数据仓库中历史拉链表的更新方法
在之前介绍过数据仓库中的历史拉链表《极限存储–历史拉链表》,
使用这种方式即可以记录历史,而且最大程度的节省存储。这里简单介绍一下这种历史拉链表的更新方法。
本文中假设:
- 数据仓库中订单历史表的刷新频率为一天,当天更新前一天的增量数据;
- 如果一个订单在一天内有多次状态变化,则只会记录最后一个状态的历史;
- 订单状态包括三个:创建、支付、完成;
- 创建时间和修改时间只取到天,如果源订单表中没有状态修改时间,那么抽取增量就比较麻烦,需要有个机制来确保能抽取到每天的增量数据;
- 本文中的表和SQL都使用Hive的HQL语法;
- 源系统中订单表结构为:
CREATE TABLE orders (
orderid INT,
createtime STRING,
modifiedtime STRING,
status STRING
) stored AS textfile;
7.在数据仓库的ODS层,有一张订单的增量数据表,按天分区,存放每天的增量数据:
CREATE TABLE t_ods_orders_inc (
orderid INT,
createtime STRING,
modifiedtime STRING,
status STRING
) PARTITIONED BY (day STRING)
stored AS textfile;
8. 在数据仓库的DW层,有一张订单的历史数据拉链表,存放订单的历史状态数据:
CREATE TABLE t_dw_orders_his (
orderid INT,
createtime STRING,
modifiedtime STRING,
status STRING,
dw_start_date STRING,
dw_end_date STRING
) stored AS textfile;
9. 暂未考虑Hive上表的查询性能问题,只实现功能;
10. 2015-08-21至2015-08-23,每天原系统订单表的数据如下,红色标出的为当天发生变化的订单,即增量数据:
全量初始化
在数据从源业务系统每天正常抽取和刷新到DW订单历史表之前,需要做一次全量的初始化,就是从源订单表中昨天以前的数据全部抽取到ODW,并刷新到DW。
以上面的数据为例,比如在2015-08-21这天做全量初始化,那么我需要将包括2015-08-20之前的所有的数据都抽取并刷新到DW:
第一步,抽取全量数据到ODS:
INSERT overwrite TABLE t_ods_orders_inc PARTITION (day = ‘2015-08-20′)
SELECT orderid,createtime,modifiedtime,status
FROM orders
WHERE createtime <= ‘2015-08-20′;
第二步,从ODS刷新到DW:
INSERT overwrite TABLE t_dw_orders_his
SELECT orderid,createtime,modifiedtime,status,
createtime AS dw_start_date,
‘9999-12-31′ AS dw_end_date
FROM t_ods_orders_inc
WHERE day = ‘2015-08-20′;
完成后,DW订单历史表中数据如下:
- spark-sql> select * from t_dw_orders_his;
- 1 2015-08-18 2015-08-18 创建 2015-08-18 9999-12-31
- 2 2015-08-18 2015-08-18 创建 2015-08-18 9999-12-31
- 3 2015-08-19 2015-08-21 支付 2015-08-19 9999-12-31
- 4 2015-08-19 2015-08-21 完成 2015-08-19 9999-12-31
- 5 2015-08-19 2015-08-20 支付 2015-08-19 9999-12-31
- 6 2015-08-20 2015-08-20 创建 2015-08-20 9999-12-31
- 7 2015-08-20 2015-08-21 支付 2015-08-20 9999-12-31
- Time taken: 2.296 seconds, Fetched 7 row(s)
增量抽取
每天,从源系统订单表中,将前一天的增量数据抽取到ODS层的增量数据表。
这里的增量需要通过订单表中的创建时间和修改时间来确定:
INSERT overwrite TABLE t_ods_orders_inc PARTITION (day = ‘${day}‘)
SELECT orderid,createtime,modifiedtime,status
FROM orders
WHERE createtime = ‘${day}’ OR modifiedtime = ‘${day}';
注意:在ODS层按天分区的增量表,最好保留一段时间的数据,比如半年,为了防止某一天的数据有问题而回滚重做数据。
增量刷新历史数据
从2015-08-22开始,需要每天正常刷新前一天(2015-08-21)的增量数据到历史表。
第一步,通过增量抽取,将2015-08-21的数据抽取到ODS:
INSERT overwrite TABLE t_ods_orders_inc PARTITION (day = ‘2015-08-21′)
SELECT orderid,createtime,modifiedtime,status
FROM orders
WHERE createtime = ‘2015-08-21′ OR modifiedtime = ‘2015-08-21′;
ODS增量表中2015-08-21的数据如下:
- spark-sql> select * from t_ods_orders_inc where day = '2015-08-21';
- 3 2015-08-19 2015-08-21 支付 2015-08-21
- 4 2015-08-19 2015-08-21 完成 2015-08-21
- 7 2015-08-20 2015-08-21 支付 2015-08-21
- 8 2015-08-21 2015-08-21 创建 2015-08-21
- Time taken: 0.437 seconds, Fetched 4 row(s)
第二步,通过DW历史数据(数据日期为2015-08-20),和ODS增量数据(2015-08-21),刷新历史表:
先把数据放到一张临时表中:
- DROP TABLE IF EXISTS t_dw_orders_his_tmp;
- CREATE TABLE t_dw_orders_his_tmp AS
- SELECT orderid,
- createtime,
- modifiedtime,
- status,
- dw_start_date,
- dw_end_date
- FROM (
- SELECT a.orderid,
- a.createtime,
- a.modifiedtime,
- a.status,
- a.dw_start_date,
- CASE WHEN b.orderid IS NOT NULL AND a.dw_end_date > '2015-08-21' THEN '2015-08-20' ELSE a.dw_end_date END AS dw_end_date
- FROM t_dw_orders_his a
- left outer join (SELECT * FROM t_ods_orders_inc WHERE day = '2015-08-21') b
- ON (a.orderid = b.orderid)
- UNION ALL
- SELECT orderid,
- createtime,
- modifiedtime,
- status,
- modifiedtime AS dw_start_date,
- '9999-12-31' AS dw_end_date
- FROM t_ods_orders_inc
- WHERE day = '2015-08-21'
- ) x
- ORDER BY orderid,dw_start_date;
其中:
UNION ALL的两个结果集中,第一个是用历史表left outer join 日期为 ${yyy-MM-dd} 的增量,能关联上的,并且dw_end_date > ${yyy-MM-dd},说明状态有变化,则把原来的dw_end_date置为(${yyy-MM-dd} – 1), 关联不上的,说明状态无变化,dw_end_date无变化。
第二个结果集是直接将增量数据插入历史表。
最后把临时表中数据插入历史表:
INSERT overwrite TABLE t_dw_orders_his
SELECT * FROM t_dw_orders_his_tmp;
刷新完后,历史表中数据如下
- spark-sql> select * from t_dw_orders_his order by orderid,dw_start_date;
- 1 2015-08-18 2015-08-18 创建 2015-08-18 9999-12-31
- 2 2015-08-18 2015-08-18 创建 2015-08-18 9999-12-31
- 3 2015-08-19 2015-08-21 支付 2015-08-19 2015-08-20
- 3 2015-08-19 2015-08-21 支付 2015-08-21 9999-12-31
- 4 2015-08-19 2015-08-21 完成 2015-08-19 2015-08-20
- 4 2015-08-19 2015-08-21 完成 2015-08-21 9999-12-31
- 5 2015-08-19 2015-08-20 支付 2015-08-19 9999-12-31
- 6 2015-08-20 2015-08-20 创建 2015-08-20 9999-12-31
- 7 2015-08-20 2015-08-21 支付 2015-08-20 2015-08-20
- 7 2015-08-20 2015-08-21 支付 2015-08-21 9999-12-31
- 8 2015-08-21 2015-08-21 创建 2015-08-21 9999-12-31
- Time taken: 0.717 seconds, Fetched 11 row(s)
由于在2015-08-21做了8月20日以前的数据全量初始化,而订单3、4、7在2015-08-21的增量数据中也存在,因此都有两条记录,但不影响后面的查询。
再看将2015-08-22的增量数据刷新到历史表:
- INSERT overwrite TABLE t_ods_orders_inc PARTITION (day = '2015-08-22')
- SELECT orderid,createtime,modifiedtime,status
- FROM orders
- WHERE createtime = '2015-08-22' OR modifiedtime = '2015-08-22';
- DROP TABLE IF EXISTS t_dw_orders_his_tmp;
- CREATE TABLE t_dw_orders_his_tmp AS
- SELECT orderid,
- createtime,
- modifiedtime,
- status,
- dw_start_date,
- dw_end_date
- FROM (
- SELECT a.orderid,
- a.createtime,
- a.modifiedtime,
- a.status,
- a.dw_start_date,
- CASE WHEN b.orderid IS NOT NULL AND a.dw_end_date > '2015-08-22' THEN '2015-08-21' ELSE a.dw_end_date END AS dw_end_date
- FROM t_dw_orders_his a
- left outer join (SELECT * FROM t_ods_orders_inc WHERE day = '2015-08-22') b
- ON (a.orderid = b.orderid)
- UNION ALL
- SELECT orderid,
- createtime,
- modifiedtime,
- status,
- modifiedtime AS dw_start_date,
- '9999-12-31' AS dw_end_date
- FROM t_ods_orders_inc
- WHERE day = '2015-08-22'
- ) x
- ORDER BY orderid,dw_start_date;
- INSERT overwrite TABLE t_dw_orders_his
- SELECT * FROM t_dw_orders_his_tmp;
刷新完后历史表数据如下:
- spark-sql> select * from t_dw_orders_his order by orderid,dw_start_date;
- 1 2015-08-18 2015-08-18 创建 2015-08-18 2015-08-21
- 1 2015-08-18 2015-08-22 支付 2015-08-22 9999-12-31
- 2 2015-08-18 2015-08-18 创建 2015-08-18 2015-08-21
- 2 2015-08-18 2015-08-22 完成 2015-08-22 9999-12-31
- 3 2015-08-19 2015-08-21 支付 2015-08-19 2015-08-20
- 3 2015-08-19 2015-08-21 支付 2015-08-21 9999-12-31
- 4 2015-08-19 2015-08-21 完成 2015-08-19 2015-08-20
- 4 2015-08-19 2015-08-21 完成 2015-08-21 9999-12-31
- 5 2015-08-19 2015-08-20 支付 2015-08-19 9999-12-31
- 6 2015-08-20 2015-08-20 创建 2015-08-20 2015-08-21
- 6 2015-08-20 2015-08-22 支付 2015-08-22 9999-12-31
- 7 2015-08-20 2015-08-21 支付 2015-08-20 2015-08-20
- 7 2015-08-20 2015-08-21 支付 2015-08-21 9999-12-31
- 8 2015-08-21 2015-08-21 创建 2015-08-21 2015-08-21
- 8 2015-08-21 2015-08-22 支付 2015-08-22 9999-12-31
- 9 2015-08-22 2015-08-22 创建 2015-08-22 9999-12-31
- 10 2015-08-22 2015-08-22 支付 2015-08-22 9999-12-31
- Time taken: 0.66 seconds, Fetched 17 row(s)
查看2015-08-21的历史快照数据:
- spark-sql> select * from t_dw_orders_his where dw_start_date <= '2015-08-21' and dw_end_date >= '2015-08-21';
- 1 2015-08-18 2015-08-18 创建 2015-08-18 2015-08-21
- 2 2015-08-18 2015-08-18 创建 2015-08-18 2015-08-21
- 3 2015-08-19 2015-08-21 支付 2015-08-21 9999-12-31
- 4 2015-08-19 2015-08-21 完成 2015-08-21 9999-12-31
- 5 2015-08-19 2015-08-20 支付 2015-08-19 9999-12-31
- 6 2015-08-20 2015-08-20 创建 2015-08-20 2015-08-21
- 7 2015-08-20 2015-08-21 支付 2015-08-21 9999-12-31
- 8 2015-08-21 2015-08-21 创建 2015-08-21 2015-08-21
订单1在2015-08-21的时候还处于创建的状态,在2015-08-22的时候状态变为支付。
再刷新2015-08-23的增量数据:
按照上面的方法刷新完后,历史表数据如下:
- spark-sql> select * from t_dw_orders_his order by orderid,dw_start_date;
- 1 2015-08-18 2015-08-18 创建 2015-08-18 2015-08-21
- 1 2015-08-18 2015-08-22 支付 2015-08-22 2015-08-22
- 1 2015-08-18 2015-08-23 完成 2015-08-23 9999-12-31
- 2 2015-08-18 2015-08-18 创建 2015-08-18 2015-08-21
- 2 2015-08-18 2015-08-22 完成 2015-08-22 9999-12-31
- 3 2015-08-19 2015-08-21 支付 2015-08-19 2015-08-20
- 3 2015-08-19 2015-08-21 支付 2015-08-21 2015-08-22
- 3 2015-08-19 2015-08-23 完成 2015-08-23 9999-12-31
- 4 2015-08-19 2015-08-21 完成 2015-08-19 2015-08-20
- 4 2015-08-19 2015-08-21 完成 2015-08-21 9999-12-31
- 5 2015-08-19 2015-08-20 支付 2015-08-19 2015-08-22
- 5 2015-08-19 2015-08-23 完成 2015-08-23 9999-12-31
- 6 2015-08-20 2015-08-20 创建 2015-08-20 2015-08-21
- 6 2015-08-20 2015-08-22 支付 2015-08-22 9999-12-31
- 7 2015-08-20 2015-08-21 支付 2015-08-20 2015-08-20
- 7 2015-08-20 2015-08-21 支付 2015-08-21 9999-12-31
- 8 2015-08-21 2015-08-21 创建 2015-08-21 2015-08-21
- 8 2015-08-21 2015-08-22 支付 2015-08-22 2015-08-22
- 8 2015-08-21 2015-08-23 完成 2015-08-23 9999-12-31
- 9 2015-08-22 2015-08-22 创建 2015-08-22 9999-12-31
- 10 2015-08-22 2015-08-22 支付 2015-08-22 9999-12-31
- 11 2015-08-23 2015-08-23 创建 2015-08-23 9999-12-31
- 12 2015-08-23 2015-08-23 创建 2015-08-23 9999-12-31
- 13 2015-08-23 2015-08-23 支付 2015-08-23 9999-12-31
订单1从20号-23号,状态变化了三次,历史表中有三条记录。
- //查看2015-08-22当天的历史快照,可以看出,和上面图中2015-08-22时候订单表中的数据是一样的
- spark-sql> select * from t_dw_orders_his where dw_start_date <= '2015-08-22' and dw_end_date >= '2015-08-22';
- 1 2015-08-18 2015-08-22 支付 2015-08-22 2015-08-22
- 2 2015-08-18 2015-08-22 完成 2015-08-22 9999-12-31
- 3 2015-08-19 2015-08-21 支付 2015-08-21 2015-08-22
- 4 2015-08-19 2015-08-21 完成 2015-08-21 9999-12-31
- 5 2015-08-19 2015-08-20 支付 2015-08-19 2015-08-22
- 6 2015-08-20 2015-08-22 支付 2015-08-22 9999-12-31
- 7 2015-08-20 2015-08-21 支付 2015-08-21 9999-12-31
- 8 2015-08-21 2015-08-22 支付 2015-08-22 2015-08-22
- 9 2015-08-22 2015-08-22 创建 2015-08-22 9999-12-31
- 10 2015-08-22 2015-08-22 支付 2015-08-22 9999-12-31
- Time taken: 0.328 seconds, Fetched 10 row(s)
- //查看当前所有订单的最新状态
- spark-sql> select * from t_dw_orders_his where dw_end_date = '9999-12-31';
- 1 2015-08-18 2015-08-23 完成 2015-08-23 9999-12-31
- 2 2015-08-18 2015-08-22 完成 2015-08-22 9999-12-31
- 3 2015-08-19 2015-08-23 完成 2015-08-23 9999-12-31
- 4 2015-08-19 2015-08-21 完成 2015-08-21 9999-12-31
- 5 2015-08-19 2015-08-23 完成 2015-08-23 9999-12-31
- 6 2015-08-20 2015-08-22 支付 2015-08-22 9999-12-31
- 7 2015-08-20 2015-08-21 支付 2015-08-21 9999-12-31
- 8 2015-08-21 2015-08-23 完成 2015-08-23 9999-12-31
- 9 2015-08-22 2015-08-22 创建 2015-08-22 9999-12-31
- 10 2015-08-22 2015-08-22 支付 2015-08-22 9999-12-31
- 11 2015-08-23 2015-08-23 创建 2015-08-23 9999-12-31
- 12 2015-08-23 2015-08-23 创建 2015-08-23 9999-12-31
- 13 2015-08-23 2015-08-23 支付 2015-08-23 9999-12-31
- Time taken: 0.293 seconds, Fetched 13 row(s)
实际业务中,有可能某一天的数据有问题,需要回滚或重做,这点有点麻烦,后续文章再介绍。
posted on 2017-08-16 16:39 running_wolf 阅读(2042) 评论(0) 编辑 收藏 举报