Modeling Slowly Changing Dimensions in Data Warehouses

Modern data warehouse design assumes that business transactions such as sales, orders, shipments, and receivables can occur at a rapid rate and each the details of each transaction needs to be recorded. Hence a fact table with a dimensional model contains a separate record for each business transaction. While in contrast, the describing or text-based values of the transaction or dimension often remain fairly constant. Often, dimensional tables within the dimensional model do not take changes into account.

But in reality, some dimensional values do change over time and numerous fields of a given row within a dimension table will need to be updated. This phenomenon in data modeling is known as “slowly changing dimensions” and it can be applied to any dimension table within a data warehouse schema. Moreover, both simple and advanced modeling techniques have been established and can be implemented for handling updates and changes within a dimension table. In addition, slowly changing dimensions assist the data warehouse in precisely recording the past values, providing an efficient method for tracking history, and allowing for the ability to respond to changes to descriptive values of transactions.

Examples of slowly changing dimensions include:: – account name; – customer phone number; – vendor address; – product description

These are good examples as they are text-based values that remain relatively constant, but can change and commonly do change over time. Names, phone numbers, addresses are fairly intuitive and it is easy to see how these values can change slowly over time. But let’s see how a product description could change… A simple ingredient change or a packaging change in a product may be so trivial that the organization does not decide to give the product a new product id. Rather the source system provides the data warehouse with a revised description of the product. Hence the data warehouse needs to track both the old and new descriptions of the product.

“The data warehouse must accept the responsibility of accurately describing the past. By doing so, the data warehouse simplifies the responsibilities of the OLTP system. Not only does the data warehouse relieve the OLTP system of almost all forms of reporting, but the data warehouse contains special structures that have several ways of tracking historical data.”
Ralph Kimball, PHD

Originally pioneered by Ralph Kimball, PhD, four main data modeling techniques have been established for managing dimension tables that contain slowly changing dimensions:

Type One – Overwrite the Record

Type Two – Update Record to Inactive / Create an Active Record

Type Three – Leverage Previous and Current Value Fields

Type Four – Insert Into a History Table

These four data modeling techniques range from the complete loss of historical data with relative ease of implementation to methods of saving almost all historical data but containing more technical challenges. Choice of the appropriate technique by the database designer can ensure that the data warehouse contains required historical values and allows for comparisons of current data or data from other time periods.

Type One Modeling – Overwrite the Record

Type one modeling is very simple and effectively handles updates to incorrect and/or outdated values within the dimension table. Type one models do not retain a history of changes and do not store previous values in any way. With this modeling technique, the data field that has changed is simply updated to reflect the most current value. Since this technique does not preserve any previous value, it should only be used when there is no requirement to retain historical data.

Type one models makes sense when correcting data issues, but not when the system requires retrieval of historical values. Without a doubt, type one is the simplest method for handling slowly changing dimensions. However, best practice is to utilize this technique on a very limited basis. Moreover, it should only be implemented when there is no need to track the history of changes.

Example of Type One Model
First an electronics company loads its list of products within its data warehouse. Later the company discovers that a television which was originally manufactured in the USA is now manufactured in China, and the company has in fact never received a shipment of the older American-made televisions. The product description column in the table is simply updated, and the date and time of the update is entered in the “Last Update” column.

To correct the product description, the existing description is simply overwritten with the correct value, and the “Last Update” field is revised with the current date. Moreover, the product description is updated because the item is no longer manufactured in the United States. Utilizing the type one method of managing slowly changing dimensions, the record on the product dimension table is updated in-place in the following manner…: – The product description is updated in-place to the new description.; – The last update is updated to the date and time of the change.

Type Two Modeling – Update Record to Inactive / Create an Active Record

Type two modeling is a very reliable and straightforward technique for preserving history of changes to dimensional tables. It is commonly utilized when a full or partial set of the previous values of a dimension’s attributes must be retained. In this modeling technique, every time an update occurs to any of the values in a dimensional table, a new record is physically inserted as active or current into the same table and the old records are marked inactive or historical.

The main advantage of type two modeling is that it can accommodate and record a massive amount of history and nearly an unlimited changes to slowly changing dimensions. But this advantage also can become a significant drawback. Within this modeling technique, it is possible for dimension tables to grow to massive sizes and adversely affect both system and query performance. In addition, since upon every change this modeling technique requires an update to the linked fact table, implementation is fairly difficult and may require substantial effort to design, develop, and support.

Two unique sub-methods have been established for distinguishing the active (or current) record from the inactive (or historical) records: active flagging and tuple versioning.

In the type two – active flagging sub-method, an “Active” column is constructed within the dimensional table and acts as a flag. Flags are binary fields with permissible values being Y and N, T and F, or 1 and 0. A positive value (Y, T, 1) indicates an active record, while a negative value (N, F, 0) indicates an inactive record.

Example of Type Two – Active Flagging Sub-Method

Suppose that a vendor changes his phone number to 858-555-6555 from 202-555-8639 because the phone company has added a new area code. Using active flagging, the change to the vendor dimension table would be processed as follows…

• The current active record in the vendor dimension table is updated:: – The active flag is changed from T to F.; – The phone number field remains as 202-555-6555.
• A new record is inserted into the vendor dimension table:: – The vendor key is given the next available integer number.; – The vendor name field remains the same as in the old record.; – The phone number has the new value of 858-555-8639.; – The active flag is set to T.

In the type two – tuple versioning sub-method, start date and end date columns are included on the dimension table. The values of these date fields then define the period during which that record has been active. For the most part, the start date is the date the record has been created. Moreover, the end date is the date the record has become inactive, either because a newer record has replaced it or because the original record in the source system no longer exists. On the active record, the end date will either be left null, blank, or identified in another way such as all-nines, according to modeler’s preferences and standards.

The main advantage of the tuple versioning method over active flagging is that it provides an audit trail of the date and sequence of all updates.

Example of Type Two – Tuple Versioning Sub-Method

Continuing with the same example above but using the tuple versioning method, the updated phone number in a record in the vendor dimension table is processed as follows:

• The current active record in the vendor dimension table is updated:: – The end date is updated to the date of the change.; – The phone number field remains as 202-555-6555.
• A new record is inserted into the vendor dimension table:: – The vendor key is given the next available integer number.; – The vendor name field remains the same as in the old record.; – The phone number has the new value of 858-555-8639.; – The start date is the date the record is inserted into the table.; – The end date is intentionally left blank.

Type Three Model – Leverage Previous and Current Value Fields

Type three models are defined by adding one or more columns to the dimension table, so that the new (or current or active) and old (or historical or inactive) value of an field are stored. In addition, multiple previous values can be stored within the table depending on how many previous columns are included in the the dimension table.

Type three modeling is a middle state between complete history loss of type one models and the numerous additional records of type two models. Moreover, type three models only provide only a limited view of history as only a predefined number previous value of any attribute can be retained, rather than a complete history. This modeling technique is fairly useful when changes to the dimension table are made on a regular schedule, such as annually, and when archival copies of the database are stored offline for historical and audit purposes. Type three modeling is not so useful when changes are more frequent and unpredictable. Multiple previous value fields could be added to the record to provide a longer historical trail, but it may be a challenge to design the table with the optimum number of previous fields. For the most part, type three modeling makes most sense to be implemented only when only there are a limited number of previous values that need to be retained.

Continuing with the examples of type two models, the phone number of a vendor changes from 202-555-6555 to 858-555-8639 because the phone company has added a new area code. But in this model the vendor dimension table contains a current phone and previous phone fields. Initially the vendor record contains a null value for the previous phone number and a value of 202-555-6555 for the current phone number.

In order to process this update using the previous and current fields, the vendor dimension table is updated:: – The previous phone field is updated to 202-555-6555, which is now the vendor’s last previous phone number.; – The current phone field is updated to 858-555-8639, which is now the vendor’s current number.; – The effective date is updated to the date of the change.

Type Four Model – Insert Into a History Table

Type four models, also known as leveraging “history tables”, is the most technically sophisticated of the four models and may be the most difficult to implement. This modeling technique provides for nearly unlimited tracking of historical records while having less storage requirement than type two models. Rather than storing the changes in the same table, a second “history” table is created which stores only the previous values of slowly changing dimensions.

Similar to type two models, type four models accommodate infinite changes to dimensional fields and create an additional record for every change to a dimensional attribute. But in contrast to type two, type four models allow for every change to an attribute to be generated within a new record in a relatively compact history table. The history table is subsequently more efficient in capturing a large amount of historical data.

Another key advantage of type four models is an efficient manner to query against a timeframe as the related search index only requires two fields (key and date fields). Other modeling techniques require more fields in the search index for date queries. Thus the search index utilizing a type four model is smaller, more intuitive, and quicker to retrieve the relevant record in the dimension table than the search index using other modeling techniques.

Type four models do have some important disadvantages. Namely type four models require implementation of multiple tables, are less intuitive for query developers, require more effort to develop and maintain than other types of models, and allow for history tables to grow to massive size.

Continuing with the examples of type two and type three models and utilizing the history table of type four models, the vendor dimension table would be updated and a new record will be inserted into the vendor history table in the following manner…

• The vendor dimension table is updated:: – The phone number is updated from 202-555-8638 to 858-555-6555.
• A new record is inserted into the vendor history table:: – The vendor key is copied from the vendor dimension table.; – The phone number 858-555-6555 is inserted.; – The effective date of 12/15/2008 is inserted.