Expert Cube Development with Microsoft SQL Server 2008 Analysis Services — Save 50%
Design and implement fast, scalable and maintainable cubes with Microsoft SQL Server 2008 Analysis Services with this book and eBook
In the previous part of this article by Chris Webb, we covered useful properties of measures, along with built-in measure aggregation types and dimension calculations. In this part, we will focus on measure groups and dimension/measure group relationships.
All but the simplest data warehouses will contain multiple fact tables, and Analysis Services allows you to build a single cube on top of multiple fact tables through the creation of multiple measure groups. These measure groups can contain different dimensions and be at different granularities, but so long as you model your cube correctly, your users will be able to use measures from each of these measure groups in their queries easily and without worrying about the underlying complexity.
Creating multiple measure groups
To create a new measure group in the Cube Editor, go to the Cube Structure tab and right-click on the cube name in the Measures pane and select 'New Measure Group'. You'll then need to select the fact table to create the measure group from and then the new measure group will be created; any columns that aren't used as foreign key columns in the DSV will automatically be created as measures, and you'll also get an extra measure of aggregation type Count. It's a good idea to delete any measures you are not going to use at this stage.
Once you've created a new measure group, BIDS will try to set up relationships between it and any existing dimensions in your cube based on the relationships you've defined in your DSV. Since doing this manually can be time-consuming, this is another great reason for defining relationships in the DSV. You can check the relationships that have been created on the Dimension Usage tab of the Cube Editor:
In Analysis Services 2005, it was true in some cases that query performance was better on cubes with fewer measure groups, and that breaking a large cube with many measure groups up into many smaller cubes with only one or two measure groups could result in faster queries. This is no longer the case in Analysis Services 2008. Although there are other reasons why you might want to consider creating separate cubes for each measure group, this is still something of a controversial subject amongst Analysis Services developers. The advantages of a single cube approach are:
- All of your data is in one place. If your users need to display measures from multiple measure groups, or you need to create calculations that span measure groups, everything is already in place.
- You only have one cube to manage security and calculations on; with multiple cubes the same security and calculations might have to be duplicated.
The advantages of the multiple cube approach are:
- If you have a complex cube but have to use Standard Edition, you cannot use Perspectives to hide complexity from your users. In this case, creating multiple cubes might be a more user-friendly approach.
- Depending on your requirements, security might be easier to manage with multiple cubes. It's very easy to grant or deny a role access to a cube; it's much harder to use dimension security to control which measures and dimensions in a multi-measure group cube a role can access.
- If you have complex calculations, especially MDX Script assignments, it's too easy to write a calculation that has an effect on part of the cube you didn't want to alter. With multiple cubes, the chances of this happening are reduced.
Creating measure groups from dimension tables
Measure groups don't always have to be created from fact tables. In many cases, it can be useful to build measure groups from dimension tables too. One common scenario where you might want to do this is when you want to create a measure that counts the number of days in the currently selected time period, so if you had selected a year on your Time dimension's hierarchy, the measure would show the number of days in the year. You could implement this with a calculated measure in MDX, but it would be hard to write code that worked in all possible circumstances, such as when a user multi-selects time periods. In fact, it's a better idea to create a new measure group from your Time dimension table containing a new measure with AggregateFunction Count, so you're simply counting the number of days as the number of rows in the dimension table. This measure will perform faster and always return the values you expect. This post on Mosha Pasumansky's blog discusses the problem in more detail: http://tinyurl.com/moshadays
MDX formulas vs pre-calculating values
If you can somehow model a calculation into the structure of your cube, or perform it in your ETL, you should do so in preference to doing it in MDX only so long as you do not compromise the functionality of your cube. A pure MDX approach will be the most flexible and maintainable since it only involves writing code, and if calculation logic needs to change, then you just need to redeploy your updated MDX Script; doing calculations upstream in the ETL can be much more time-consuming to implement and if you decide to change your calculation logic, then it could involve reloading one or more tables. However, an MDX calculation, even one that is properly tuned, will of course never perform as well as a pre-calculated value or a regular measure. The day count measure, discussed in the previous paragraph, is a perfect example of where a cube-modeling approach trumps MDX. If your aim was to create a measure that showed average daily sales, though, it would make no sense to try to pre-calculate all possible values since that would be far too time-consuming and would result in a non-aggregatable measure. The best solution here would be a hybrid: create real measures for sales and day count, and then create an MDX calculated measure that divided the former by the latter. However, it's always necessary to consider the type of calculation, the volume of data involved and the chances of the calculation algorithm changing in the future before you can make an informed decision on which approach to take.
Handling different dimensionality
When you have different measure groups in a cube, they are almost always going to have different dimensions associated with them; indeed, if you have measure groups that have identical dimensionality, you might consider combining them into a single measure group if it is convenient to do so. As we've already seen, the Dimension Usage tab shows us which dimensions have relationships with which measure groups.
When a dimension has a relationship with a measure group it goes without saying that making a selection on that dimension will affect the values that are displayed for measures on that measure group. But what happens to measures when you make a selection on a dimension that has no relationship with a measure group? In fact, you have two options here, controlled by the IgnoreUnrelatedDimensions property of a measure group:
- IgnoreUnrelatedDimensions=False displays a null value for all members below the root (the intersection of all of the All Members or default members on every hierarchy) of the dimension, except the Unknown member, or
- IgnoreUnrelatedDimensions=True repeats the value displayed at the root of the dimension for every member on every hierarchy of the dimension. This is the default state.
The screenshot below shows what happens for two otherwise identical measures from measure groups which have IgnoreUnrelatedDimensions set to True and to False when they're displayed next to a dimension they have no relationship with:
It's usually best to keep IgnoreUnrelatedDimensions set to True since if the users are querying measures from multiple measure groups, then they don't want some of their selected measures suddenly returning null if they slice by a dimension that has a regular relationship with their other selected measures.
Handling different granularities
Even when measure groups share the same dimensions, they may not share the same granularity. For example, we may hold sales information in one fact table down to the day level, but also hold sales quotas in another fact table at the quarter level. If we created measure groups from both these fact tables, then they would both have regular relationships with our Time dimension but at different granularities.
Normally, when you create a regular relationship between a dimension and a measure group, Analysis Services will join the columns specified in the KeyColumns property of the key attribute of the dimension with the appropriate foreign key columns of the fact table (note that during processing, Analysis Services won't usually do the join in SQL, it does it internally). However, when you have a fact table of a higher granularity, you need to change the granularity attribute property of the relationship to choose the attribute from the dimension you do want to join on instead:
In the previous screenshot, we can see an amber warning triangle telling us that by selecting a non-key attribute, the server may have trouble aggregating measure values. What does this mean exactly? Let's take a look at the attribute relationships defined on our Time dimension again:
If we're loading data at the Quarter level, what do we expect to see at the Month and Date level? We can only expect to see useful values at the level of the granularity attribute we've chosen, and for only those attributes whose values can be derived from that attribute; this is yet another good reason to make sure your attribute relationships have been optimized. Below the granularity attribute, we've got the same options regarding what gets displayed as we had with dimensions that have no relationship at all with a measure group: either repeated values or null values. The IgnoreUnrelatedDimensions property is again used to control this behavior.
Unfortunately, the default True setting for IgnoreUnrelatedDimensions is usually not the option you want to use in this scenario (users usually prefer to see nulls below the granularity of a measure in our experience) and this may conflict with how we want to set IgnoreUnrelatedDimensions to control the behavior of dimensions which have no relationship with a measure group. There are ways of resolving this conflict such as using MDX Script assignments to set cell values to null or by using the ValidMeasure() MDX function, but none are particularly elegant.
Non-aggregatable measures: a different approach
We've already seen how we can use parent/child hierarchies to load non-aggregatable measure values into our cube. However, given the problems associated with using parent/child hierarchies and knowing what we now know about measure groups, let's consider a different approach to solving this problem.
A non-aggregatable measure will have, by its very nature, data stored for many different granularities of a dimension. Rather than storing all of these different granularities of values in the same fact table, we could create multiple fact tables for each granularity of value. Having built measure groups from these fact tables, we would then be able to join our dimension to each of them with a regular relationship but at different granularities.
We'd then be in the position of having multiple measures representing the different granularities of a single, logical measure. What we actually want is a single non-aggregatable measure, and we can get this by using MDX Script assignments to combine the different granularities. Let's say we have a regular (non-parent/child) dimension called Employee with three attributes Manager, Team Leader and Sales Person, and a logical non-aggregatable measure called Sales Quota appearing in three measure groups as three measures called Sales Amount Quota_Manager, Sales Amount Quota_TeamLead and Sales Amount Quota for each of these three granularities. Here's a screenshot showing what a query against this cube would show at this stage:
We can combine the three measures into one like this:
SCOPE([Measures].[Sales Amount Quota]);
THIS=[Measures].[Sales Amount Quota_TeamLead];
THIS=[Measures].[Sales Amount Quota_Manager];
This code takes the lowest granularity measure Sales Amount Quota, and then overwrites it twice: the first assignment replaces all of the values above the Sales Person granularity with the value of the measure containing Sales Amount Quota for Team Leaders; the second assignment then replaces all of the values above the Team Leader granularity with the value of the measure containing Sales Quotas for Managers. Once we've set Visible=False for the Sales Amount Quota_TeamLead and Sales Amount Quota_Manager measures, we're left with just the Sales Amount Quota measure visible, thus displaying the non-aggregatable values that we wanted. The user would then see this:
Using linked dimensions and measure groups
Creating linked dimensions and measure groups allows you to share the same dimensions and measure groups across separate Analysis Services databases, and the same measure group across multiple cubes. To do this, all you need to do is to run the 'New Linked Object' wizard from the Cube Editor, either by clicking on the button in the toolbar on the Cube Structure or Dimension Usage tabs, or by selecting it from the right-click menu in the Measures pane of the Cube Structure tab.
Doing this has the advantage of reducing the amount of processing and maintenance needed: instead of having many identical dimensions and measure groups to maintain and keep synchronized, all of which need processing separately, you can have a single object which only needs to be changed and processed once. At least that's the theory—in practice, linked objects are not as widely used as they could be because there are a number of limitations in their use:
- Linked objects represent a static snapshot of the metadata of the source object, and any changes to the source object are not passed through to the linked object. So for example, if you create a linked dimension and then add an attribute to the source dimension, you then have to delete and recreate the linked dimension—there's no option to refresh a linked object.
- You can also import the calculations defined in the MDX Script of the source cube using the wizard. However, you can only import the entire script and this may include references to objects present in the source cube that aren't in the target cube, and which may need to be deleted to prevent errors. The calculations that remain will also need to be updated manually when those in the source cube are changed, and if there are a lot, this can add an unwelcome maintenance overhead.
- A linked measure group can only be used with dimensions from the same database as the source measure group. This isn't a problem when you're sharing measure groups between cubes in the same database, but could be if you wanted to share measure groups across databases.
- As you would expect, when you query a linked measure group, your query is redirected to the source measure group. If the source measure group is on a different server, this may introduce some latency and hurt query performance. Analysis Services does try to mitigate this by doing some caching on the linked measure group's database, though. By default, it will cache data on a per-query basis, but if you change the RefreshPolicy property from ByQuery to ByInterval you can specify a time limit for data to be held in cache.
Linked objects can be useful when cube development is split between multiple development teams, or when you need to create multiple cubes containing some shared data, but, in general, we recommend against using them widely because of these limitations.
It's also possible to add the same dimension to a cube more than once, and give each instance a different relationship to the same measure group. For example, in our Sales fact table, we might have several different foreign key columns that join to our Time dimension table: one which holds the date an order was placed on, one which holds the date it was shipped from the warehouse, and one which holds the date the order should arrive with the customer. In Analysis Services, we can create a single physical Time dimension in our database, which is referred to as a database dimension, and then add it three times to the cube to create three 'cube dimensions', renaming each cube dimension to something like Order Date, Ship Date and Due Date. These three cube dimensions are referred to as role-playing dimensions: the same dimension is playing three different roles in the same cube.
Role playing dimensions are a very useful feature. They reduce maintenance overheads because you only need to edit one dimension, and unlike linked dimensions, any changes made to the underlying database dimension are propagated to all of the cube dimensions that are based on it. They also reduce processing time because you only need to process the database dimension once. However, there is one frustrating limitation with role-playing dimensions and that is that while you can override certain properties of the database dimension on a per-cube dimension basis, you can't change the name of any of the attributes or hierarchies of a cube dimension. So if you have a user hierarchy called 'Calendar' on your database dimension, all of your cube dimensions will also have a user hierarchy called 'Calendar', and your users might find it difficult to tell which hierarchy is which in certain client tools (Excel 2003 is particularly bad in this respect) or in reports. Unfortunately, we have seen numerous cases where this problem alone meant role-playing dimensions couldn't be used.
Dimension/measure group relationships
So far we've seen dimensions either having no relationship with a measure group or having a regular relationship, but that's not the whole story: there are many different types of relationships that a dimension can have with a measure group. Here's the complete list:
- No relationship
- Data Mining
Fact or degenerate dimensions are dimensions that are built directly from columns in a fact table, not from a separate dimension table. From an Analysis Services dimension point of view, they are no different from any other kind of dimension, except that there is a special fact relationship type that a dimension can have with a measure group. There are in fact very few differences between a fact relationship and a regular relationship, and they are:
- A fact relationship will result in marginally more efficient SQL being generated when the fact dimension is used in ROLAP drillthrough.
- Fact relationships are visible to client tools in the cube's metadata, so client tools may choose to display fact dimensions differently.
- A fact relationship can only be defined on dimensions and measure groups that are based on the same table in the DSV.
- A measure group can only have a fact relationship with one database dimension. It can have more than one fact relationship, but all of them have to be with cube dimensions based on the same database dimension.
It still makes sense though to define relationships as fact relationships when you can. Apart from the reasons given above, the functionality might change in future versions of Analysis Services and fact relationship types might be further optimized in some way.
A referenced relationship is where a dimension joins to a measure group through another dimension. For example, you might have a Customer dimension that includes geographic attributes up to and including a customer's country; also, your organization might divide the world up into international regions such as North America, Europe, Middle East and Africa (EMEA), Latin America (LATAM) and Asia-Pacific and so on for financial reporting, and you might build a dimension for this too. If your sales fact table only contained a foreign key for the Customer dimension, but you wanted to analyze sales by international region, you would be able to create a referenced relationship from the Region dimension through the Customer dimension to the Sales measure group.
When setting up a referenced relationship in the Define Relationship dialog in the Dimension Usage tab, you're asked to first choose the dimension that you wish to join through and then which attribute on the reference dimension joins to which attribute on the intermediate dimension:
When the join is made between the attributes you've chosen on the reference dimension, once again it's the values in the columns that are defined in the KeyColumns property of each attribute that you're in fact joining on.
The Materialize checkbox is automatically checked, and this ensures maximum query performance by resolving the join between the dimensions at processing time, which can lead to a significant decrease in processing performance. Unchecking this box means that no penalty is paid at processing time but query performance may be worse.
The question you may well be asking yourself at this stage is: why bother to use referenced relationships at all? It is in fact a good question to ask, because, in general, it's better to include all of the attributes you need in a single Analysis Services dimension built from multiple tables rather than use a referenced relationship. The single dimension approach will perform better and is more user-friendly: for example, you can't define user hierarchies that span a reference dimension and its intermediate dimension.
That said, there are situations where referenced relationships are useful because it's simply not feasible to add all of the attributes you need to a dimension. You might have a Customer dimension, for instance, that has a number of attributes representing dates—the date of a customer's first purchase, the date of a customer's tenth purchase, the date of a customer's last purchase and so on. If you had created these attributes with keys that matched the surrogate keys of your Time dimension, you could create multiple, referenced (but not materialized) role-playing Time dimensions joined to each of these attributes that would give you the ability to analyze each of these dates. You certainly wouldn't want to duplicate all of the attributes from your Time dimension for each of these dates in your Customer dimension. Another good use for referenced relationships is when you want to create multiple parent/child hierarchies from the same dimension table
Data mining relationships
The data mining functionality of Analysis Services is outside the scope of this article, so we won't spend much time on the data mining relationship type. Suffice to say that when you create an Analysis Services mining structure from data sourced from a cube, you have the option of using that mining structure as the source for a special type of dimension, called a data mining dimension. The wizard will also create a new cube containing linked copies of all of the dimensions and measure groups in the source cube, plus the new data mining dimension, which then has a data mining relationships with the measure groups.
In this part, we focused on how to create new measure groups and handle the problems of different dimensionality and granularity, and looked at the different types of relationships that are possible between dimensions and measure groups.
|Design and implement fast, scalable and maintainable cubes with Microsoft SQL Server 2008 Analysis Services with this book and eBook|
eBook Price: $29.99
Book Price: $49.99
About the Author :
Chris Webb (firstname.lastname@example.org) has been working with Microsoft Business Intelligence tools for 15 years in a variety of roles and industries. He is an independent consultant (www.crossjoin.co.uk) and trainer (www.technitrain.com) based in the UK, specializing in SQL Server Analysis Services, MDX, DAX, Power Pivot, and the whole Power BI stack. He is the co-author of MDX Solutions with Microsoft SQL Server Analysis Services 2005 and SQL Server Analysis Services 2012: The BISM Tabular Model. He is a regular speaker at user groups and conferences, and blogs about Microsoft BI at http://cwebbbi.wordpress.com/.