Handle Slowly Changing Dimensions in SQL Server Integration Services

By:   |   Comments (16)   |   Related: More > Integration Services Analysis Services Tasks


Problem

We have experimented with the Slowly Changing Dimension (SCD) Data Flow Transformation that is available in the SSIS designer and have found a few issues with it.  Our major concern is the use of the OLE DB Command Data Flow Transformation for all updates (inferred, type 1 and type 2) to rows in the dimension table.  Do you have any suggestions?

Solution

You have hit upon the one issue with the Slowly Changing Dimension Data Flow Transformation that likely requires an alternative approach.  The issue with the OLE DB Command is that it executes a single SQL statement for each row passing through the Data Flow.  When the volume of rows is substantial, this creates an unnecessary and probably unacceptable performance hit.  Let's take a step back and analyze the task at hand then discuss an alternative solution.

The term slowly changing dimensions encompasses the following three different methods for handling changes to columns in a data warehouse dimension table:

  • Type 1 - update the columns in the dimension row without preserving any change history.
  • Type 2 - preserve the change history in the dimension table and create a new row when there are changes. 
  • Type 3 - some combination of Type 1 and Type 2, usually maintaining multiple instances of a column in the dimension row; e.g. a current value and one or more previous values.

A dimension that implements Type 2 changes would typically have the following housekeeping columns to identify the current row and the effective date range for each row:

  • Natural Key - the unique source system key that identifies the entity; e.g. CustomerID in the source system would be called nk_CustomerID in the dimension.
  • Surrogate Key (or warehouse key) - typically an identity value used to uniquely identify the row in the dimension.  For a given natural key there will be an instance of a row for each Type 2 change so the natural key will not be unique in the dimension.
  • CurrentMember - a bit column to indicate if the row is the current row.
  • EffectiveDate - a datetime (or smalldatetime) column to indicate when the row became the current row.
  • ExpirationDate - a datetime (or smalldatetime) column to indicate when the row ceased being the current row.

The effective date range columns retain the history of a natural key in the dimension, allowing us to see the column values at any point in time.  Fact table rows can be joined to the dimension row where the fact row transaction date is between the effective date range of the dimension row.

When you add the SCD Data Flow Transformation to the Data Flow designer, you step through a wizard to configure the task, and you will wind up with the Slowly Changing Dimension task and everything that follows below being added to the Data Flow designer (the task names generated by the SCD wizard have been updated to add clarification):

scd transform

Main points about the above screen shot:

  • The Excel Source is a sample data source representing data extracted from a source system that is used to update a dimension table in a data warehouse.
  • The Type 1 OLE DB Command task updates dimension rows one at a time by executing an UPDATE statement on the dimension table.
  • The Type 2 OLE DB Command task "expires" the current dimension rows one at a time (sets the ExpirationDate or CurrentMember flag) by executing an UPDATE statement.
  • The Insert New OLE DB Destination task inserts a new row into the dimension table when there is a new row in the source system or a Type 2 change. 
  • The Inferred OLE DB Command task performs a Type 1 update to a dimension row that was created with default values as a result of an early arriving fact.  An early arriving fact is one where the fact row has a source system key value that does not exist in the dimension; we will discuss Inferred processing in part two of this tip.

Now that we have described how the SCD transformation implements slowly changing dimension processing, we can discuss an alternative solution.  As an example we will use a Customer dimension that is updated with source system data in an Excel spreadsheet.  The SSIS package Control Flow looks like this:

alt transform

Main points about the above solution:

  • Truncate Customer Staging Table is an Execute SQL task that clears out the Customer dimension staging table. 
  • Stage Customer Data from Source System is a Data Flow task that extracts the rows from the Excel spreadsheet, cleanses and transforms the data, and writes the data out to the staging table.
  • Update Customer Dimension is an Execute SQL task that invokes a stored procedure that implements the Type 1 and Type 2 handling on the Customer dimension.

An additional detail about Type 1 and Type 2 processing is that a dimension may implement both.  In other words some column changes may be handled as Type 1 and other column changes may be handled as Type 2. An elegant way to implement this is to take advantage of the SQL Server CHECKSUM function.  CHECKSUM calculates a unique integer hash value based on the values of every column in a row or a subset of columns.  We can use a hash value comparison to determine whether anything has changed in our list of columns in the staging table versus the dimension table.  

Let's take a look at our Customer dimension table:

dim customer

The checksum columns are defined as follows:

  • [Type1Checksum] AS CHECKSUM([ContactName],[ContactTitle],[Phone],[Fax])
  • [Type2Checksum] AS CHECKSUM([Address],[City],[Region],[PostalCode],[Country])

There is a separate CHECKSUM value calculated for the list of Type 1 columns and the list of Type 2 columns.  In our staging table we have the same two CHECKSUM computed columns; the column lists must match exactly in order for this to work. 

As a general rule the staging table schema mirrors the dimension table schema but includes a couple of other housekeeping columns as shown below:

stg dim customer

The housekeeping columns in staging are as follows:

  • IsNew is set to 1 if this is a new dimension row.

  • IsType1 is set to 1 if there is a change to any column handled as Type 1.

  • IsType2 is set to 1 if there is a change to any column handled as Type 2.

Finally let's review the single stored procedure that implements the Type 1 and Type 2 processing and is invoked in the Update Customer Dimension Execute SQL task as noted above.  The first step is to update the housekeeping columns in the staging table to specify whether the row is new, has a Type 1 change, or a Type 2 change.  Remember that Type 1 and Type 2 changes are not mutually exclusive; you can have one, both, or neither.  We simply join the staging table to the dimension on the natural key and CurrentMember = 1 to set the housekeeping flags. 

 UPDATE stg SET
  wk_Customer = dim.wk_Customer
 ,IsNew = CASE WHEN dim.wk_Customer IS NULL 
    THEN 1 ELSE 0 END
 ,IsType1 = CASE WHEN dim.wk_Customer IS NOT NULL 
    AND stg.Type1Checksum <> dim.Type1Checksum 
     THEN 1 ELSE 0 END
 ,IsType2 = CASE WHEN dim.wk_Customer IS NOT NULL 
    AND stg.Type2Checksum <> dim.Type2Checksum
     THEN 1 ELSE 0 END
 FROM dbo.stg_dim_Customer stg
 LEFT OUTER JOIN dbo.dim_Customer dim ON 
  dim.nk_CustomerID = stg.nk_CustomerID 
  AND dim.CurrentMember = 1

The Type 1 changes are handled by updating the dimension table from staging where the IsType1 column = 1.  Note that if there are multiple rows for the natural key in the dimension, all rows will be updated.  This is typically how Type1 changes are handled but you can easily restrict the update to the current row if desired.

 UPDATE dim SET
  [ContactName] = stg.[ContactName]
 ,[ContactTitle] = stg.[ContactTitle]
 ,[Phone] = stg.[Phone]
 ,[Fax] = stg.[Fax]
 FROM dbo.stg_dim_Customer stg
 JOIN dbo.dim_Customer dim ON dim.nk_CustomerID = stg.nk_CustomerID
 WHERE IsType1 = 1

The Type 2 changes are handled by expiring the current dimension row.  The ExpirationDate is set to the ModifiedDate per the staging table less 1 minute.

 UPDATE dim SET
  CurrentMember = 0
 ,ExpirationDate = DATEADD(minute, -1, stg.ModifiedDate)
 FROM dbo.stg_dim_Customer stg
 JOIN dbo.dim_Customer dim ON dim.wk_Customer = stg.wk_Customer
 WHERE IsType2 = 1

A row is inserted into the dimension table for new rows as well as Type 2 changes.  Typically the EffectiveDate in new rows may be set to the minimum value of the datetime column as a convenience instead of the actual ModifiedDate (i.e. created date) just so that if a fact row had a transaction date before the dimension row's EffectiveDate it would still be in the range of the earliest dimension row.  The ExpirationDate is set to the maximum value of the datetime column; some folks prefer NULL which also works.

 INSERT INTO dbo.dim_Customer (
  nk_CustomerID
 ,CurrentMember
 ,EffectiveDate
 ,ExpirationDate
 ,CompanyName
 ,ContactName
 ,ContactTitle
 ,Address
 ,City
 ,Region
 ,PostalCode
 ,Country
 ,Phone
 ,Fax
 )
 SELECT
  nk_CustomerID
 ,1
 ,CASE WHEN IsNew = 1 THEN '1900-01-01' -- MIN of smalldatetime
  ELSE ModifiedDate END
 ,'2079-06-06' -- MAX of smalldatetime
 ,CompanyName
 ,ContactName
 ,ContactTitle
 ,Address
 ,City
 ,Region
 ,PostalCode
 ,Country
 ,Phone
 ,Fax
 FROM dbo.stg_dim_Customer stg
 WHERE IsType2 = 1 OR IsNew = 1

Let's take a look at an example of Type 2 processing in the dim_Customer table.  The following query results show a customer after the region has been updated.  Region is one of the columns that is handled as a Type 2 change.  As you can see a new row has been inserted with CurrentMember = 1, an EffectiveDate = the MdifiedDate when the change was processed, and an ExpirationDate which is the maximum value for a smalldatetime.  The original row was expired and its CurrentMember = 0 and ExpirationDate is set to the ModifiedDate from the source system less 1 minute.  The 1 minute subtraction eliminates any overlap in the effective date range.

query results
Next Steps
  • Stay tuned for part two of this tip; there are some additional points to discuss such as transaction handling, using different databases and/or instances for staging and the warehouse, and handling early arriving facts (aka Inferred member processing).
  • Download a copy of the sample SSIS package here to experiment with slowly changing dimensions.  Note that the package assumes you have a SQL Server instance running locally with a database called MSSQLTips.  Unzip the files into the folder C:\MSSQLTips to minimize changes to get the sample to run.


sql server categories

sql server webinars

subscribe to mssqltips

sql server tutorials

sql server white papers

next tip



About the author
MSSQLTips author Ray Barley Ray Barley is a Principal Architect at IT Resource Partners and a MSSQLTips.com BI Expert.

This author pledges the content of this article is based on professional experience and not AI generated.

View all my tips



Comments For This Article




Thursday, July 11, 2013 - 7:26:21 AM - Raymond Barley Back To Top (25791)

The beginning of the stored proc dbo.stp_UpdateCustomer has the update statement below which sets the value of wk_Customer in staging if it alreadys exists in the dimension; if the dimension does exist then a type 2 change is possible

 

-- Set flags to indicate Type 1 or Type 2 change, or New row

-- Join stg to dim on the natural key and CurrentMember = 1

UPDATE stg SET

wk_Customer = dim.wk_Customer

,IsNew = CASE WHEN dim.wk_Customer IS NULL 

THEN 1 ELSE 0 END

,IsType1 = CASE WHEN dim.wk_Customer IS NOT NULL 

AND stg.Type1Checksum <> dim.Type1Checksum 

THEN 1 ELSE 0 END

,IsType2 = CASE WHEN dim.wk_Customer IS NOT NULL 

AND stg.Type2Checksum <> dim.Type2Checksum

THEN 1 ELSE 0 END

FROM dbo.stg_dim_Customer stg

LEFT OUTER JOIN dbo.dim_Customer dim ON 

dim.nk_CustomerID = stg.nk_CustomerID 

AND dim.CurrentMember = 1


Tuesday, July 9, 2013 - 5:06:12 PM - DHRUV ARORA Back To Top (25765)

I could not understand how the following statement is valid. Everytime, staging table would be having null values in wk_Customer field.How teh staging table's wk_Cutomer field would ever match with Dim_Customer's wk_Cutomer field ? Even if I suppose it matches, how does this statement know which is the current record that has to be updated with currentMember = 0 if there are multiple historical records of the same customer.

 

UPDATE dim SET
  CurrentMember = 0
 ,ExpirationDate = DATEADD(minute, -1, stg.ModifiedDate)
 FROM dbo.stg_dim_Customer stg
 JOIN dbo.dim_Customer dim ON dim.wk_Customer = stg.wk_Customer
 WHERE IsType2 = 1


Monday, May 30, 2011 - 8:23:03 AM - Ray Barley Back To Top (13933)

I've used it to store the hash value of a password instead of the password itself.  You could use it in lieu of checksum but it's awkward having to concatenate all of the columns into a variable that you pass in to the hashbytes function.

CHECKSUM is certainly more convenient.

Since the original tip was written SQL Server 2008 came out and it supports change data capture which I typically use instead of CHECKSUM.


Monday, May 30, 2011 - 12:05:19 AM - Rich Back To Top (13932)

What are your thoughts on using hashbytes?


Sunday, May 29, 2011 - 8:07:35 AM - Ray Barley Back To Top (13931)

According to the documentation, talking about how CHECKSUM computes a value based on all of the columns in a table or a specific list of columns:

If one of the values in the expression list changes, the checksum of the list also usually changes. However, there is a small chance that the checksum will not change.

So the answer is yes it could give the wrong result.


Sunday, May 29, 2011 - 12:57:06 AM - Rich Back To Top (13930)

Is there a risk in this implementation on the checksum saying something is the same when in actuality it has changed - due to the algorithm associated with checksum()?


Tuesday, April 5, 2011 - 8:09:04 AM - Sam Back To Top (13435)

Cheers Ray,

thats what I was kind of thinking. The staging environment won't use keys between the dimensions and fact tables.

 


Tuesday, April 5, 2011 - 7:51:48 AM - Ray Barley Back To Top (13434)

The purpose of truncate is to clear out the staging table so that you can add in the rows from the source system tht have changed.  The fact staging tables don't have foreign keys that reference the dimension staging tables.  The ETL process performs lookups to supply the surrogate keys for the dimensions; the process also deals with missing dimensions (aka early arriving facts) so you don't need to rely on foreign keys for integrity.

No magic - there are no foreign keys.

 


Tuesday, April 5, 2011 - 7:38:57 AM - Sam Back To Top (13433)

Thanks for the prompt response Ray

What I mean is how can you truncate a dimension if you have foreign key constraints referencing the dimension table?

Thanks a lot


Tuesday, April 5, 2011 - 7:27:24 AM - Ray Barley Back To Top (13432)

This tip only talked about dimensions.  Fact tables would typically go through a staging process where you extract data from the source system into staging then at a minimum join the data with the dimensions via the surrogate keys.  In addition there may be other cleanup, transformations, calculations, etc. that are done in staging before populating the fact tables.  

One thing different about facts is that you usually only have inserts.  Facts don't change, they represent a historical transaction of some sort - think of an invoice.  When a purchase is made the invoice is generated.  If a customer were to return the purchase they get a new invoice; the original one doesn't change.  


Tuesday, April 5, 2011 - 4:23:50 AM - Sam Back To Top (13431)

This is an excellent article. One thing I need to know: How do I deal with fact tables? I am assuming since you are using truncate that you are not staging the fact data? Only the Dimension data? So in that case are you going to go straight from source to DW for the fact data? Much appreciated.


Monday, September 20, 2010 - 2:16:05 PM - Ray Barley Back To Top (10186)
You're referring to this T-SQL:

 UPDATE dim SET
  [ContactName] = stg.[ContactName]
 ,[ContactTitle] = stg.[ContactTitle]
 ,[Phone] = stg.[Phone]
 ,[Fax] = stg.[Fax]
 FROM dbo.stg_dim_Customer stg
 JOIN dbo.dim_Customer dim ON dim.nk_CustomerID = stg.nk_CustomerID
 WHERE IsType1 = 1

There is a subtle thing going on here; I don't recall if I pointed this out in the article.  The reason for joining on nk_CustomerID
is so that every dimension row for the customer gets updated.  In other words my definition of Type 1 is that every version of the
customer dimension gets overwritten.  So if you had say 10 versions of a customer dimension, every one of them gets overwritten
with the Type 1 changes.  Because of the Type 2 processing you can wind up with multiple rows for a particular customer.

I'm not sure everybody handles Type 1 this way but I have worked on projects where this is what the customer decided.

Monday, September 20, 2010 - 1:49:26 PM - andygarner Back To Top (10185)
I think this is a great tip - many thanks for sharing. Just one suggestion ... When updating the Type 1 fields the JOIN as published uses the natural key fields: dim.nk_CustomerID = stg.nk_CustomerID which is fine if there is only one field, but if the natural key is multipart it might be preferable to use the warehouse key: dim.wk_Customer = stg.wk_Customer (as is published for the Type 2 update).

 


Thursday, March 25, 2010 - 12:01:42 PM - toddmcdermid Back To Top (5114)

The other slow part about using the included Slowly Changing Dimension component in SSIS is that it performs a lookup for each and every row that passes through it - and those lookups are all uncached, resulting in very slow performance.

There are alternatives:

  • The T-SQL MERGE statement - lightning fast, but hard to understand
  • Coding your own in a Script component - completely customizable, difficult to implement, hard to reuse
  • Designing your own using Lookups and Conditional Splits - completely customizable, takes time to implement, impossible to reuse
  • The Kimball Method SCD Component - very fast, easy to understand

Friday, June 20, 2008 - 2:38:30 AM - raybarley Back To Top (1211)

 Are you saying that you deleted a row from the dimension table then when you rerun the package it gets inserted again from the source?  If so then you want to add a deleted flag column to the dimension.  Rather than physically deleting the row you set the deleted flag.  Then the package won't add the deleted row back.


Friday, June 20, 2008 - 12:32:50 AM - bhanurapeti Back To Top (1210)

hi

i had a problem with dealing this if i delete any row then the primary key column generating new column which is an identity column so when ever i running the package a new row is inserting can you help on this problem















get free sql tips
agree to terms