ROW_NUMBER 的性能

Question

我有一个使用 ROW_NUMBER() 的查询。我有这样的事情：

ROW_NUMBER() OVER (ORDER BY publish_date DESC) rnum

查询运行得非常快。但是，如果我添加对“rnum”列的任何引用，查询就会变慢。所以，看起来只有 ROW_NUMBER() 不是问题，但是当我在实际查询中使用“rnum”时，它会爬行大约 30 秒。

有什么想法吗？

供参考，这里是查询：

  WITH aquire AS (
    SELECT rtnum, trans_id, source, provider, publish_date, story_link, industry_name, sector_name, subject, teaser, tickers
    FROM (SELECT d.trans_id, d.source, 'AquireMedia' AS provider,
                 d.trans_time AS publish_date, '/research/get_news.php?id=' || d.trans_id AS story_link,
                 i.name AS industry_name, s.sector_name, d.headline AS subject, NULL AS teaser,
                 NEWS.NEWS_FUNCTIONS.CONCATENATE_TICKERS(d.trans_id,'AQUIREMEDIA') AS tickers,
                 ROW_NUMBER() OVER (PARTITION BY d.trans_id ORDER BY d.trans_time DESC) as rtnum
          FROM   story_descriptions_3m d, story_tickers_3m t, uber_master_mv m, industry i, ind_sector ix, sectors s, comp_ind c
          WHERE  d.trans_id = t.trans_id
            AND  t.m_ticker = m.m_ticker
            AND  t.m_ticker = c.m_ticker(+)
            AND  c.ind_code = i.ind_code(+)
            AND  i.ind_code = ix.ind_code(+)
            AND  ix.sector_id = s.sector_id(+)  AND s.sector_id = 10 )
    WHERE rtnum = 1),
partner AS (
  SELECT rtnum, trans_id, source, provider, publish_date, story_link, industry_name, sector_name, subject, teaser, tickers
  FROM (SELECT CAST(n.story_id AS VARCHAR2(20)) trans_id, n.provider AS source, 'Partner News' AS provider,
               n.story_date AS publish_date, n.link AS story_link, i.name AS industry_name, s.sector_name, n.title AS subject,
               CAST(substr(n.teaser,1,4000) AS VARCHAR2(4000)) AS teaser, NEWS.NEWS_FUNCTIONS.CONCATENATE_TICKERS(n.story_id,'OTHER') AS tickers,
               ROW_NUMBER() OVER (PARTITION BY n.story_id ORDER BY n.story_date DESC) as rtnum
        FROM   news_stories_3m n, news_stories_lookup_3m t, comp_ind c, uber_master_mv m, industry i, ind_sector ix, sectors s
        WHERE  t.story_id = n.story_id
          AND  t.ticker   = m.ticker
          AND  m.m_ticker = c.m_ticker(+)
          AND  c.ind_code = i.ind_code(+)
          AND  i.ind_code = ix.ind_code(+)
          AND  ix.sector_id = s.sector_id(+)  AND s.sector_id = 10 )
   WHERE rtnum = 1)
  SELECT  trans_id, source, provider,
         TO_CHAR(publish_date,'MM/DD/YYYY HH24:MI:SS') AS publish_date,
         UNIX_TIMESTAMP(publish_date) AS timestamp,
         story_link, industry_name, sector_name, subject, teaser, tickers
  FROM (SELECT trans_id, source, provider, publish_date, story_link, industry_name, sector_name, subject, teaser, tickers,
                         ROW_NUMBER() OVER (ORDER BY publish_date DESC) rnum
             FROM (SELECT trans_id, source, provider, publish_date, story_link, industry_name, sector_name, subject, teaser, tickers
                        FROM   aquire WHERE rtnum <= 5
                        UNION ALL
                        SELECT trans_id, source, provider, publish_date, story_link, industry_name, sector_name, subject, teaser, tickers
                        FROM   partner WHERE rtnum <= 5)) 
WHERE rnum BETWEEN 1 AND 1 * 5;

Answer 1

让我们通过一个简单的示例来模拟您的查询，以演示和解释您遇到可预期的结果。

样本数据

create table tab1 as
select rownum id, lpad('x',3000,'y') pad from dual connect by level <= 1000000;

现在，如果您在 IDE 中运行以下查询，您将立即看到结果集的第一页。

注意，你定义了row_number但不要使用它。

select id, pad from (
 select id, pad,
  row_number() over (order by id) as rnum
 from tab1
)

答案在下面的执行计划中

--------------------------------------------------------------------------
| Id  | Operation         | Name | Rows  | Bytes | Cost (%CPU)| Time     |
--------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |      |  1000K|  2866M|   135K  (1)| 00:00:06 |
|   1 |  TABLE ACCESS FULL| TAB1 |  1000K|  2866M|   135K  (1)| 00:00:06 |
--------------------------------------------------------------------------

你看到没有执行排序和过滤，row_number 很简单忽略。

这（只获取少数初始行并且没有排序）解释了为什么查询执行。

如果您对row_number 进行如下约束，则相反

SQL> select id, pad from (
  2   select id, pad,
  3    row_number() over (order by id) as rnum
  4   from tab1
  5  ) where rnum between 1 and 5
  6  ;

Elapsed: 00:00:07.80

您观察到可观的经过时间。 执行计划再次提供了答案。

请参阅 here 如何获取 execution plan 以进行查询。

-----------------------------------------------------------------------------------------
| Id  | Operation                | Name | Rows  | Bytes |TempSpc| Cost (%CPU)| Time     |
-----------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT         |      |     5 |  7640 |       |   762K  (1)| 00:00:30 |
|*  1 |  VIEW                    |      |     5 |  7640 |       |   762K  (1)| 00:00:30 |
|*  2 |   WINDOW SORT PUSHED RANK|      |  1000K|  2866M|  3906M|   762K  (1)| 00:00:30 |
|   3 |    TABLE ACCESS FULL     | TAB1 |  1000K|  2866M|       |   135K  (1)| 00:00:06 |
-----------------------------------------------------------------------------------------
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   1 - filter("RNUM">=1 AND "RNUM"<=5)
   2 - filter(ROW_NUMBER() OVER ( ORDER BY "ID")<=5)

结果是现在您必须遍历所有记录（在您的情况下执行所有连接），这会破坏性能。

为了证明这一点，请使用fetch all 选项或添加order by 子句运行简单的执行 查询。您很可能会得到与第二个查询相同的不良结果。

最后的评论

您可以使用row_limiting_clause 而不是ROW_NUMBER()

在order by 子句中传递来自row_number 的排序列，并使用offset 和fetch first 来限制结果。

select id, pad from (
 select id, pad
 from tab1
) order by id
fetch first 5 rows only;

在封面下，您将看到与上面使用 WINDOW SORT PUSHED RANK 相同的执行计划。