Oracle数据库9i、10g里列转行各种方法比较

把字符串 '1,2,3,4,5'，以逗号分隔，输出为行，也就是

1
2
3
4
5

随手写了一个, Oracle 10G 以上
有表如下

SQL> select * from t;
        ID NAME
---------- ----------------
         1 0,1,5,2,8,10
         2 9,7,8
         3 你好,他好,大家好

with vmaxnum as (
     select  rownum ele
     from    dual
     connect by rownum<=(select max(length(name)-length(replace(name,',',null)))+1 from t))
select id,
       decode(pos,0,substr(name,lagpos+1),substr(name,lagpos+1,pos-lagpos-1)) name
from   (select id,name,ele,pos,nvl(lag(pos) over (partition by id order by ele),0) lagpos
        from   (select id,name,ele,
                       instr(name,',',1,ele) pos
                from   (select /*+ all_rows no_merge(v2) use_merge(t,v2) */
                               t.id,t.name,v2.ele
                        from   t,vmaxnum v2
                        where  ele<=length(name)-length(replace(name,',',null))+1)))
order  by id,ele;

        ID NAME
---------- --------
         1 0
         1 1
         1 5
         1 2
         1 8
         1 10
         2 9
         2 7
         2 8
         3 你好
         3 他好
         3 大家好

12 rows selected.

然后客户那反馈说数据库说9i的，报错说connect by不能使用子查询。
于是再稍做修改，在确定逗号不会超过99个的前提下，直接写个常量100，可以支持9i
 

with   vmaxnum as (
       select  rownum ele
       from    dual
       connect by rownum<=100)
select id,
       decode(pos,0,substr(name,lagpos+1),substr(name,lagpos+1,pos-lagpos-1)) name
from   (select id,name,ele,pos,nvl(lag(pos) over (partition by id order by ele),0) lagpos
        from   (select id,name,ele,
                       instr(name,',',1,ele) pos
                from   (select /*+ all_rows no_merge(v2) use_merge(t,v2) */
                               t.*,v2.ele
                        from   t,vmaxnum v2
                        where  ele<=length(name)-length(replace(name,',',null))+1)))
order  by id,ele;

后来想想这样也不好，于是建议建立一张特定的IOT表，保存10000个数字，一般够用了。
建立一张IOT,共10000行，同样支持9i

create table tnumber(ele,constraint pk_tnumber primary key(ele)) organization index as
select rownum id from dual connect by rownum<=10000;

然后不再需要常量100，即保证准确，又保证较好的性能

create table t500 nologging as
with   vmaxnum as (
       select  ele
       from    tnumber
       where   ele<=(select max(length(name)-length(replace(name,',',null)))+1 from t))
select id,
       decode(pos,0,substr(name,lagpos+1),substr(name,lagpos+1,pos-lagpos-1)) name
from   (select id,name,ele,pos,nvl(lag(pos) over (partition by id order by ele),0) lagpos
        from   (select id,name,ele,
                       instr(name,',',1,ele) pos
                from   (select /*+ all_rows no_merge(v2) use_merge(t,v2) */
                               t.id,t.name,v2.ele
                        from   t,vmaxnum v2
                        where  ele<=length(name)-length(replace(name,',',null))+1)))
order  by id,ele;

        ID NAME
---------- --------
         1 0
         1 1
         1 5
         1 2
         1 8
         1 10
         2 9
         2 7
         2 8
         3 你好
         3 他好
         3 大家好

12 rows selected.

下面测试一下性能：转换50万行试试，测试环境：DELL D630 用了4年半的旧笔记本，磁盘都是碎片
SQL> insert into t select rownum+4,'1,2,3,4' from dual connect by rownum<=500000;

500000 rows created.

SQL> commit;

Commit complete.

set timi on
SQL> set timi on
SQL> create table t500 nologging as
     with   vmaxnum as (
            select  ele
            from    tnumber
            where   ele<=(select max(length(name)-length(replace(name,',',null)))+1 from t))
     select id,
            decode(pos,0,substr(name,lagpos+1),substr(name,lagpos+1,pos-lagpos-1)) name
     from   (select id,name,ele,pos,nvl(lag(pos) over (partition by id order by ele),0) lagpos
             from   (select id,name,ele,
                            instr(name,',',1,ele) pos
                     from   (select /*+ all_rows no_merge(v2) use_merge(t,v2) */
                                    t.id,t.name,v2.ele
                             from   t,vmaxnum v2
                             where  ele<=length(name)-length(replace(name,',',null))+1)))
     order  by id,ele;

Table created.

Elapsed: 00:00:16.48

耗时16秒，含建表的写盘时间

SQL> select count(*) from t500;

  COUNT(*)
----------
   2000012

当然还有别的写法，比如简单的可以这样

select  id,substr(name,instr(name,',',1,rownum)+1,instr(name,',',1,rownum+1)-instr(name,',',1,rownum)-1) name
from    (select  id,','||name||',' name from t)
connect by rownum<length(translate(name,','||name,','));

或者10G以上用正则表达式也可以实现
这个方法可能不是最高效的，但也还可以。
这个SQL的性能和字符串含有的逗号的个数有关，逗号越多，也就是分隔项越多，性能越差。