order by
1、order by会对输入按照指定字段做全局排序,输出结果有序,因此只有一个reducer(多个reducer无法保证全局排序,手工设定reduce数量无效);
只有一个reducer会导致当输入规模较大时,需要较长的计算时间,速度很非常慢;在数据量大的情况下慎用order by;
2、hive.mapred.mode(默认值是nonstrict)对order by的影响
1)当hive.mapred.mode=nonstrict时,order by和关系型数据库中的order by功能一致,按照指定的某一列或多列排序输出;
2)当hive.mapred.mode=strict时,order by必须要使用limit,否则执行会报错;
set hive.mapred.mode=strict; select * from emp order by empno desc;
FAILED: SemanticException 1:27 In strict mode, if ORDER BY is specified, LIMIT must also be specified. Error encountered near token 'empno'
3)当hive.mapred.mode=strict时,分区表必须要指定partition进行查询,否则执行会报错;
FAILED: SemanticException [Error 10041]: No partition predicate found for Alias "xxx" Table "xxxx";
3、select * from emp where empno > 7782 order by empno desc;的执行过程
从表中读取数据,执行where条件,然后以empno的值做为key,其他列值做为value,然后把数据传到同一个reduce中,根据需要的排序方式进行排序;
注意:where是在map阶段执行的,而非reduce阶段执行。
sort by
1、使用sort by可以指定reduce的个数:set mapred.reduce.tasks=10; 对输出的数据在进行归并排序即可得到全部结果;
2、sort by非分区表不受hive.mapred.mode是否是strict还是nostrict的影响(不需要强制使用limit限制);如果是分区表,必须指定partition字段;
3、sort by的数据只能保证在同一个reduce中的数据可以按照指定字段排序;
4、可以用limit子句大大减少数据量。使用limit n后,传输到reduce端的数据记录数就减少到n* (map个数);
set mapred.reduce.tasks = 3; select * from emp sort by empno;
7654 MARTIN SALESMAN 7698 1981-9-28 1250.0 1400.0 30 7698 BLAKE MANAGER 7839 1981-5-1 2850.0 NULL 30 7782 CLARK MANAGER 7839 1981-6-9 2450.0 NULL 10 7788 SCOTT ANALYST 7566 1987-4-19 3000.0 NULL 20 7839 KING PRESIDENT NULL 1981-11-17 5000.0 NULL 10 7844 TURNER SALESMAN 7698 1981-9-8 1500.0 0.0 30 7499 ALLEN SALESMAN 7698 1981-2-20 1600.0 300.0 30 7521 WARD SALESMAN 7698 1981-2-22 1250.0 500.0 30 7566 JONES MANAGER 7839 1981-4-2 2975.0 NULL 20 7876 ADAMS CLERK 7788 1987-5-23 1100.0 NULL 20 7900 JAMES CLERK 7698 1981-12-3 950.0 NULL 30 7934 MILLER CLERK 7782 1982-1-23 1300.0 NULL 10 7369 SMITH CLERK 7902 1980-12-17 800.0 NULL 20 7902 FORD ANALYST 7566 1981-12-3 3000.0 NULL 20
把上面的结果写到文件中再观察:每个文件中都是按照empno的升序排列的
set mapred.reduce.tasks = 3; insert overwrite local directory '/home/spark/data' select * from emp sort by empno; cd /home/spark/data ls 000000_0 000001_0 000002_0
more 000000_0 7654MARTINSALESMAN76981981-9-281250.01400.030 7698BLAKEMANAGER78391981-5-12850.0\N30 7782CLARKMANAGER78391981-6-92450.0\N10 7788SCOTTANALYST75661987-4-193000.0\N20 7839KINGPRESIDENT\N1981-11-175000.0\N10 7844TURNERSALESMAN76981981-9-81500.00.030
more 000001_0 7499ALLENSALESMAN76981981-2-201600.0300.030 7521WARDSALESMAN76981981-2-221250.0500.030 7566JONESMANAGER78391981-4-22975.0\N20 7876ADAMSCLERK77881987-5-231100.0\N20 7900JAMESCLERK76981981-12-3950.0\N30 7934MILLERCLERK77821982-1-231300.0\N10
more 000002_0 7369SMITHCLERK79021980-12-17800.0\N20 7902FORDANALYST75661981-12-33000.0\N20
可见每个reduce内部的数据是经过排序的。
distribute by
1、按照指定的字段把数据分散到不同的reduce文件中(可以指定map到reduce端分发的key,这样可以充分利用hadoop资源,在多个reduce中局部按需要排序的字段进行排序);可以设定reduce个数;
set mapred.reduce.tasks = 3; insert overwrite local directory '/home/spark/data' select * from emp distribute by length(ename) sort by empno; cd /home/spark/data ls 000000_0 000001_0 000002_0
more 000000_0 7654MARTINSALESMAN76981981-9-281250.01400.030 7844TURNERSALESMAN76981981-9-81500.00.030 7934MILLERCLERK77821982-1-231300.0\N10
more 000001_0 7521WARDSALESMAN76981981-2-221250.0500.030 7839KINGPRESIDENT\N1981-11-175000.0\N10 7902FORDANALYST75661981-12-33000.0\N20
more 000002_0 7369SMITHCLERK79021980-12-17800.0\N20 7499ALLENSALESMAN76981981-2-201600.0300.030 7566JONESMANAGER78391981-4-22975.0\N20 7698BLAKEMANAGER78391981-5-12850.0\N30 7782CLARKMANAGER78391981-6-92450.0\N10 7788SCOTTANALYST75661987-4-193000.0\N20 7876ADAMSCLERK77881987-5-231100.0\N20 7900JAMESCLERK76981981-12-3950.0\N30
length是内建函数,也可以指定其他的函数或者使用UDF;
2、distribute by与sort by连用:按照ename指定到reduce,每个reduce中按照ename升序排列;
set mapred.reduce.tasks = 3; insert overwrite local directory '/home/spark/data' select * from emp distribute by ename sort by ename; cd /home/spark/data ls 000000_0 000001_0 000002_0
more 000000_0 7698BLAKEMANAGER78391981-5-12850.0\N30 7839KINGPRESIDENT\N1981-11-175000.0\N10
more 000001_0 7876ADAMSCLERK77881987-5-231100.0\N20 7499ALLENSALESMAN76981981-2-201600.0300.030 7654MARTINSALESMAN76981981-9-281250.01400.030 7934MILLERCLERK77821982-1-231300.0\N10 7788SCOTTANALYST75661987-4-193000.0\N20 7844TURNERSALESMAN76981981-9-81500.00.030
more 000002_0 7782CLARKMANAGER78391981-6-92450.0\N10 7902FORDANALYST75661981-12-33000.0\N20 7900JAMESCLERK76981981-12-3950.0\N30 7566JONESMANAGER78391981-4-22975.0\N20 7369SMITHCLERK79021980-12-17800.0\N20 7521WARDSALESMAN76981981-2-221250.0500.030
3、如果sort by只有一个reduce时,和order by输出结果一致;
4、select * from emp where empno>7782 distribute by ename sort by ename;执行过程
从表中读取数据,执行where条件;以distribute by列的值做为key,其他列值做为value,然后把数据根据key值传到不同的reduce,然后按照sort by字段进行排序;
cluster by
1、cluster by除了具有distribute by的功能外还兼备sort by功能;
2、当distribute by col1与sort by col1连用,且跟随的字段相同时,可使用cluster by简写;
select * from emp cluster by ename;
Hive排序总结
1、在hive中进行字段排序统计过程中,使用ORDER BY是全局排序,hive只能通过一个reduce进行排序,效率很低;
2、sort by实现部分排序,单个reduce输出的结果是有序的、效率高,通常与distribute by关键字一起使用,distribute by关键字可以指定map到reduce端的分发key, 这样可以充分利用hadoop资源, 在多个reduce中局部按需要排序的字段进行排序;
3、cluster by col1等同于distributed by col1与sort by col1组合。