Question

我正在使用 PostgreSQL ，我的SQL查询有一个奇怪的问题。 取决于我正在使用的日期参数。我的请求没有做同样的操作。

这是我的工作查询：

SELECT DISTINCT app.id_application 
FROM stat sj
LEFT OUTER JOIN groupe gp ON gp.id_groupe = sj.id_groupe 
LEFT OUTER JOIN application app ON app.id_application = gp.id_application 
WHERE date_stat >= '2016/3/01' 
AND date_stat <= '2016/3/31' 
AND ( date_stat = date_gen-1 or (date_gen = '2016/04/01' AND date_stat = '2016/3/31')) 
AND app.id_application IS NOT NULL

此查询需要大约2个secondes（这对我来说是好的，因为我有很多行）。当我为这个查询运行EXPLAIN ANALYZE时，我有这个：

HashAggregate  (cost=375486.95..375493.62 rows=667 width=4) (actual time=2320.541..2320.656 rows=442 loops=1)
    ->  Hash Join  (cost=254.02..375478.99 rows=3186 width=4) (actual time=6.144..2271.984 rows=263274 loops=1)
    Hash Cond: (gp.id_application = app.id_application)
    ->  Hash Join  (cost=234.01..375415.17 rows=3186 width=4) (actual time=5.926..2200.671 rows=263274 loops=1)
          Hash Cond: (sj.id_groupe = gp.id_groupe)
          ->  Seq Scan on stat sj  (cost=0.00..375109.47 rows=3186 width=8) (actual time=3.196..2068.357 rows=263274 loops=1)
                Filter: ((date_stat >= '2016-03-01'::date) AND (date_stat <= '2016-03-31'::date) AND ((date_stat = (date_gen - 1)) OR ((date_gen = '2016-04-01'::date) AND (date_stat = '2016-03-31'::date))))
                Rows Removed by Filter: 7199514
          ->  Hash  (cost=133.45..133.45 rows=8045 width=12) (actual time=2.677..2.677 rows=8019 loops=1)
                Buckets: 1024  Batches: 1  Memory Usage: 345kB
                ->  Seq Scan on groupe gp  (cost=0.00..133.45 rows=8045 width=12) (actual time=0.007..1.284 rows=8019 loops=1)
    ->  Hash  (cost=11.67..11.67 rows=667 width=4) (actual time=0.206..0.206 rows=692 loops=1)
          Buckets: 1024  Batches: 1  Memory Usage: 25kB
          ->  Seq Scan on application app  (cost=0.00..11.67 rows=667 width=4) (actual time=0.007..0.101 rows=692 loops=1)
                Filter: (id_application IS NOT NULL)
    Total runtime: 2320.855 ms

现在，当我正在尝试使用相同查询的当前月份（我们是4月6日，所以我试图获得4月份的所有application_id ）的相同查询时/ p>

SELECT DISTINCT app.id_application 
FROM stat sj
LEFT OUTER JOIN groupe gp ON gp.id_groupe = sj.id_groupe 
LEFT OUTER JOIN application app ON app.id_application = gp.id_application 
WHERE date_stat >= '2016/04/01' 
AND date_stat <= '2016/04/30' 
AND ( date_stat = date_gen-1 or ( date_gen = '2016/05/01' AND date_job = '2016/04/30')) 
AND app.id_application IS NOT NULL

此查询现在需要120秒。所以我也在这个查询上运行了EXPLAIN ANALYZE，现在它没有相同的操作：

HashAggregate  (cost=375363.50..375363.51 rows=1 width=4) (actual time=186716.468..186716.532 rows=490 loops=1)
->  Nested Loop  (cost=0.00..375363.49 rows=1 width=4) (actual time=1.945..186619.404 rows=118990 loops=1)
    Join Filter: (gp.id_application = app.id_application)
    Rows Removed by Join Filter: 82222090
    ->  Nested Loop  (cost=0.00..375343.49 rows=1 width=4) (actual time=1.821..171458.237 rows=118990 loops=1)
          Join Filter: (sj.id_groupe = gp.id_groupe)
          Rows Removed by Join Filter: 954061820
          ->  Seq Scan on stat sj  (cost=0.00..375109.47 rows=1 width=8) (actual time=0.235..1964.423 rows=118990 loops=1)
                Filter: ((date_stat >= '2016-04-01'::date) AND (date_stat <= '2016-04-30'::date) AND ((date_stat = (date_gen - 1)) OR ((date_gen = '2016-05-01'::date) AND (date_stat = '2016-04-30'::date))))
                Rows Removed by Filter: 7343798
          ->  Seq Scan on groupe gp  (cost=0.00..133.45 rows=8045 width=12) (actual time=0.002..0.736 rows=8019 loops=118990)
    ->  Seq Scan on application app  (cost=0.00..11.67 rows=667 width=4) (actual time=0.003..0.073 rows=692 loops=118990)
          Filter: (id_application IS NOT NULL)
  Total runtime: 186716.635 ms

所以我决定通过减少查询条件的数量来搜索问题的来源，直到表演再次被接受为止。

仅使用此参数

WHERE date_stat >= '2016/04/01'

只需1.9秒（就像第一个工作查询一样）它还使用2个参数：

WHERE date_stat >= '2016/04/01' 
AND app.id_application IS NOT NULL

但是当我尝试添加其中一行时，我在Explain中有了嵌套循环

AND date_stat <= '2016/04/30' 
AND ( date_stat = date_gen-1 or ( date_gen = '2016/05/01' AND date_stat = '2016/04/30'))

有人知道它可能来自哪里吗？

Answer 1

好的，看起来优化器估算存在问题。他说四月只有1 row所以他选择NESTED LOOP这对于大量行（118,990来说效率非常低）。

为每张桌子执行VACUUM ANALYZE。这将清理死元组并刷新统计信息。
考虑根据dates 等CREATE INDEX date_stat_idx ON <table with date_stat> USING btree (date_stat);添加索引
重新运行查询，

使用嵌套循环 - PostgreSQL提高SQL查询的性能

1 个答案: