Hive的分桶详解
2018-09-26 本文已影响0人
数据萌新
Hive分桶通俗点来说就是将表（或者分区，也就是hdfs上的目录而真正的数据是存储在该目录下的文件）中文件分成几个文件去存储。比如表buck(目录，里面存放了某个文件如sz.data)文件中本来是1000000条数据，由于在处理大规模数据集时，在开发和修改查询的阶段，如果能在数据集的一小部分数据上试运行查询，会带来很多方便，所以我们可以分4个文件去存储。
下面记录了从头到尾以及出现问题的操作
进行连接，创建数据库myhive2,使用该数据库
[root@mini1 ~]# cd apps/hive/bin
[root@mini1 bin]# ./beeline
Beeline version 1.2.1 by Apache Hive
beeline> !connect jdbc:hive2://localhost:10000
Connecting to jdbc:hive2://localhost:10000
Enter username for jdbc:hive2://localhost:10000: root
Enter password for jdbc:hive2://localhost:10000: ******
Connected to: Apache Hive (version 1.2.1)
Driver: Hive JDBC (version 1.2.1)
Transaction isolation: TRANSACTION_REPEATABLE_READ
0: jdbc:hive2://localhost:10000> show databases;
+----------------+--+
| database_name |
+----------------+--+
| default |
| myhive |
+----------------+--+
2 rows selected (1.795 seconds)
0: jdbc:hive2://localhost:10000> create database myhive2;
No rows affected (0.525 seconds)
0: jdbc:hive2://localhost:10000> use myhive2;
No rows affected (0.204 seconds)
创建分桶表，导入数据，查看表内容
0: jdbc:hive2://localhost:10000> create table buck(id string,name string)
0: jdbc:hive2://localhost:10000> clustered by (id) sorted by (id) into 4 buckets
0: jdbc:hive2://localhost:10000> row format delimited fields terminated by ',';
No rows affected (0.34 seconds)
0: jdbc:hive2://localhost:10000> desc buck;
+-----------+------------+----------+--+
| col_name | data_type | comment |
+-----------+------------+----------+--+
| id | string | |
| name | string | |
+-----------+------------+----------+--+
2 rows selected (0.55 seconds)
load data local inpath '/root/sz.data' into table buck;
INFO : Loading data to table myhive2.buck from file:/root/sz.data
INFO : Table myhive2.buck stats: [numFiles=1, totalSize=91]
No rows affected (1.411 seconds)
0: jdbc:hive2://localhost:10000> select * from buck;
+----------+------------+--+
| buck.id | buck.name |
+----------+------------+--+
| 1 | zhangsan |
| 2 | lisi |
| 3 | wangwu |
| 4 | furong |
| 5 | fengjie |
| 6 | aaa |
| 7 | bbb |
| 8 | ccc |
| 9 | ddd |
| 10 | eee |
| 11 | fff |
| 12 | ggg |
+----------+------------+--+
如果分桶了的话，那么buck目录下应该有4个文件，页面查看
image.png
然而并没有，还是自己导入的那个文件。
这是因为分桶不是hive活着hadoop自动给我们划分文件来分桶的，而应该是我们分好之后导入才好。
需要设置开启分桶，设置reducetask数量（跟分桶数量一致）
0: jdbc:hive2://localhost:10000> set hive.enforce.bucketing = true;
No rows affected (0.063 seconds)
0: jdbc:hive2://localhost:10000> set hive.enforce.bucketing ;
+------------------------------+--+
| set |
+------------------------------+--+
| hive.enforce.bucketing=true |
+------------------------------+--+
1 row selected (0.067 seconds)
0: jdbc:hive2://localhost:10000> set mapreduce.job.reduces=4;
那么创建另外一个表tp,将该表数据放入到buck中（select出来insert 进去），放入的时候指定进行分桶，那么会分四桶，每个里面进行排序。那么最后buck表就进行了分桶（分桶是导入的时候就分桶的而不是自己实现分桶（文件划分））。
接下来，清空buck表信息，创建表tp，将tp中数据查询出来insert into到buck中。
0: jdbc:hive2://localhost:10000> truncate table buck;
No rows affected (0.316 seconds)
0: jdbc:hive2://localhost:10000> create table tp(id string,name string)
0: jdbc:hive2://localhost:10000> row format delimited fields terminated by ',';
No rows affected (0.112 seconds)
0: jdbc:hive2://localhost:10000> load data local inpath '/root/sz.data' into table tp;
INFO : Loading data to table myhive2.tp from file:/root/sz.data
INFO : Table myhive2.tp stats: [numFiles=1, totalSize=91]
No rows affected (0.419 seconds)
0: jdbc:hive2://localhost:10000> show tables;
+-----------+--+
| tab_name |
+-----------+--+
| buck |
| tp |
+-----------+--+
2 rows selected (0.128 seconds)
0: jdbc:hive2://localhost:10000> select * from tp;
+--------+-----------+--+
| tp.id | tp.name |
+--------+-----------+--+
| 1 | zhangsan |
| 2 | lisi |
| 3 | wangwu |
| 4 | furong |
| 5 | fengjie |
| 6 | aaa |
| 7 | bbb |
| 8 | ccc |
| 9 | ddd |
| 10 | eee |
| 11 | fff |
| 12 | ggg |
+--------+-----------+--+
12 rows selected (0.243 seconds)
0: jdbc:hive2://localhost:10000> insert into buck
0: jdbc:hive2://localhost:10000> select id,name from tp distribute by (id) sort by (id);
INFO : Number of reduce tasks determined at compile time: 4
INFO : In order to change the average load for a reducer (in bytes):
INFO : set hive.exec.reducers.bytes.per.reducer=<number>
INFO : In order to limit the maximum number of reducers:
INFO : set hive.exec.reducers.max=<number>
INFO : In order to set a constant number of reducers:
INFO : set mapreduce.job.reduces=<number>
INFO : number of splits:1
INFO : Submitting tokens for job: job_1508216103995_0028
INFO : The url to track the job: http://mini1:8088/proxy/application_1508216103995_0028/
INFO : Starting Job = job_1508216103995_0028, Tracking URL = http://mini1:8088/proxy/application_1508216103995_0028/
INFO : Kill Command = /root/apps/hadoop-2.6.4/bin/hadoop job -kill job_1508216103995_0028
INFO : Hadoop job information for Stage-1: number of mappers: 1; number of reducers: 4
INFO : 2017-10-19 03:57:23,631 Stage-1 map = 0%, reduce = 0%
INFO : 2017-10-19 03:57:29,349 Stage-1 map = 100%, reduce = 0%, Cumulative CPU 1.18 sec
INFO : 2017-10-19 03:57:40,096 Stage-1 map = 100%, reduce = 25%, Cumulative CPU 2.55 sec
INFO : 2017-10-19 03:57:41,152 Stage-1 map = 100%, reduce = 75%, Cumulative CPU 5.29 sec
INFO : 2017-10-19 03:57:42,375 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 6.61 sec
INFO : MapReduce Total cumulative CPU time: 6 seconds 610 msec
INFO : Ended Job = job_1508216103995_0028
INFO : Loading data to table myhive2.buck from hdfs://192.168.25.127:9000/user/hive/warehouse/myhive2.db/buck/.hive-staging_hive_2017-10-19_03-57-14_624_1985499545258899177-1/-ext-10000
INFO : Table myhive2.buck stats: [numFiles=4, numRows=12, totalSize=91, rawDataSize=79]
No rows affected (29.238 seconds)
0: jdbc:hive2://localhost:10000> select * from buck;
+----------+------------+--+
| buck.id | buck.name |
+----------+------------+--+
| 11 | fff |
| 4 | furong |
| 8 | ccc |
| 1 | zhangsan |
| 12 | ggg |
| 5 | fengjie |
| 9 | ddd |
| 2 | lisi |
| 6 | aaa |
| 10 | eee |
| 3 | wangwu |
| 7 | bbb |
+----------+------------+--+
到这应该就知道已经分桶了，否则id应该是1-12出来的，这是因为在4个桶中，分别进行了各自的排序，而不是跟order by一样会进行全局排序，页面查看下吧。
image.png
能看到确实分了4桶，客户端查看下内容吧（可以直接解析hdfs操作的）
0: jdbc:hive2://localhost:10000> dfs -ls /user/hive/warehouse/myhive2.db/buck;
+-----------------------------------------------------------------------------------------------------------+--+
| DFS Output |
+-----------------------------------------------------------------------------------------------------------+--+
| Found 4 items |
| -rwxr-xr-x 2 root supergroup 22 2017-10-19 03:57 /user/hive/warehouse/myhive2.db/buck/000000_0 |
| -rwxr-xr-x 2 root supergroup 34 2017-10-19 03:57 /user/hive/warehouse/myhive2.db/buck/000001_0 |
| -rwxr-xr-x 2 root supergroup 13 2017-10-19 03:57 /user/hive/warehouse/myhive2.db/buck/000002_0 |
| -rwxr-xr-x 2 root supergroup 22 2017-10-19 03:57 /user/hive/warehouse/myhive2.db/buck/000003_0 |
+-----------------------------------------------------------------------------------------------------------+--+
5 rows selected (0.028 seconds)
0: jdbc:hive2://localhost:10000> dfs -cat /user/hive/warehouse/myhive2.db/buck/000000_0;
+-------------+--+
| DFS Output |
+-------------+--+
| 11,fff |
| 4,furong |
| 8,ccc |
+-------------+--+
3 rows selected (0.02 seconds)
0: jdbc:hive2://localhost:10000> dfs -cat /user/hive/warehouse/myhive2.db/buck/000001_0;
+-------------+--+
| DFS Output |
+-------------+--+
| 1,zhangsan |
| 12,ggg |
| 5,fengjie |
| 9,ddd |
+-------------+--+
4 rows selected (0.08 seconds)
0: jdbc:hive2://localhost:10000> dfs -cat /user/hive/warehouse/myhive2.db/buck/000002_0;
+-------------+--+
| DFS Output |
+-------------+--+
| 2,lisi |
| 6,aaa |
+-------------+--+
2 rows selected (0.088 seconds)
0: jdbc:hive2://localhost:10000> dfs -cat /user/hive/warehouse/myhive2.db/buck/000003_0;
+-------------+--+
| DFS Output |
+-------------+--+
| 10,eee |
| 3,wangwu |
| 7,bbb |
+-------------+--+
3 rows selected (0.062 seconds)
注： select id,name from tp distribute by (id) sort by (id)语句中distribute by (id) sort by (id)知道根据id进行分桶（根据id进行hash散列），根据id进行排序默认升序。如果两者字段相同那么可以使用cluster by (id);也就是说可以写成
insert into buck select id ,name from p cluster by (id);
效果是一样的。
分桶的作用
观察下面的语句。
select a.id,a.name,b.addr from a join b on a.id = b.id;
如果a表和b表已经是分桶表，而且分桶的字段是id字段，那么做这个操作的时候就不需要再进行全表笛卡尔积了。
