ORACLE数据库分页查询/翻页
最佳实践
一、示例数据:
Select Count(*) From dba_objects ;
-------------------------------------------------
COUNT(*)
6254
Create table ao as select * f rom dba_objects ;
Create index i_ao_created on ao(created) ;
Create index i_ao_id on ao(object_id) ;
Create table bo as select * from dba_objects where rownum<=100 ;
Analyze table ao compute statistics for table for all indexes ;
Analyze table bo compute statistics ;
二、求页数
无论以本文中哪种SQL写法,为了获得查询的总页数,不可避免的总要进行一次COUNT(*) 的查询过程,SQL如下:
单表:
Select count(*) into rowcount from ao where created > … and object_id between …
多表:
Select count(*) into rowcount from ao a,bo b where ao.object_id = bo.object_id and ao.object_name like …. And bo.ojbect_id > …
设单页行数为 pagesize (缺省为 20) ,则总页数 pagecount = ceil(rowcount / pagesize)
CEIL函数: 返回大于或等于给出数字的最小整数
SQL> select ceil(21.000001),ceil(1),ceil(10.1234) from dual;
CEIL(21.000001) CEIL(1) CEIL(10.1234)
--------------- ---------- -------------
22 1 11
提示:在具体项目中做分页实现时,设计上如果不要求算出总页数,则可以省去此COUNT(*) 的查询,则整个翻页性能将有所提升(因为少做了一个查询)
综上所述,对于每次的分页查询,需要输入的参数有2个:
页号: pageindex
页行数: PAGESIZE (一般为整个系统设一个常量,如前面定义的20)
则每个页面应该显示的记录为:
IF pageindex <= pagecount THEN
Firstrow = (pageindex – 1)*PAGESIZE + 1
Lastrow = pageindex * PAGESIZE
END IF;
三、分页查询实现SQL方案
方案一:(单表分页查询)
适用场景:
A、 WHERE部分条件仅仅通过索引查询就可满足所有查询过滤条件;
B、满足A条件,且有ORDER BY ,且WHERE 条件中所使用的索引,正好适合 ORDER BY 字段的要求。
SQL如下:
select /*+ ordered use_nl(v3 t) */
t.*
from (select v2.rd
from (select v1.rd, rownum rn
from (select a.rowid rd
from ao a
where a.created >= sysdate - 100
order by created) v1
where rownum <= 40) v2 -- 40 :lastrow
where rn >= 21) v3, -- 21 :firstrow
ao t
where v3.rd = t.rowid
执行计划:
0 SELECT STATEMENT Optimizer=CHOOSE
1 0 NESTED LOOPS
2 1 VIEW
3 2 COUNT (STOPKEY)
4 3 VIEW
5 4 INDEX (RANGE SCAN) OF 'I_AO_CREATED' (NON-UNIQUE)
6 1 TABLE ACCESS (BY USER ROWID) OF 'AO'
评价:该种场景下的分页SQL写法执行效率最高效,但限制较多(见适用场景)。
不适用场景:
以上述SQL为例,如果WHERE 条件部分为 a.created >= sysdate – 100 and a.object_name like ‘ABC%’ and … 即不适应于场景A条件所述,或有ORDER BY ,如:ORDER BY OBJECT_NAME ,CREATED ,即不适应于场景B条件所述,则方案一SQL不适用。
方案二:(单表/多表分页查询)
适用场景:
所有场景都适用,包括方案一所指场景。
SQL如下(单表):
select v2.*
from (select v1.*, rownum rn
from (select object_name,
object_id,
owner,
data_object_id,
created
from ao
where created >= sysdate - 100
and owner = 'SYS'
order by object_name, created) v1
where rownum <= 40) v2 -- 40 :lastrow
where rn >= 21 -- 21 :firstrow
执行计划:
0 SELECT STATEMENT Optimizer=CHOOSE
1 0 VIEW
2 1 COUNT (STOPKEY)
3 2 VIEW
4 3 SORT (ORDER BY STOPKEY)
5 4 TABLE ACCESS (BY INDEX ROWID) OF 'AO'
6 5 INDEX (RANGE SCAN) OF 'I_AO_CREATED' (NON-UNIQUE)
SQL如下(多表)
select v2.*
from (select v1.*, rownum rn
from (select ao.object_name,
ao.owner,
ao.created,
bo.object_type,
bo.last_ddl_time
from ao, bo
where ao.object_id = bo.object_id
and ao.created >= sysdate - 300
and ao.owner = 'SYS'
and bo.status = 'VALID'
order by ao.object_name, bo.created) v1
where rownum <= 40) v2 -- 40:lastrow
where rn >= 21 -- 21: firstrow
执行计划:
0 SELECT STATEMENT Optimizer=CHOOSE
1 0 VIEW
2 1 COUNT (STOPKEY)
3 2 VIEW
4 3 SORT (ORDER BY STOPKEY)
5 4 TABLE ACCESS (BY INDEX ROWID) OF 'AO'
6 5 NESTED LOOPS
7 6 TABLE ACCESS (FULL) OF 'BO'
8 6 INDEX (RANGE SCAN) OF 'I_AO_ID'
评价:以上两种(单表/多表)的分页查询写法很类似,从性能上来说都逊于方案一的做法,但这和场景有关,是没有办法的事。
综述:本文所述的两种分页方案,中心是阐述分页的实现思想,即翻页由数据库查询来完成,不占用额外的前后端会话内存,不在前端/中间层缓存数据,实现翻页时方法尽量简单、实用,需要强调的是,不论以哪种方案实现翻页,都要求查询的主体部分(SQL中粗体部分)的WHERE条件要求高效,这样整个翻页的执行速度才会快。方案一是取合理的利用了索引的特性;方案二则在主查询不变的基础上通过灵活运用ROWNUM伪列,实现查询行的定位,具体应用当中能够使用哪种方案,需具体评估场景,在之前请深刻理解本文两种方案中示例SQL的实现。
应当特别注意一个问题,在获得分页的查询SQL时,注意WHERE条件部分不应该是拼出来的,而应该是以绑定变量的方式设置条件字段值!!
SQL形如:
select v2.*
from (select v1.*, rownum rn
from (select object_name,
object_id,
owner,
data_object_id,
created
from ao
where created >= :created
and owner = :owner
order by object_name, created) v1
where rownum <= :lastrow) v2
where rn >= :firstrow