Recommended Practices for SQL Server Analysis Services 2005/2008 Design
- Recommended Practices for Dimensions
- Recommended Practices with Partitions and Aggregations
Microsoft’s SQL Server Analysis Services (MSAS) 2005 was a revolutionary release that dramatically changed how software developers and database administrators build and manage business intelligence (BI) solutions. A huge number of new features and changes introduced with this release instigated a steep learning curve even for those IT professionals who felt comfortable building cubes with previous releases of software.
MSAS 2008, on the other hand, is an evolutionary release building on the strong foundation laid by 2005 version, fixing some issues and introducing a few additional features. By now most BI folks know how to build Analysis Services cubes; however, documentation and literature for troubleshooting and performance tuning is still sparse. This article attempts to cover some gotchas that apply to large-scale MSAS implementations where both processing and querying performance is of paramount importance to the application’s success.
Tuning Analysis Services could involve many aspects, including but not limited to hardware configuration, dimension and partition design, re-engineering MDX queries and processing XMLA statements, and tweaking MSAS configuration settings. In this article, I focus on most commonly encountered design pitfalls. The article presumes that the readers are familiar with Analysis Services terms and concepts such as measures, dimensions, measure groups, cubes, and so on.
Recommended Practices for Dimensions
The following sections offer recommendations for working with dimensions.
#1: Limit the Number of Dimensions Per Measure Group and Number of Attributes Per Dimension
Although it may be tempting to build a data warehousing solution that encompasses all business data and allows examining such data from every possible angle, the “kitchen sink” approach doesn’t always work. In fact, business users who demand such flexibility often become confused if a query filters data by more than a few dimensions. The latest releases of Analysis Services allow more than one fact table and more than one measure group per cube, so each cube could easily support dozens of dimensions. However, for best performance and manageability, try to limit each measure group to fewer than 20 dimensions.
Furthermore, be sure to also limit the number attributes in each dimension to only those attributes that are necessary for analytics. Dimension tables in the relational data source may have numerous columns that may be used for transaction level reporting, or could simply be obsolete and left over from legacy implementations of the data warehouse. If you know that a given column will never be used for analytics, do not add it to the Analysis Services dimension structure. If the column will be seldom used, be sure to set AttributeHierarchyOptimizedState property to “Not Optimized” for this attribute. MSAS will not build indexes for such attributes, thereby saving storage space and processing time.
#2: Use Dimension Properties Effectively
Each dimension has a number of properties that need to be handled with care. For large dimensions that expose millions of rows and have a large number of attributes, pay particular attention to the ProcessingGroup property. By default, this property is assigned a value of ByAttribute. During processing, Analysis Services issues a separate SELECT DISTINCT … query for each dimension attribute. Running multiple expensive SELECT DISTINCT queries could place a significant load on the relational database engine and can be time-consuming. If you change the ProcessingGroup property to ByTable, MSAS will only execute a single query without the DISTINCT keyword to populate the dimension structure.
Note that the ByTable option may use more memory on the Analysis Services host compared to ByAttribute; nevertheless, if processing performance is important and you have plenty of memory (as you should for large-scale implementations), be sure to experiment with this property to see which value works best for you. Unfortunately, there is no hard and fast rule.
#3: Use Regular Dimension Relationship Whenever Possible
Latest versions of MSAS give us the flexibility of using referenced, many-to-many, fact, and other types of relationships between fact and dimension tables. However, this does not mean that star and snowflake schemas are obsolete or that you should be building a data warehouse using the third normal form. Clearly if you don’t manage the relational data source and don’t have the flexibility of building a star schema, then you may have no option but to use the data warehouse in its current shape. However, if you care about performance, try to stick with a regular relationship whereby fact tables have foreign keys to each dimension table’s primary key. If you must use a referenced relationship (snowflake schema), then be sure to check the Materialize check box in the Dimension Relationship dialog. By doing so, you advise Analysis Services to build additional indexes for resolving the relationship/join from fact table to referenced (intermediate) dimension table to the primary dimension table. MSAS would have to join partition records to referenced dimension dynamically during query execution for un-materialized dimensions.
Fact dimensions (sometimes referred to as degenerate dimensions) have a valid business use. For example, we often need to expose invoice numbers or customer identifiers on reports. I suggest you test performance of such reports with Analysis Services and compare to the same report generated directly from the relational data source. Analysis Services is still best suited for its original intended purposeaggregating and presenting summarized dataas opposed to reporting transaction-level details. For huge transaction-level reports, you may find that SQL queries will perform just as well or better than reports created with Analysis Services. See the next recommended practice for more tuning advice concerning fact-level dimensions.
Many-to-many dimensions are necessary for several common business scenarios; however, this flexibility comes with a performance penalty. Investigate the possibility of alternative implementation without using many-to-many relationships. If you have to use a many-to-many relationship, be sure to read and follow optimization guidelines in this Microsoft article.
#4: Use Integer Data Type for Attribute Keys If at All Possible
For best processing performance, try to use the smallest data type for attribute key columns. Although you could use string data types for key columns, they do not perform nearly as well as the numeric data types.
At times you may need to use multiple columns for the composite key for a dimension attribute. Although this practice is almost always acceptable, for large dimensions (and particularly for fact level dimensions) you may have to add an alternate key to the dimension table to ensure processing is efficient. For example, let’s suppose you have a fact level “transaction” attribute and you use company identifier, effective date, and journal entry identifier string columns as keys for this attribute. While processing this dimension, MSAS must retrieve a unique combination of three string columns, which can be very time-consuming for a dimension that contains millions of records. Instead, you can add an identity column (or another integer column uniquely identifying each transaction) to the dimension table and use it as the attribute key. Reading a single integer column will be considerably faster than reading three string columns. Storing integer keys will also be much more efficient.
#5: Use Natural Hierarchies When Possible
I trust that you have already heard this piece of advice because it has been heavily stressed in a number of white papers and presentations. Sadly, there are still developers out there who do not follow this rule, so it makes sense to repeat it here. Each dimension can have three types of hierarchies:
- Attribute hierarchies, which expose a single attributefor example, product price
- User hierarchies, which group multiple unrelated attributesfor example, product category → product color
- Natural hierarchies, where each attribute is related either directly or indirectly to all other attributes in the same hierarchy, as in product category → product subcategory → product name
Defining attribute relationships helps with both processing and querying performance, particularly when users drill down from the top to the intermediate level, and then to the lower level of the hierarchy. As a rule of thumb, you should have at least one natural hierarchy in every dimension that does not include parent-child relationships. Parent-child dimensions expose self-referencing relationships, such as employee-manager, and normally do not require any additional hierarchies.