分享 ETL with Ruby

hooopo for Shopper+ · 2015年04月15日 · 最后由 hooopo 回复于 2016年06月14日 · 7435 次阅读

本帖已被管理员设置为精华贴

Why ETL

前面两篇介绍了数据仓库相关的基本概念和星型模型等常用建模方式。

这篇主要来谈 ETL。星型模型让分析查询变得容易，但存在以下问题：

一般先有 OLTP 系统，后有 OLAP 系统，分析系统只是辅助工具。
我们的事务系统不是按星型模型设计的，数据来源是 OLTP 类型应用。
数据源除了 OLTP 系统的数据，还需要与其他来源数据进行组合。

所以，需要引入一套 ETL 流程来把 OLTP 数据转化为星型模型，然后才能方便分析。

 +------------------+          +-----------+      +-----------+      +---------+ 
 |OLTP Data         +--------->+           |      |           |      |         | 
 +------------------+          |           |      |           |      |         | 
                               |           +----> |           +----> |         | 
 +------------------+          | Extract   |      |           |      |         | 
 |Google Analytics  +--------->+           |      |           |      |         | 
 +------------------+          |           |      |           |      |Analytics| 
                               | Transform |      |Star Schema|      |   &&    | 
 +------------------+          |           |      |           |      |Reporting| 
 |CRM Data          +--------->+           |      |           |      |         | 
 +------------------+          | Load      |      |           |      |         | 
                               |           +----> |           +----> |         | 
 +------------------+          |           |      |           |      |         | 
 |Others            +--------->+           |      |           |      |         | 
 +------------------+          +-----------+      +-----------+      +---------+

那么，ETL 是什么？

Extract data from a source
Transform the data for storing it in proper format or structure for querying and analysis purpose
Load it into the target

其实 ETL 工作我们并不陌生，正像 thbar 所说的 Rubyists — Are you doing ETL unknowingly?,下面这些都属于 ETL 范畴：

Write a script to migrate a legacy database to a new schema.
Automate processing of your data to generate a report.
Synchronize all or part of the data between 2 systems on a regular basis.
Prepare your data for indexing/searching.
Aggregate heterogeneous data sources into one consistent database.
Clean-up dirty or bogus data.
Geocode rows in an app to present them through a map app.
Implement a data export process for your users.

ETL Tools

和 Ruby 相关的 ETL 开源工具只有以下几个：

与 AR 结合紧密的 activewarehouse-etl：https://github.com/activewarehouse/activewarehouse-etl
Square 出品的轻量级 ETL 工具：https://github.com/square/ETL
activewarehouse-etl 维护者 thbar 新作：https://github.com/thbar/kiba

值得说明一下的是，由于 ActiveRecord 完全是针对 OLTP 场景设计的 ORM 工具，我们用 AR 来导数据一定会觉得巨慢，即使做了各种优化手段。这不是 AR 的错，是你没选对工具。ETL 工具针对这种场景做了优化，一般都是批量加载数据，速度可以说是有了质的提升。可以参考之前写过的Fastest way to load data in MySQL。

如果想了解 Activewarehosue ETL，这里有一个专门介绍 Activewarehouse ETL 的 slide 也值得一看：https://speakerdeck.com/thbar/transforming-data-with-ruby-and-activewarehouse-etl

Incremental Update & MySQL stream

针对大数据场景，Activewarehouse ETL 会提供增量更新和流读取的方式，只需要简单配置：

source :in, {
  :type => :database,
  :target => :operational_database
  :table => "people",
  :join => "addresses on people.address_id = addresses.id",
  :select => "people.email, addresses.city, addresses.state, people.created_at"
  :conditions => "people.unsubscribed = 0 AND people.date_of_death IS NULL"
  :order => "people.created_at",
  :new_records_only => "people.updated_at", # 增量导入
  :mysqlstream => true                      # 按流的方式
},
[
  :email,
  :city,
  :state
]

处理缓慢变化维度

维度建模的数据仓库中，有一个概念叫 Slowly Changing Dimensions，中文一般翻译成“缓慢变化维”，经常被简写为 SCD。缓慢变化维的提出是因为在现实世界中，维度的属性并不是静态的，它会随着时间的流失发生缓慢的变化。这种随时间发生变化的维度我们一般称之为缓慢变化维，并且把处理维度表的历史变化信息的问题称为处理缓慢变化维的问题，有时也简称为处理 SCD 的问题。常见类型有两种（Wikipedia 里面介绍了 6 种）：

类型 1 (Type 1): 覆盖旧记录。

有些 Dimension 表从业务上讲不需要保存历史记录或者只需要对原有记录进行修改。比如说 Customer 表中有 Customer 地址的属性，原有的地址输入错误我们需要修改这个属性而不需要对原有的错误地址进行保存，这个时候就可以使用 SCD Type 1

类型 2 (Type 2): 增加新记录。

Type 2 是精确捕获 Dimension 表历史变化的一种标准的方法，它通过对数据源表的 Change Data Capture (CDC) 机制来捕获数据源的变化，然后在 Dimension 表中插入一个新的记录再使旧的相应的记录失效。

缓慢变化维度一般设计成这样的结构：

# customer dimension
+-----------------+------------+------+-----+---------+----------------+
| id              | int(11)    | NO   | PRI | NULL    | auto_increment |
| customer_key    | int(11)    | YES  | MUL | NULL    |                |
| customer_type   | int(11)    | YES  | MUL | 1       |                |
| effective_start | datetime   | YES  |     | NULL    |                |
| effective_end   | datetime   | YES  |     | NULL    |                |
| latest_version  | tinyint(1) | YES  |     | 1       |                |
+-----------------+------------+------+-----+---------+----------------+

Activewarehouse ETL 对 SCD 问题也提供了解决方案。