本文描述了在oracle 11g release 2 版本中新增的listagg函数,listagg是一个实现字符串聚合的oracle内建函数。作为一种普遍的技术,网络上也有多种实现字符串聚合的方法。本文会首先介绍listagg函数,最后会拿这些方法与listagg进行性能方面的对比。
样例数据
本文的例子将使用如下的样例数据:
DEPTNO ENAME HIREDATE
10 CLARK 09/06/1981
10 KING 17/11/1981
10 MILLER 23/01/1982
20 ADAMS 12/01/1983
20 FORD 03/12/1981
20 JONES 02/04/1981
20 SCOTT 09/12/1982
20 SMITH 17/12/1980
30 ALLEN 20/02/1981
30 BLAKE 01/05/1981
30 JAMES 03/12/1981
30 MARTIN 28/09/1981
30 TURNER 08/09/1981
30 WARD 22/02/1981
字符串聚合
字符串聚合就是按照分组把多行数据串联成一行,以下面的结果集为例:
DEPTNO ENAME
10 CLARK
10 KING
10 MILLER
20 ADAMS
20 FORD
20 JONES
按照DEPTNO字段分组,对结果集进行字符串聚合,结果如下:
DEPTNO AGGREGATED_ENAMES
10 CLARK,KING,MILLER
20 ADAMS,FORD,JONES
可以看到,employee names基于deptno分组实现了串联,如前所述,有很多种方法实现聚合功能(文章最后提供相关链接),但是listagg更为简单,易用。
listagg 语法概述
listagg函数的语法结构如下:
LISTAGG( [,]) WITHIN GROUP (ORDER BY ) [OVER (PARTITION BY )]
listagg虽然是聚合函数,但可以提供分析功能(比如可选的OVER()子句)。使用listagg中,下列中的元素是必须的:
需要聚合的列或者表达式
WITH GROUP 关键词
分组中的ORDER BY子句
下面将演示listagg函数使用的例子
listagg 作为聚合函数
下面以EMP表为例,按照部门分组聚合employee name,并以,为分隔符。
SQL> SELECT deptno,
LISTAGG(ename, ',') WITHIN GROUP (ORDER BY ename) AS employees
FROM emp
GROUP BY deptno;
DEPTNO EMPLOYEES
---------- ------------------------------------------------------------
10 CLARK,KING,MILLER
20 ADAMS,FORD,JONES,SCOTT,SMITH 30 ALLEN,BLAKE,JAMES,MARTIN,TURNER,WARD
3 rows selected.
注:在每个聚合元素中,本例选用empolyee name字段进行排序,不过需要说明的是,在其它实现字符串聚合方法中,排序可是重量级的任务。
下面的例子中,empolyee name的聚合将按照hire date来排序。
SQL> SELECT deptno
, LISTAGG(ename, ',') WITHIN GROUP (ORDER BY hiredate) AS employees
FROM emp
GROUP BY deptno;
DEPTNO EMPLOYEES
---------- ------------------------------------------------------------
10 CLARK,KING,MILLER
20 SMITH,JONES,FORD,SCOTT,ADAMS 30 ALLEN,WARD,BLAKE,TURNER,MARTIN,JAMES
3 rows selected.
可以看到,每组中empolyee names的排序与前面的例子截然不同。
listagg作为分析函数
与许多的聚合函数类似,listagg通过加上over()子句可以实现分析功能,下面的例子将展示分析功能:
SQL> SELECT deptno
, ename, hiredate,
LISTAGG(ename, ',') WITHIN GROUP (ORDER BY hiredate) OVER (PARTITION BY deptno) AS employees
FROM emp;
DEPTNO ENAME HIREDATE EMPLOYEES
---------- ---------- ----------- -------------------------------------
10 CLARK 09/06/1981 CLARK,KING,MILLER
10 KING 17/11/1981 CLARK,KING,MILLER
10 MILLER 23/01/1982 CLARK,KING,MILLER
20 SMITH 17/12/1980 SMITH,JONES,FORD,SCOTT,ADAMS
20 JONES 02/04/1981 SMITH,JONES,FORD,SCOTT,ADAMS
20 FORD 03/12/1981 SMITH,JONES,FORD,SCOTT,ADAMS
20 SCOTT 19/04/1987 SMITH,JONES,FORD,SCOTT,ADAMS
20 ADAMS 23/05/1987 SMITH,JONES,FORD,SCOTT,ADAMS
30 ALLEN 20/02/1981 ALLEN,WARD,BLAKE,TURNER,MARTIN,JAMES
30 WARD 22/02/1981 ALLEN,WARD,BLAKE,TURNER,MARTIN,JAMES
30 BLAKE 01/05/1981 ALLEN,WARD,BLAKE,TURNER,MARTIN,JAMES
30 TURNER 08/09/1981 ALLEN,WARD,BLAKE,TURNER,MARTIN,JAMES
30 MARTIN 28/09/1981 ALLEN,WARD,BLAKE,TURNER,MARTIN,JAMES
30 JAMES 03/12/1981 ALLEN,WARD,BLAKE,TURNER,MARTIN,JAMES
14 rows selected.
切记:分析函数不会丢失结果集的每一行,而字符串的聚合却并非如此。
排序
如前所述,ORDER BY 子句是必选项,如下例所示:
SQL> SELECT deptno 2 , LISTAGG(ename, ',') WITHIN GROUP () AS employees 3 FROM emp 4 GROUP BY 5 deptno;
, LISTAGG(ename, ',') WITHIN GROUP () AS employees * ERROR at line 2: ORA-30491: missing ORDER BY clause
如果所要聚合字段的排序无关紧要,那么可以可以使用NULL代替:
SQL> SELECT deptno 2 , LISTAGG(ename, ',') WITHIN GROUP (ORDER BY NULL) AS employees 3 FROM emp 4 GROUP BY 5 deptno;
DEPTNO EMPLOYEES ---------- ------------------------------------------------------------ 10 CLARK,KING,MILLER 20 ADAMS,FORD,JONES,SCOTT,SMITH 30 ALLEN,BLAKE,JAMES,MARTIN,TURNER,WARD 3 rows selected.
在这个例子当中,虽然使用的是NULL来进行排序,但结果集中聚合的元素还是按字母的顺序排序的,这是因为listagg的默认排序行为。
分隔符
在字符串的聚合中,可以使用静态变量或者表达式作为分隔符,事实上,分隔符是可选项,例如:
SQL> SELECT deptno 2 , LISTAGG(ename) WITHIN GROUP (ORDER BY ename) AS employees 3 FROM emp 4 GROUP BY 5 deptno;
DEPTNO EMPLOYEES ---------- ------------------------------------------------------------ 10 CLARKKINGMILLER 20 ADAMSFORDJONESSCOTTSMITH 30 ALLENBLAKEJAMESMARTINTURNERWARD 3 rows selected.
唯一的限制是,分隔符要么是静态变量(如字母),要么是建立在分组字段上的确定性表达式,比如,不能使用ROWNUM作为分隔符,如下所示:
SQL> SELECT deptno 2 , LISTAGG(ename, '(' || ROWNUM || ')') 3 WITHIN GROUP (ORDER BY hiredate) AS employees 4 FROM emp 5 GROUP BY 6 deptno;
, LISTAGG(ename, '(' || ROWNUM || ')') * ERROR at line 2: ORA-30497: Argument should be a constant or a function of expressions in GROUP BY.
-- 错误信息非常清楚:ROWNUM既不是静态变量,也不是建立在分组字段上的表达式,如果使用了分组字段,那就限制了表达式的类型,例如:
SQL> SELECT deptno 2 , LISTAGG(ename, '(' || MAX(deptno) || ')') 3 WITHIN GROUP (ORDER BY hiredate) AS employees 4 FROM emp 5 GROUP BY 6 deptno;
, LISTAGG(ename, '(' || MAX(deptno) || ')') * ERROR at line 2: ORA-30496: Argument should be a constant.
-- 这个例子当中,oracle分析到分隔符试图使用分组字段,但是是一个非法的表达式,下面的例子中,使用了oracle接受的确定性表达式:
SQL> SELECT deptno 2 , LISTAGG(ename, '(' || CHR(deptno+55) || '); ') 3 WITHIN GROUP (ORDER BY hiredate) AS employees 4 FROM emp 5 GROUP BY 6 deptno;
DEPTNO EMPLOYEES ---------- ------------------------------------------------------------ 10 CLARK(A); KING(A); MILLER 20 SMITH(K); JONES(K); FORD(K); SCOTT(K); ADAMS 30 ALLEN(U); WARD(U); BLAKE(U); TURNER(U); MARTIN(U); JAMES 3 rows selected.
-- 这里把DETPNO转化成ASCII字符作为分隔符,这个一个在分组列上的确定性表达式。
其它限制listagg聚合的结果列大小限制在varchar2类型的最大值内(比如4000),例如:
SQL> SELECT LISTAGG(object_name) WITHIN GROUP (ORDER BY NULL) 2 FROM all_objects;
FROM all_objects * ERROR at line 2: ORA-01489: result of string concatenation is too long
--这里没有clob或者更大的varchar2类型类代替,所以更大的字符串必须使用替代方案(比如COLLECTION或者用户自定义的PL/SQL函数)
性能方面
下面将比较几种常用的字符串聚合方法的性能,类比的有:
LISTAGG (11g Release 2);
COLLECT + PL/SQL function(10g);
Oracle Data Cartridge – user-defined aggregate function (9i)
MODEL SQL (10g).
这里最主要的不同是listagg是一个内建函数,所以其至少与其它方案有可比性。
建立环境
为了性能比较,下面将建立一张有2000个分组,100万行数据的表,具体如下:
SQL> CREATE TABLE t 2 AS 3 SELECT ROWNUM AS id 4 , MOD(ROWNUM,2000) AS grp 5 , DBMS_RANDOM.STRING(‘u’,5) AS val 6 , DBMS_RANDOM.STRING(‘u’,30) AS pad 7 FROM dual 8 CONNECT BY ROWNUM <= 1000000;
Table created.
SQL> SELECT COUNT() FROM t;
COUNT() ———- 1000000 1 row selected.
SQL> exec DBMS_STATS.GATHER_TABLE_STATS(USER, ‘T’);
PL/SQL procedure successfully completed.
这里将使用wall-clock和autotrace进行性能方面的类比。注:样例的数据已被缓存,准备环境如下:
SQL> set autotrace traceonly statistics SQL> set timing on SQL> set arrays 500
listagg
第一个测试的是listagg,下面将对2000个分组进行聚合,并按照value排序:
SQL> SELECT grp 2 , LISTAGG(val, ‘,’) WITHIN GROUP (ORDER BY val) AS vals 3 FROM t 4 GROUP BY 5 grp;
2000 rows selected. Elapsed: 00:00:05.85 Statistics ———————————————————- 1 recursive calls 0 db block gets 7092 consistent gets 0 physical reads 0 redo size 6039067 bytes sent via SQLNet to client 552 bytes received via SQLNet from client 5 SQL*Net roundtrips to/from client 1 sorts (memory) 0 sorts (disk) 2000 rows processed
测试机上,这条语句执行了不到6秒,没有磁盘物理I/O,所有的sorting都在内存当中。
stragg/wm_concat
下面将使用广为流传的字符串聚合,Tom Kyte的定义的聚合函数STRAGG。在 oracle的10g版本中,oracle在WMSYS的用户下实现了类似功能的函数,这里直接使用这个函数来测试。注:STRAGG函数不支持字符串的排序。
SQL> SELECT grp 2 , WMSYS.WM_CONCAT(val) AS vals –<– WM_CONCAT ~= STRAGG 3 FROM t 4 GROUP BY 5 grp;
2000 rows selected. Elapsed: 00:00:19.45 Statistics ———————————————————- 1 recursive calls 0 db block gets 7206 consistent gets 0 physical reads 0 redo size 6039067 bytes sent via SQLNet to client 552 bytes received via SQLNet from client 5 SQL*Net roundtrips to/from client 1 sorts (memory) 0 sorts (disk) 2000 rows processed
这个方法花费了三倍于listagg的时间(没有排序),用户自定义的函数会比这个PL/SQL函数效率更低(比如:上下文切换)
collect(without ordering)
当10g发布的时候,我就立即使用collect函数和一个“collection-to-string”PL/SQL函数来替代STRAGG。不过10g版本中的collect没有排序功能。注;To_STRING的源码可以在相关文档中查到。
SQL> SELECT grp 2 , TO_STRING( 3 CAST(COLLECT(val) AS varchar2_ntt) 4 ) AS vals 5 FROM t 6 GROUP BY 7 grp;
2000 rows selected. Elapsed: 00:00:02.90 Statistics ———————————————————- 10 recursive calls 0 db block gets 7197 consistent gets 0 physical reads 0 redo size 6039067 bytes sent via SQLNet to client 552 bytes received via SQLNet from client 5 SQLNet roundtrips to/from client 1 sorts (memory) 0 sorts (disk) 2000 rows processed
没有排序的情况下,collect/TO_STRING方法比listagg快了两倍,但是listagg花费了大量的时间在排序上,如果说排序时无关紧要的,那么可以说collect技术是最快的。
collect (with ordering)
公平起见,在collect中引入排序(11g中的一个新特性),如下;
SQL> SELECT grp 2 , TO_STRING( 3 CAST(COLLECT(val ORDER BY val) AS varchar2_ntt) 4 ) AS vals 5 FROM t 6 GROUP BY 7 grp;
2000 rows selected. Elapsed: 00:00:07.08 Statistics ———————————————————- 10 recursive calls 0 db block gets 7197 consistent gets 0 physical reads 0 redo size 6039067 bytes sent via SQLNet to client 552 bytes received via SQLNet from client 5 SQLNet roundtrips to/from client 1 sorts (memory) 0 sorts (disk) 2000 rows processed 这次,引入了排序后,collect方法确实比listagg慢多了。
model
最后一个性能比较是与使用了model子句的实现方法。下面的例子的源代码来自于Rob van Wijk’s About Oracle blog 并做了些修改以适应样例数据。
SQL> SELECT grp 2 , vals 3 FROM ( 4 SELECT grp 5 , RTRIM(vals, ‘,’) AS vals 6 , rn 7 FROM t 8 MODEL 9 PARTITION BY (grp) 10 DIMENSION BY (ROW_NUMBER() OVER (PARTITION BY grp ORDER BY val) AS rn) 11 MEASURES (CAST(val AS VARCHAR2(4000)) AS vals) 12 RULES 13 ( vals[ANY] ORDER BY rn DESC = vals[CV()] || ‘,’ || vals[CV()+1] 14 ) 15 ) 16 WHERE rn = 1 17 ORDER BY 18 grp;
2000 rows selected. Elapsed: 00:03:28.15 Statistics ———————————————————- 3991 recursive calls 0 db block gets 7092 consistent gets 494791 physical reads 0 redo size 6039067 bytes sent via SQLNet to client 553 bytes received via SQLNet from client 5 SQL*Net roundtrips to/from client 130 sorts (memory) 0 sorts (disk) 2000 rows processed
这个例子执行了三分钟,统计信息显示发生了大量的I/O读,递归调用和内存排序,事实上,这个糟糕的表现主要是由于在查询中,大量的对临时表空间的读和写(尽管统计信息并未显示磁盘排序)。
MODEL字符串聚合方法的执行计划如下:
————————————————————– | Id | Operation | Name | Rows | Bytes |TempSpc| ————————————————————– | 0 | SELECT STATEMENT | | | | | | 1 | SORT ORDER BY | | 1000K| 1934M| 1953M| |* 2 | VIEW | | 1000K| 1934M| | | 3 | SQL MODEL ORDERED | | 1000K| 9765K| | | 4 | WINDOW SORT | | 1000K| 9765K| 19M| | 5 | TABLE ACCESS FULL| T | 1000K| 9765K| | ————————————————————– Predicate Information (identified by operation id): ————————————————— 2 – filter(“RN”=1) 通过SQL的监控报告(使用DBMS_SQLTUNE.REPORT_SQL_MONITOR)在SQL MODEL操作的第三步中,数据的排序使用4Gb的临时空间,在Gary Myers’ Sydney Oracle Lab blog中也阐述了这个现象。
性能总结
从以上事例中可以看出,listagg函数是字符串聚合方法中效率最高的一个,并且还是一个内建的函数。如果不需要排序的情况下,collect会更快一些,但如果是需要排序的话,listagg绝对是最快的。
深度阅读
更多关于listagg的信息,详见online sql reference。