::text)
 planning time: 0.270 ms
 execution time: 2.295 ms
(7 rows)
 
time: 3.444 ms
explain analyze select * from gallery_map where author like '曹志耘%';
             query plan             
-----------------------------------------------------------------------------------------------------------------
 seq scan on gallery_map (cost=0.00..7002.32 rows=1028 width=621) (actual time=0.015..41.389 rows=1031 loops=1)
 filter: ((author)::text ~~ '曹志耘%'::text)
 rows removed by filter: 71315
 planning time: 0.260 ms
 execution time: 41.518 ms
(5 rows)
 
time: 42.430 ms
explain analyze select * from gallery_map where author like '%研究室';
             query plan             
-----------------------------------------------------------------------------------------------------------------
 seq scan on gallery_map (cost=0.00..7002.32 rows=2282 width=621) (actual time=0.064..52.824 rows=2152 loops=1)
 filter: ((author)::text ~~ '%研究室'::text)
 rows removed by filter: 70194
 planning time: 0.254 ms
 execution time: 53.064 ms
(5 rows)
 
time: 53.954 ms

可以看到，等于、like的全匹配是用到索引的，like的模糊查询还是全表扫描

三、创建gin索引

create extension pg_trgm;
 
create index ix_gallery_map_author on gallery_map using gin (author gin_trgm_ops);
explain analyze select * from gallery_map where author like '曹%'; 
                query plan                
-------------------------------------------------------------------------------------------------------------------------------------
 bitmap heap scan on gallery_map (cost=19.96..2705.69 rows=1028 width=621) (actual time=0.419..1.771 rows=1031 loops=1)
 recheck cond: ((author)::text ~~ '曹%'::text)
 heap blocks: exact=438
 -> bitmap index scan on ix_gallery_map_author (cost=0.00..19.71 rows=1028 width=0) (actual time=0.312..0.312 rows=1031 loops=1)
   index cond: ((author)::text ~~ '曹%'::text)
 planning time: 0.358 ms
 execution time: 1.916 ms
(7 rows)
 
time: 2.843 ms
explain analyze select * from gallery_map where author like '%耘%'; 
             query plan             
-----------------------------------------------------------------------------------------------------------------
 seq scan on gallery_map (cost=0.00..7002.32 rows=1028 width=621) (actual time=0.015..51.641 rows=1031 loops=1)
 filter: ((author)::text ~~ '%耘%'::text)
 rows removed by filter: 71315
 planning time: 0.268 ms
 execution time: 51.957 ms
(5 rows)
 
time: 52.899 ms
explain analyze select * from gallery_map where author like '%研究室%';
                query plan                
-------------------------------------------------------------------------------------------------------------------------------------
 bitmap heap scan on gallery_map (cost=31.83..4788.42 rows=2559 width=621) (actual time=0.914..4.195 rows=2402 loops=1)
 recheck cond: ((author)::text ~~ '%研究室%'::text)
 heap blocks: exact=868
 -> bitmap index scan on ix_gallery_map_author (cost=0.00..31.19 rows=2559 width=0) (actual time=0.694..0.694 rows=2402 loops=1)
   index cond: ((author)::text ~~ '%研究室%'::text)
 planning time: 0.306 ms
 execution time: 4.403 ms
(7 rows)
 
time: 5.227 ms

gin_trgm索引的效果好多了

由于pg_trgm的索引是把字符串切成多个3元组，然后使用这些3元组做匹配，所以gin_trgm索引对于少于3个字符（包括汉字）的查询，只有前缀匹配会走索引

另外，还测试了btree_gin，效果和btree一样

注意：

gin_trgm要求数据库必须使用utf-8编码

demo_v1 # \l demo_v1
        list of databases
 name | owner | encoding | collate | ctype | access privileges
---------+-----------+----------+-------------+-------------+-------------------
 demo_v1 | wmpp_user | utf8  | en_us.utf-8 | en_us.utf-8 |
 

以上为个人经验，希望能给大家一个参考，也希望大家多多支持。如有错误或未考虑完全的地方，望不吝赐教。