此查询的 Postgresql 索引？

Question

是否有任何索引有助于此类查询？

select p.name 
from person p 
where 'abjohncde' like '%'||p.name||'%'

使用 gin 和 gin_trgm_ops 的索引可用于此查询：

select p.name from person p where p.name like '%john%'

但 gin 不适用于第一个查询。

Answer 1

pg_trgm 适用于select * from table where field like '%pattern%' 之类的情况，但不幸的是，它在select * from table where 'abcde' like '%'||field||'%' 或select * from table where 'abcde' ~ field 之类的情况下无济于事（语义相同，但第一个更快）

相反，您可以尝试手动创建三元组索引：

create or replace function my_trgm(x text)
    returns text[]
    language sql
    immutable strict
as $$
    select
        case when length(x) > 3 then
            (select array_agg(distinct lower(substr(x, i, 3))) from generate_series(1, length(x)-2) as i)
            else array[x]
        end
$$;

create or replace function public.my_trgm_1(x text)
    returns text[]
    language sql
    immutable strict
as $$
    select
        (select array_agg(distinct lower(substr(x, i, 3))) from generate_series(1, length(x)-2) as i) ||
        (select array_agg(distinct lower(substr(x, i, 2))) from generate_series(1, length(x)-1) as i) ||
        (select array_agg(distinct lower(substr(x, i, 1))) from generate_series(1, length(x)) as i)
$$;

并测试：

postgres=# create table t as select ((random()*1000)::int)::text as x from generate_series(1, 1000000);
SELECT 1000000
postgres=# create index idx_t_x on t using gin (my_trgm(x));
CREATE INDEX

postgres=# explain (analyze, verbose) select * from t where '16662' like '%'||x||'%';
                                                         QUERY PLAN                                                         
----------------------------------------------------------------------------------------------------------------------------
 Gather  (cost=1000.00..13216.67 rows=5000 width=3) (actual time=0.338..65.335 rows=8971 loops=1)
   Output: x
   Workers Planned: 2
   Workers Launched: 2
   ->  Parallel Seq Scan on public.t  (cost=0.00..11716.67 rows=2083 width=3) (actual time=0.106..59.276 rows=2990 loops=3)
         Output: x
         Filter: ('16662'::text ~~ (('%'::text || t.x) || '%'::text))
         Rows Removed by Filter: 330343
         Worker 0: actual time=0.089..57.952 rows=3126 loops=1
         Worker 1: actual time=0.124..58.830 rows=3308 loops=1
 Planning Time: 0.061 ms
 Execution Time: 65.803 ms

postgres=# explain (analyze, verbose) select * from t where my_trgm(x) <@ my_trgm_1('16662') and '16662' ~ x;
                                                       QUERY PLAN                                                       
------------------------------------------------------------------------------------------------------------------------
 Bitmap Heap Scan on public.t  (cost=184.40..6916.58 rows=43 width=3) (actual time=4.655..70.851 rows=8971 loops=1)
   Output: x
   Recheck Cond: (my_trgm(t.x) <@ '{166,666,662,16,66,62,1,6,2}'::text[])
   Filter: ('16662'::text ~ t.x)
   Heap Blocks: exact=3876
   ->  Bitmap Index Scan on idx_t_x  (cost=0.00..184.39 rows=8586 width=0) (actual time=3.534..3.534 rows=8971 loops=1)
         Index Cond: (my_trgm(t.x) <@ '{166,666,662,16,66,62,1,6,2}'::text[])
 Planning Time: 0.578 ms
 Execution Time: 71.228 ms

请注意，它在单列表上执行，因此在此特定测试用例中扫描整个表可能比扫描 GIN 索引更有效。

PS：使用一些 PL 语言，如 plpythonu 或 plperl 可能更有效地插入/更新数据。