为什么这个 LEFT JOIN 运行缓慢？

Question

PostgreSQL 查询的LEFT JOIN 部分运行非常缓慢，我不知道为什么。

完整的查询：

SELECT t.id FROM tests t
LEFT JOIN tests c ON c.parent_id IN (t.id, t.parent_id)
INNER JOIN responses r ON (
    r.test_id IN (t.id, t.parent_id, c.id)
) WHERE r.user_id = 333

tests.id 和 tests.parent_id 上有索引。

测试包含 28876 行（其中有 1282 行 WHERE parent_id IS NOT NULL）。

查询的 LEFT JOIN 部分生成 32098 行，大约需要 700 毫秒。

SELECT t.id FROM tests t
LEFT JOIN tests c ON c.parent_id IN (t.id, t.parent_id)

查询的其余部分花费的时间可以忽略不计。

关于为什么它可能很慢的任何想法，或者实现相同目标的更好方法？

谢谢！

---

选择版本（）

PostgreSQL 9.1.9 on x86_64-unknown-linux-gnu, compiled by gcc (Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3, 64-bit

解释分析

（注意：这里使用的是真实的表名usability_tests，我在前面的示例中将其简化为tests。）

Nested Loop  (cost=5.18..158692.45 rows=80 width=4) (actual time=107.873..5718.295 rows=103 loops=1)
  Join Filter: ((r.usability_test_id = t.id) OR (r.usability_test_id = t.parent_id) OR (r.usability_test_id = c.id))
  ->  Nested Loop Left Join  (cost=0.56..136015.63 rows=28876 width=12) (actual time=0.091..486.496 rows=32098 loops=1)
        Join Filter: ((c.parent_id = t.id) OR (c.parent_id = t.parent_id))
        ->  Seq Scan on usability_tests t  (cost=0.00..1455.76 rows=28876 width=8) (actual time=0.042..39.558 rows=28876 loops=1)
        ->  Bitmap Heap Scan on usability_tests c  (cost=0.56..4.60 rows=4 width=8) (actual time=0.010..0.011 rows=0 loops=28876)
              Recheck Cond: ((parent_id = t.id) OR (parent_id = t.parent_id))
              ->  BitmapOr  (cost=0.56..0.56 rows=4 width=0) (actual time=0.008..0.008 rows=0 loops=28876)
                    ->  Bitmap Index Scan on index_usability_tests_on_parent_id  (cost=0.00..0.28 rows=2 width=0) (actual time=0.003..0.003 rows=0 loops=28876)
                          Index Cond: (parent_id = t.id)
                    ->  Bitmap Index Scan on index_usability_tests_on_parent_id  (cost=0.00..0.28 rows=2 width=0) (actual time=0.001..0.001 rows=0 loops=28876)
                          Index Cond: (parent_id = t.parent_id)
  ->  Materialize  (cost=4.62..153.63 rows=39 width=4) (actual time=0.001..0.076 rows=70 loops=32098)
        ->  Bitmap Heap Scan on responses r  (cost=4.62..153.44 rows=39 width=4) (actual time=0.053..0.187 rows=70 loops=1)
              Recheck Cond: (user_id = 3649)
              ->  Bitmap Index Scan on index_responses_on_user_id  (cost=0.00..4.61 rows=39 width=0) (actual time=0.040..0.040 rows=70 loops=1)
                    Index Cond: (user_id = 3649)
Total runtime: 5718.592 ms

Answer 1

更新：看起来您的查询基本上是这样的

with cte as (
    select r.test_id
    from responses as r
    where r.user_id = 333
    union all
    select c.parent_id
    from tests as c
        inner join responses as r on r.test_id = c.id
    where r.user_id = 333
)
select
    t.id
from tests as t
where
    t.id in (select c.test_id from cte as c) or
    t.parent_id in (select c.test_id from cte as c)

old：试着把它变成这个查询，看看会不会更快：

select t.id 
from tests t
    inner join tests c on c.parent_id = t.id

union all

select t.id 
from tests t
    inner join tests c oN c.parent_id = t.parent_id

执行其中一项查询需要多长时间？

Answer 2

我认为查询可以简化为：

SELECT t.id FROM tests t
WHERE EXISTS ( 
        SELECT * FROM responses r
        WHERE (r.test_id = t.id OR r.test_id = t.parent_id )
        AND r.user_id = 333
        )
OR EXISTS (
        SELECT * FROM responses r 
        JOIN tests c ON r.test_id = c.id
            -- Note: the ... OR sibling makes no sense to me
        WHERE (c.parent_id = t.id OR c.parent_id = t.parent_id)
        AND r.user_id = 333
        );

注意：问题中的查询可能为t.id 生成重复值。这个只报告不同的值。

更新：我刚刚测试了它（在合成数据上），上面的查询返回的结果与原始的完全相同减去重复项。

UPDATE2：添加了兄弟匹配。