问题描述
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，则查询会变慢为爬行。因此，似乎只有 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？
作为参考，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,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,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,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,TO_CHAR(publish_date,'MM/DD/YYYY HH24:MI:SS') AS publish_date,UNIX_TIMESTAMP(publish_date) AS timestamp,tickers
FROM (SELECT trans_id,tickers,ROW_NUMBER() OVER (ORDER BY publish_date DESC) rnum
FROM (SELECT trans_id,tickers
FROM aquire WHERE rtnum <= 5
UNION ALL
SELECT trans_id,tickers
FROM partner WHERE rtnum <= 5))
WHERE rnum BETWEEN 1 AND 1 * 5;
解决方法
让我们在一个简单的示例上模拟您的查询，以演示和解释您遇到可预期的结果。
示例数据
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 中运行下面的查询，您将立即看到结果集的第一页。
注意，您定义了 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 |
--------------------------------------------------------------------------
您看到没有执行排序和过滤，row_number 很简单被忽略。
这（只获取少数初始行且不进行排序）解释了为什么查询执行。
相反，如果您对 row_number 进行如下约束
SQL> select id,pad from (
2 select id,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
您观察到了可观的经过时间。 执行计划再次提供了答案。
请参阅 here 如何为您的查询获取 execution plan。
-----------------------------------------------------------------------------------------
| 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。
要证明这一点，请使用 fetch all 选项或添加的 order by 子句运行简单的 performat 查询。您很可能会得到与第二个查询相同的性能不佳结果。
最后评论
您可以使用 row_limiting_clause 代替 ROW_NUMBER()
从 row_number 子句中的 order by 传递排序列，并使用 offset 和 fetch first 来限制结果。
select id,pad
from tab1
) order by id
fetch first 5 rows only;
在封面下，您将看到使用上述 WINDOW SORT PUSHED RANK 的相同执行计划。