在Hive上实现SCD

一、问题提出
        官方一直称Hive是Hadoop数据仓库解决方案。既然是数据仓库就离不开多维、CDC、SCD这些概念，于是尝试了一把在Hive上实现SCD1和SCD2。这有两个关键点，一个是行级更新，一个是生成代理键。行级更新hive本身就是支持的，但需要一些配置，还有一些限制。具体可参考http://blog.csdn.net/wzy0623/article/details/51483674。生成代理键在RDBMS上一般都用自增序列。Hive也有一些对自增序列的支持，本实验分别使用了窗口函数ROW_NUMBER()和hive自带的UDFRowSequence实现生成代理键。

二、软件版本
Hadoop 2.7.2
Hive 2.0.0

三、实验步骤
1. 准备初始数据文件a.txt，内容如下：

1,张三,US,CA2,李四,US,CB3,王五,CA,BB4,赵六,CA,BC5,老刘,AA,AA

2. 用ROW_NUMBER()方法实现初始装载和定期装载

（1）建立初始装载脚本init_row_number.sql，内容如下：

USE test;-- 建立过渡表DROP TABLE IF EXISTS tbl_stg;CREATE TABLE tbl_stg (    id INT,    name STRING,    cty STRING,    st STRING)ROW FORMAT DELIMITED FIELDS TERMINATED BY ',';-- 建立维度表DROP TABLE IF EXISTS tbl_dim;CREATE TABLE tbl_dim (    sk INT,    id INT,    name STRING,    cty STRING,    st STRING,    version INT,    effective_date DATE,    expiry_date DATE)CLUSTERED BY (id) INTO 8 BUCKETSSTORED AS ORC TBLPROPERTIES ('transactional'='true');-- 向过渡表加载初始数据LOAD DATA LOCAL INPATH '/home/grid/BigDataDWTest/a.txt' INTO TABLE tbl_stg;-- 向维度表装载初始数据INSERT INTO tbl_dimSELECT    ROW_NUMBER() OVER (ORDER BY tbl_stg.id) + t2.sk_max,    tbl_stg.*,    1,    CAST('1900-01-01' AS DATE),    CAST('2200-01-01' AS DATE)from tbl_stg CROSS JOIN (SELECT COALESCE(MAX(sk),0) sk_max FROM tbl_dim) t2;

（2）执行初始装载

hive -S -f /home/grid/BigDataDWTest/init_row_number.sql

（3）修改数据文件a.txt，内容如下：

1,张,U,C3,王五,CA,BB4,赵六,AC,CB5,刘,AA,AA6,老杨,DD,DD

说明：
1. 新增了第6条数据
2. 删除了第2条数据
3. 修改了第1条数据的name列、cty列和st列（name列按SCD2处理，cty列和st列按SCD1处理）
4. 修改了第4条数据的cty列和st列（按SCD1处理）
5. 修改了第5条数据的name列（按SCD2处理）

（4）建立定期装载脚本scd_row_number.sql，内容如下：

USE test;-- 设置日期变量SET hivevar:pre_date = DATE_ADD(CURRENT_DATE(),-1);SET hivevar:max_date = CAST('2200-01-01' AS DATE);-- 向过渡表加载更新后的数据LOAD DATA LOCAL INPATH '/home/grid/BigDataDWTest/a.txt' OVERWRITE INTO TABLE tbl_stg;-- 向维度表装载更新后的数据-- 设置已删除记录和SCD2的过期UPDATE tbl_dimSET expiry_date = ${hivevar:pre_date}WHERE sk IN(SELECT a.sk FROM (SELECT sk,id,name FROM tbl_dim WHERE expiry_date = ${hivevar:max_date}) a LEFT JOIN tbl_stg b ON a.id=b.idWHERE b.id IS NULL OR a.name<>b.name);-- 处理SCD2新增行INSERT INTO tbl_dimSELECT    ROW_NUMBER() OVER (ORDER BY t1.id) + t2.sk_max,    t1.id,    t1.name,    t1.cty,    t1.st,    t1.version,    t1.effective_date,    t1.expiry_dateFROM(SELECT    t2.id id,    t2.name name,    t2.cty cty,    t2.st st,    t1.version + 1 version,    ${hivevar:pre_date} effective_date,    ${hivevar:max_date} expiry_dateFROM tbl_dim t1 INNER JOIN tbl_stg t2ON t1.id=t2.id AND t1.name<>t2.name AND t1.expiry_date = ${hivevar:pre_date}LEFT JOIN tbl_dim t3 ON T1.id = t3.id AND t3.expiry_date = ${hivevar:max_date}WHERE t3.sk IS NULL) t1CROSS JOIN(SELECT COALESCE(MAX(sk),0) sk_max FROM tbl_dim) t2;-- 处理SCD1-- 因为hive的update还不支持子查询，所以这里使用了一个临时表存储需要更新的记录，用先delete再insert代替update-- 因为SCD1本身就不保存历史数据，所以这里更新维度表里的所有cty或st改变的记录，而不是仅仅更新当前版本的记录DROP TABLE IF EXISTS tmp;CREATE TABLE tmp ASSELECT a.sk,a.id,a.name,b.cty,b.st,a.version,a.effective_date,a.expiry_date FROM tbl_dim a, tbl_stg bWHERE a.id=b.id AND (a.cty <> b.cty OR a.st <> b.st);DELETE FROM tbl_dim WHERE sk IN (SELECT sk FROM tmp);INSERT INTO tbl_dim SELECT * FROM tmp;-- 处理新增记录INSERT INTO tbl_dimSELECT    ROW_NUMBER() OVER (ORDER BY t1.id) + t2.sk_max,    t1.id,    t1.name,    t1.cty,    t1.st,    1,    ${hivevar:pre_date},    ${hivevar:max_date}FROM(SELECT t1.* FROM tbl_stg t1 LEFT JOIN tbl_dim t2 ON t1.id = t2.idWHERE t2.sk IS NULL) t1CROSS JOIN(SELECT COALESCE(MAX(sk),0) sk_max FROM tbl_dim) t2;

（5）执行定期装载

hive -S -f /home/grid/BigDataDWTest/scd_row_number.sql

查询维度表结果如图1所示。

select * from tbl_dim order by id,version;

图1

（6）再次执行定期装载，维度表的数据没有变化

hive -S -f /home/grid/BigDataDWTest/scd_row_number.sql

2. 用UDFRowSequence方法实现初始装载和定期装载
实验过程和ROW_NUMBER()方法基本一样，只是先要将hive-contrib-2.0.0.jar传到HDFS上，否则会报错。

hadoop dfs -put /home/grid/hive/lib/hive-contrib-2.0.0.jar /user

初始装载脚本init_UDFRowSequence.sql，内容如下：

USE test;ADD JAR hdfs:///user/hive-contrib-2.0.0.jar;CREATE TEMPORARY FUNCTION row_sequence as 'org.apache.hadoop.hive.contrib.udf.UDFRowSequence';-- 建立过渡表DROP TABLE IF EXISTS tbl_stg;CREATE TABLE tbl_stg (    id INT,    name STRING,    cty STRING,    st STRING)ROW FORMAT DELIMITED FIELDS TERMINATED BY ',';-- 建立维度表DROP TABLE IF EXISTS tbl_dim;CREATE TABLE tbl_dim (    sk INT,    id INT,    name STRING,    cty STRING,    st STRING,    version INT,    effective_date DATE,    expiry_date DATE)CLUSTERED BY (id) INTO 8 BUCKETSSTORED AS ORC TBLPROPERTIES ('transactional'='true');-- 向过渡表加载初始数据LOAD DATA LOCAL INPATH '/home/grid/BigDataDWTest/a.txt' INTO TABLE tbl_stg;-- 向维度表装载初始数据INSERT INTO tbl_dimSELECT    t2.sk_max + row_sequence(),    tbl_stg.*,    1,    CAST('1900-01-01' AS DATE),    CAST('2200-01-01' AS DATE)from tbl_stg CROSS JOIN (SELECT COALESCE(MAX(sk),0) sk_max FROM tbl_dim) t2;

定期装载脚本scd_UDFRowSequence.sql，内容如下：

USE test;ADD JAR hdfs:///user/hive-contrib-2.0.0.jar;CREATE TEMPORARY FUNCTION row_sequence as 'org.apache.hadoop.hive.contrib.udf.UDFRowSequence';-- 设置日期变量SET hivevar:pre_date = DATE_ADD(CURRENT_DATE(),-1);SET hivevar:max_date = CAST('2200-01-01' AS DATE);-- 向过渡表加载更新后的数据LOAD DATA LOCAL INPATH '/home/grid/BigDataDWTest/a.txt' OVERWRITE INTO TABLE tbl_stg;-- 向维度表装载更新后的数据-- 设置已删除记录和SCD2的过期UPDATE tbl_dimSET expiry_date = ${hivevar:pre_date}WHERE sk IN(SELECT a.sk FROM (SELECT sk,id,name FROM tbl_dim WHERE expiry_date = ${hivevar:max_date}) a LEFT JOIN tbl_stg b ON a.id=b.idWHERE b.id IS NULL OR a.name<>b.name);-- 处理SCD2新增行INSERT INTO tbl_dimSELECT    t2.sk_max + row_sequence(),    t1.id,    t1.name,    t1.cty,    t1.st,    t1.version,    t1.effective_date,    t1.expiry_dateFROM(SELECT    t2.id id,    t2.name name,    t2.cty cty,    t2.st st,    t1.version + 1 version,    ${hivevar:pre_date} effective_date,    ${hivevar:max_date} expiry_dateFROM tbl_dim t1 INNER JOIN tbl_stg t2ON t1.id=t2.id AND t1.name<>t2.name AND t1.expiry_date = ${hivevar:pre_date}LEFT JOIN tbl_dim t3 ON T1.id = t3.id AND t3.expiry_date = ${hivevar:max_date}WHERE t3.sk IS NULL) t1CROSS JOIN(SELECT COALESCE(MAX(sk),0) sk_max FROM tbl_dim) t2;-- 处理SCD1-- 因为hive的update还不支持子查询，所以这里使用了一个临时表存储需要更新的记录，用先delete再insert代替update-- 因为SCD1本身就不保存历史数据，所以这里更新维度表里的所有cty或st改变的记录，而不是仅仅更新当前版本的记录DROP TABLE IF EXISTS tmp;CREATE TABLE tmp ASSELECT a.sk,a.id,a.name,b.cty,b.st,a.version,a.effective_date,a.expiry_date FROM tbl_dim a, tbl_stg bWHERE a.id=b.id AND (a.cty <> b.cty OR a.st <> b.st);DELETE FROM tbl_dim WHERE sk IN (SELECT sk FROM tmp);INSERT INTO tbl_dim SELECT * FROM tmp;-- 处理新增记录INSERT INTO tbl_dimSELECT    t2.sk_max + row_sequence(),    t1.id,    t1.name,    t1.cty,    t1.st,    1,    ${hivevar:pre_date},    ${hivevar:max_date}FROM(SELECT t1.* FROM tbl_stg t1 LEFT JOIN tbl_dim t2 ON t1.id = t2.idWHERE t2.sk IS NULL) t1CROSS JOIN(SELECT COALESCE(MAX(sk),0) sk_max FROM tbl_dim) t2;

参考：
http://blog.csdn.net/wzy0623/article/details/49616643
http://www.remay.com.br/blog/hdp-2-2-how-to-create-a-surrogate-key-on-hive/
http://www.aboutechnologies.com/2016/02/hive-auto-increment-column.html