大資料012-Hive的分桶詳解
Hive分桶通俗點來說就是將表(或者分割槽,也就是hdfs上的目錄而真正的資料是儲存在該目錄下的檔案)中檔案分成幾個檔案去儲存。比如表buck(目錄,裡面存放了某個檔案如sz.data)檔案中本來是1000000條資料,由於在處理大規模資料集時,在開發和修改查詢的階段,如果能在資料集的一小部分資料上試執行查詢,會帶來很多方便,所以我們可以分4個檔案去儲存。
下面記錄了從頭到尾以及出現問題的操作
進行連線,建立資料庫myhive2,使用該資料庫
[[email protected] ~]# cd apps/hive/bin [[email protected] 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)
建立分桶表,匯入資料,查看錶內容
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 | +----------+------------+--+
如果分桶了的話,那麼buck目錄下應該有4個檔案,頁面檢視
然而並沒有,還是自己匯入的那個檔案。
這是因為分桶不是hive活著hadoop自動給我們劃分檔案來分桶的,而應該是我們分好之後匯入才好。
需要設定開啟分桶,設定reducetask數量(跟分桶數量一致)
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;
那麼建立另外一個表tp,將該表資料放入到buck中(select出來insert 進去),放入的時候指定進行分桶,那麼會分四桶,每個裡面進行排序。那麼最後buck表就進行了分桶(分桶是匯入的時候就分桶的而不是自己實現分桶(檔案劃分))。
接下來,清空buck表資訊,建立表tp,將tp中資料查詢出來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 |
+----------+------------+--+
到這應該就知道已經分桶了,否則id應該是1-12出來的,這是因為在4個桶中,分別進行了各自的排序,而不是跟order by一樣會進行全域性排序,頁面檢視下吧。
能看到確實分了4桶,客戶端檢視下內容吧(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)語句中distribute by (id) sort by (id)知道根據id進行分桶(根據id進行hash雜湊),根據id進行排序預設升序。如果兩者欄位相同那麼可以使用cluster by (id);也就是說可以寫成
insert into buck select id ,name from p cluster by (id);
效果是一樣的。
分桶的作用
觀察下面的語句。
select a.id,a.name,b.addr from a join b on a.id = b.id;
如果a表和b表已經是分桶表,而且分桶的欄位是id欄位,那麼做這個操作的時候就不需要再進行全表笛卡爾積了。