使用参数而不是硬编码字符串时,Postgres查询非常慢

Question

我遇到了这个Postgres问题,如果我在查询字符串上使用参数vs硬编码它的相同查询需要很长时间才能执行.列名是'media_type',它是VARCHAR(20).我正在使用Symfony2和Doctrine2 ORM从PHP运行这些查询,并且该表有大约1.000.000条记录.

我的查询有问题吗？它可能是Postgres配置问题吗？

1 - media_type的硬编码值

duration: 5.365 ms  parse pdo_stmt_00000001: SELECT id,site_id FROM item where media_type = 'Collection' AND enabled = 'true' AND site_id = $1 AND user_id = $2 ORDER BY id DESC LIMIT $3 OFFSET $4
duration: 0.142 ms  bind pdo_stmt_00000001: SELECT id,site_id FROM item where media_type = 'Collection' AND enabled = 'true' AND site_id = $1 AND user_id = $2 ORDER BY id DESC LIMIT $3 OFFSET $4
parameters: $1 = '1', $2 = '1', $3 = '100', $4 = '0'
duration: 8.667 ms  execute pdo_stmt_00000001: SELECT id,site_id FROM item where media_type = 'Collection' AND enabled = 'true' AND site_id = $1 AND user_id = $2 ORDER BY id DESC LIMIT $3 OFFSET $4
parameters: $1 = '1', $2 = '1', $3 = '100', $4 = '0'

执行计划:

duration: 8.640 ms  plan:
    Query Text: SELECT id,site_id FROM item where media_type = 'Collection' AND enabled = 'true' AND site_id = $1 AND user_id = $2 ORDER BY id DESC LIMIT $3 OFFSET $4
    Limit  (cost=8.38..8.38 rows=1 width=12) (actual time=8.516..8.595 rows=24 loops=1)
      Buffers: shared hit=10 read=15
        ->  Sort  (cost=8.38..8.38 rows=1 width=12) (actual time=8.505..8.530 rows=24 loops=1)
            Sort Key: id
            Sort Method: quicksort  Memory: 26kB
            Buffers: shared hit=10 read=15
            ->  Index Scan using item_media_type_index on item  (cost=0.00..8.37 rows=1 width=12) (actual time=7.955..8.397 rows=24 loops=1)
                    Index Cond: ((media_type)::text = 'Collection'::text)
                    Filter: (enabled AND (site_id = $1) AND (user_id = $2))
                    Buffers: shared hit=8 read=15

2 - 使用media_type参数(SLOWER)

duration: 5.557 ms  parse pdo_stmt_00000001: SELECT id,site_id FROM item where media_type = $1 AND enabled = 'true' AND site_id = $2 AND user_id = $3 ORDER BY id DESC LIMIT $4 OFFSET $5
duration: 1.322 ms  bind pdo_stmt_00000001: SELECT id,site_id FROM item where media_type = $1 AND enabled = 'true' AND site_id = $2 AND user_id = $3 ORDER BY id DESC LIMIT $4 OFFSET $5
parameters: $1 = 'Collection', $2 = '1', $3 = '1', $4 = '100', $5 = '0'
duration: 71564.998 ms  execute pdo_stmt_00000001: SELECT id,site_id FROM item where media_type = $1 AND enabled = 'true' AND site_id = $2 AND user_id = $3 ORDER BY id DESC LIMIT $4 OFFSET $5
parameters: $1 = 'Collection', $2 = '1', $3 = '1', $4 = '100', $5 = '0'

执行计划:

duration: 71564.922 ms  plan:
    Query Text: SELECT id,site_id FROM item where media_type = $1 AND enabled = 'true' AND site_id = $2 AND user_id = $3 ORDER BY id DESC LIMIT $4 OFFSET $5
    Limit  (cost=90663.16..181326.31 rows=17184 width=12) (actual time=3.667..71564.864 rows=24 loops=1)
      Buffers: shared hit=183786 read=96585
        ->  Index Scan Backward using item_pkey on item  (cost=0.00..906610.46 rows=171836 width=12) (actual time=3.655..71564.798 rows=24 loops=1)
               Filter: (enabled AND ((media_type)::text = $1) AND (site_id = $2) AND (user_id = $3))
               Buffers: shared hit=183786 read=96585

提前致谢.

Answer 1

这在PostgreSQL中是一个长期存在的瑕疵,历史上需要一些有趣的规划师调整来解决.它已经在PostgreSQL 9.2(现在测试版)中得到修复,但是像往常一样感谢Tom Lane.

E.1.3.1.3.优化

提高规划人员选择参数化计划的能力(Tom Lane)

现在,已准备好的语句被解析,分析和重写,但不一定是计划好的.当使用参数执行准备好的计划时,计划员可以为每个常量重新计划它,或者如果其成本接近于特定于常量的计划,则可以执行通用计划.

请参阅9.2 beta版发行说明以及我在lwn.net上撰写的关于此内容的快速说明.有很多关于处理准备/参数化语句的信息,这些语句在邮件列表上运行速度比普通语句慢.