Niraj Bhatt – Architect's Blog

Ruminations on .NET, Architecture & Design

Architecting Data Warehouse

As data in your applications grows it’s the database that usually becomes a bottleneck. It’s hard to scale a relational DB and the preferred approach for large scale applications is to create separate databases for writes and reads. These databases are referred as transactional database and reporting database. Though there are tools / techniques which can allow you to create snapshot of your transactional database for reporting purpose, sometimes they don’t quite fit the reporting requirements of an enterprise. These requirements typically are data analytics, effective schema (for an Information worker to self-service herself), historical data, better performance (flat data, no joins) etc. This is where a need for data warehouse or an OLAP system arises.

A Key point to remember is a data warehouse is mostly a relational database. It’s built on top of same concepts like Tables, Rows, Columns, Primary keys, Foreign Keys, etc. Before we talk about how data warehouses are typically structured let’s understand key components that can create a data flow between OLTP systems and OLAP systems. There are 3 major areas to it:

a) OLTP system should be capable of tracking its changes as all these changes should go back to data warehouse for historical recording. For e.g. if an OLTP transaction moves a customer from silver to gold category, OLTP system needs to ensure that this change is tracked and send to data warehouse for reporting purpose. A report in context could be how many customers divided by geographies moved from sliver to gold category. In data warehouse terminology this process is called Change Data Capture. There are quite a few systems that leverage database triggers to move these changes to corresponding tracking tables. There are also out of box features provided by some databases e.g. SQL Server 2008 offers Change Data Capture and Change Tracking for addressing such requirements.

b) After we make the OLTP system capable of tracking its changes we need to provision a batch process that can run periodically and takes these changes from OLTP system and dump them into data warehouse. There are many tools out there that can help you fill this gap – SQL Server Integration Services happens to be one of them.

c) So we have an OLTP system that knows how to track its changes, we have jobs that run periodically to move these changes to warehouse. The question though remains is how warehouse will record these changes? This structural change in data warehouse arena is often covered under something called Slowly Changing Dimension (SCD). While we will talk about dimensions in a while, SCD can be applied to pure relational tables too. SCD enables a database structure to capture historical data. This would create multiple records for a given entity in relational database and data warehouses prefer having their own primary key, often known as surrogate key.

As I mentioned a data warehouse is just a relational database but industry often attributes a specific schema style to data warehouses. These styles are Star Schema or Snowflake Schema. The motivation behind these styles is to create a flat database structure (as opposed to normalized one), which is easy to understand / use, easy to query and easy to slice / dice. Star schema is a database structure made up of dimensions and facts. Facts are generally the numbers (sales, quantity, etc.) that you want to slice and dice. Fact tables have these numbers and have references (foreign keys) to set of tables that provide context around those facts. E.g. if you have recorded 10,000 USD as sales that number would go in a sales fact table and could have foreign keys attached to it that refers to the sales agent responsible for sale and to time table which contains the dates between which that sale was made. These agent and time tables are called dimensions which provide context to the numbers stored in fact tables. This schema structure of fact being at center surrounded by dimensions is called Star schema. A similar structure with difference of dimension tables being normalized is called a Snowflake schema.

This relational structure of facts and dimensions serves as an input for another analysis structure called Cube. Though physically Cube is a special structure supported by commercial databases like SQL Server Analysis Services, logically it’s a multidimensional structure where dimensions define the sides of cube and facts define the content. Facts are often called as Measures inside a cube. Dimensions often tend to form a hierarchy. E.g. Product may be broken into categories and categories in turn to individual items. Category and Items are often referred as Levels and their constituents as Members with their overall structure called as Hierarchy. Measures are rolled up as per dimensional hierarchy. These rolled up measures are called Aggregates. Now this may seem like an overwhelming vocabulary to deal with but don’t worry it will sink in as you start working with Cubes and others.

To get a simplistic view of a cube you might want to think of it as a relational table – with rows representing ‘measures’ and columns representing ‘dimensions’. Taking our product example the cube can be expressed as a relational table shown below:

Above relational table, shows aggregated data at per day per item level. As you would have observed each row in the table contains a unique set of values. You can further aggregate this data as necessary in real time – e.g. to derive sales for a given item in a given month or sales a given category in given month. So, at which level in dimensional hierarchy we should aggregate data – hour level, day level or month level? The answer is – it depends on your reporting requirements. In this example, if the reporting always happens at a month level, there is no point in aggregating data at day level to associate additional runtime overhead. On the other side, if you plan to drill at an hour level you might have to rebuild the cube including hour as an additional dimension.

Let’s see few other terms that we would run into while talking about data warehouses.

ODS or an Operational Data Store is a frequently misused term. There would be few users in your organization that want to report on most current data and can’t afford to miss a single transaction for their report. Then there is another set of users that typically don’t care how current the data is. Mostly senior level executives who are interesting in trending, mining, forecasting, strategizing, etc. don’t care for that one specific transaction. This is where an ODS can come in handy. ODS can use the same star schema and the OLAP cubes we saw earlier. The only difference is that the data inside an ODS would be short lived, i.e. for few months and ODS would sync with OLTP system every few minutes. Data warehouse can periodically sync with ODS either daily or weekly depending on business drivers.

Data marts are another frequently talked about topic in data warehousing. They are subject-specific data warehouse. Data warehouses that try to span over an enterprise are normally too big to scope, build, manage, track, etc. Hence they are often scaled down to something called Data mart that supports a specific segment of business like sales, marketing, or support. Data marts too, are often designed using star schema model discussed earlier. Industry is divided when it comes to use of data marts. Some experts prefer having data marts along with a central data warehouse. Data warehouse here acts as information staging and distribution hub with spokes being data marts connected via data feeds serving summarized data. Others eliminate the need for a centralized data warehouse citing that most users want to report on detailed data.

One response to “Architecting Data Warehouse

  1. justinhay December 10, 2010 at 6:18 am

    Hi Niraj,

    This is a nice overview. I agree that “ODS” is a term applied rather loosely. I would be interested to know what you think the place is for a “real-time” data warehouse?

    As for the central normalized warehouse – I think there’s a valuable place for it – in addition to data marts. In both cases, they can hold atomic, detailed information, and not just aggregates.

    The challenge I see in many organizations is to deal with large volumes of detailed transactional data, without having to migrate it through multiple EDW layers.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: