SQL or M? – SSAS Partitions

Part 1: Using SQL Server Table-Valued Functions (UDFs)

In SQL Server Analysis Services projects, as of SQL Server Data Tools 2017, you can define table partitions using Power Query.  Of course, we still have the option to use SQL Server database objects like views or user-defined functions.  So, which of these two option makes most sense?  The same concepts and decision points apply to Power BI data models although the design experience is quite a bit different.

The following steps will bring us to a question: Using the new SSDT partition design method for SSAS 2017, should I define partition filtering logic in SQL or in Power Query/M?

The objective is to define three partitions in the data model for the Sales fact table in the ContosoDW database:

  • New transactions added in the current week
  • Adjusting entries for the current month
  • Historic records prior to the current month

New sales transactions in the source database needs to be refreshed in the data model every hour for reporting.  Reprocessing only the records since the beginning of the current week takes seconds to minutes.  If we schedule that partition to refresh every hour, users can have up-to-date reports throughout the day.  In addition to new transactions, adjusting records are made weekly but only to records in the current month before the end-of-month closing of the books.  Records in the current month that are older than the current week might be updated on occasion but changes don’t need to be available until the weekend.  Records older than a month rarely change and don’t need to be refreshed but once a month.  By scheduling only the first or second partition to process, data can be updated without requiring tens of millions of historical records to be reloaded.

Partitioning with a SQL User-Defined Function

I’ll step through the more conventional method we’ve been using for many years.  I’ve written the following T-SQL table-valued User-Defined Function named fnSalesPartitionForPeriod.  Three possible input parameter values allow the function to return rows for the past week, for the past month (up to the past week) or for all dates previous to the current month.

Here is the T-SQL script for a table-valued user-defined function created in SQL Server.  Passing in one of three parameter values will cause it to return the desired records.

/******************************
     User-defined function used to partition Sales fact table in SSAS tabular model
     @Period values:
         PriorToThisMonth
         ThisMonthPriorToThisWeek
         ThisWeek
*******************************/
create function dbo.fnSalesPartitionForPeriod
     ( @Period varchar(100) )
returns table
return
     select * from [ContosoDW].[dbo].[FactSalesCompleteDates]
     where
     (@Period = ‘BeforeThisMonth’
         and
         [DateKey] < dateadd(month, datediff(month, 0, getdate()), 0)
     )
     or
     (@Period = ‘ThisWeek’
         and
         [DateKey] >= dateadd(week, datediff(week, 0, getdate()), 0)
     )
     or
     (@Period = ‘ThisMonthBeforeThisWeek’
         and
         [DateKey] >= dateadd(month, datediff(month, 0, getdate()), 0)
         and
         [DateKey] < dateadd(week, datediff(week, 0, getdate()), 0)
     )
;
go

To create the three Sales table partitions using this UDF, I start by importing one table.  Here’s the Import Table dialog for the new Sales table in the data model.  I’ve selected the new UDF and entered the parameter value ‘BeforeThisMonth’ to define the first partition.

image

This part gets tricky and quite honestly, I rarely get the steps right the first time through.  I haven’t quite decided yet if my routine struggle with the SSDT Power Query editor are because I expect it work like it does in Power BI Desktop or if it truly has some quirks that catch me off guard.  Regardless, I’m cautious to save copies of my work and if something doesn’t work, I delete the query and repeat the steps.

The query editor was smart enough to create an M function from the UDF query and this function needs to be invoked to generate the new Sales table.  Enter the parameter value once again and click the Invoke button. 

image

Change the name of the new query to “Sales” and make sure that the query is set to “Create New Table”, then click the Import button on the toolbar.

image

After the table is imported, click the Partitions button on the SSDT toolbar.  As you can see, the Power Query “M” script for the Sales table calls the function and passes the parameter value I had set.  This default partition should be renamed and the other two partitions should be added using different parameter values.

image

Updating and adding the partitions is fairly simple, using these steps:

  1. Copy the original partition
  2. Rename the new partition
  3. Change the function parameter value

Rename the current partition with a friendly name.   Clicking the Copy button twice gives me two copies of the parameter.  You can see that I’ve commented the code with the valid UDF parameter values.

SNAGHTML3791b15

Now the table can be refreshed incrementally and only new transaction records for the current week or month can be updated during schedule refresh cycles.

Partitioning with Power Query

No matter what the data source is; whether you use table-valued UDFs, views or in-line SQL, you are still using Power Query to define tables – so why not just use Power Query without creating database objects?

In another post, I’ll repeat the exercise using only Power Query to define the same partitions.  Stay tuned.

Hands-on Workshops at the Pacific Northwest Power BI Symposium

Please join Power BI authors and community leaders for an afternoon and evening of deep learning.  Featured presenters include “Guy In A Cube” Adam Saxton and international author and Excel MVP, Matt Allington.  Deepen your skills with Power BI and get a recap from the Microsoft Business Applications Summit held in Seattle earlier in the week.

Jul 26, 2018 from 3:00 PM to 8:30 PM Portland, OR

Associated with  Portland Power BI User Group

Event Image

AGENDA –

3:00 ; check-in and registration

3:45 ; guide guests to workshops

4:00 ; workshops commence, ONLY REGISTER FOR ONE:

workshop 1 – TRISTAN MALHERBE (AZEO) – Power Query to Create a Calendar Table

register here: https://bit.ly/2K3U0lT

workshop 2 – PAUL TURLEY (CSG Pro) – Model and Visualize Financial and Accounting Data with Power BI and Excel

register here: https://bit.ly/2JOJy5v

5:00 ; workshops wrap up, seating in main event area (dinner served / networking)

5:35 ; presentation intro – RON ELLIS GAUT (CSG Pro / Portland Power BI User Group Leader)

6:00 ; presentation 1 – BRIAN GRANT (CSG Pro) – Shining a New Light on Calculate

6:45 ; presentation 2 – ADAM SAXTON (Microsoft / Guy in a Cube) – Business Applications Summit Recap

7:30 ; presentation 3 – MATT ALLINGTON (Excelerator BI) – DAX as a Query Language

8:15 ; closing remarks

8:30 PM ; event ends

Questions?  Contact Gregory Petrossian: gregp@csgpro.com for sponsorship options

SQL, M or DAX: When Does it Matter?

 

Column-based calculations are part of every BI project. Some of the most common examples include building a street address column from individual fields, concatenating a person’s full name from First, Middle and Last Name fields; or creating a location string from City, State and Country fields. More complex examples might require a lookup or join operation to get a reference value used in a calculation that is then stored as a column on each entity record.  Keep in mind that we are strictly talking about calculated column values that are stored for each row and not dynamic calculations that run in the context of filters and slicers.  Those are measures and that is a separate topic.

It Depends, or It Matters (one or the other)

If the data source is a relational database that supports queries, should you perform these calculations in SQL, Power Query “M”, or in a DAX calculated column? The standard tongue-in-cheek answer from a consultant is usually “it depends”. That was the answer I grew up with, but apparently popular language has changed in the last generation. My kids, who are all now young adults, say “it matters”. Back in the day, if I said “Hey, Dad. Can we get ice cream on the way to the store?” He would say, “It depends on whether you get your chores done.” My kids would say “it matters whether I get back from the beach on time.” So, it either depends or it matters, I guess.

Self-service BI is all about having the freedom to create reports that make an impact and bring important value to business users and leaders.  When importing, shaping and modeling data; if we can get simple and mundane tasks out of the way, this leaves time and energy to move on to more important things.  If you can just get the core table structures in-place; with unique keys, calculated columns, and numeric columns for summaries and aggregate measures, you can design the more impactful bits of the solution to support the report design.

Many BI projects start out the same way, with aspirations to import data from several different sources, to work-out the complexities of cleansing and matching records in various tables to create a nice uniform data model used to build all kinds of beautiful dashboards and interactive reports.  Our optimism about making quick progress at the beginning of the project is often squelched when we realize that the data source for a lookup table isn’t reliable, and that the system of record is an application controlled by a different business group in some remote corner of the organization. The data is in a different format, access is restricted and the person in charge of managing it is on extended leave. We get caught-up in the complications of just getting essential data into the model and then deliver far less than expected. I can’t tell you how common this scenario is – especially in larger projects.

For calculated columns that end up stored in a data model table, there is rarely a difference in performance, storage or report query speed based on the technique used to calculate the column value. In cases where there is a technical advantage, the decision should be clear – use the most optimal method that is feasible. In the majority of cases where there is no strong technical argument for one method or the other, use the method that simplifies development and maintenance, and offers more control.

You should have a standard method for managing calculated columns, so you know how to maintain them down the road. This might seem trivial so why does it matter so much?

The data model schema is the foundation for your reporting solution and making changes after the rest of the solution is designed can be catastrophic if you don’t plan and manage future changes. A semantic data model is literally a house of cards. Deleting, renaming or changing the data type for a column could break every calculation and the report visual referencing that column. Whether you should create these calculated values in a source query using hand-written SQL or a database object like a view or user-defined function, in Power Query or as a calculated column using DAX – that will depend on who maintains the Power BI or SSAS model and who should manage the design in the future.

SQL and Database Objects

As a general rule of thumb, in formal SSAS projects built on a relational data mart or data warehouse that is managed by the same project team as the BI data model, I typically recommend that every table in the model import data from a corresponding view or UDF stored and managed in the relational database. Keep in mind that is the way we’ve been designing Microsoft BI projects for several years. Performing simple tasks like renaming columns in the SSAS data model designer was slow and cumbersome. Performing this part of the data prep in T-SQL was much easier than in SSDT. With the recent advent of Power Query in SQL Server Data Tools, there is a good argument to be made for managing those transformations but the tool is still new and frankly I’m still testing the water. Again, keep changes in one place for future maintenance.

Do your absolute best to avoid writing complex SQL query logic that cannot be traced back to the sources. Complicated queries can become a black box – and a Pandora’s box if they aren’t documented, annotated and easy to decipher.

Power Query/M

For less formal projects in Power BI data models, Power Query is king. We’ve never had a tool so flexible and easy to use. If I’m importing data from multiple sources into a single model, you bet I’m going to use Power Query instead of SQL queries because I’ll know where to find and manage all the query definitions.

DAX Calculated Columns

Why not use DAX calculated columns? There is a good argument for using DAX. It’s quick and easy, and sometime more convenient. If I add a custom column to a multi-million row table defined in Power Query, I have to re-process the table to see the new column. If I use DAX, I don’t have to wait. If the calculation relies on a DAX calculation residing in the rest of the model, DAX is the clear winner. These cases are less common though. Once again, I’ll make the argument to manage calculations, as much as possible, in one place.

IT Process, Business Culture, Team Dynamics, Rules & Restrictions

Now that you have some clear criteria for always implementing column calculation in either SQL, Power Query or DAX; let me inject some reality back into the “it depends or it matters” equation.

I do my best to put cynicism aside and focus on what it takes to get IT projects over the finish line. I’ve found that good BI project practitioners positive, optimistic and tough skinned; although there are many forces at work to change this disposition. If you know what I mean, no further explanation is needed. If you don’t, you will. Wherever you choose to work, just do what you can to maintain your perspective throughout your career.

I’ll give one example that represents situations I’ve encountered on several larger, formal BI projects over the years:

The BI project Architect, Database Administrator, Lead Developer and IT Director all agree that any schema dependencies on the data warehouse or data mart should be managed using database views. This is a paramount rule in the solution architecture. The SSAS data model and Power BI data model developers should import tables from these views. The team is using a pure Agile mythology and will use JIRA to manage and assign tasks performed in two-week team sprints.

Based on high-level report requirements documented by the Business Systems Analyst, the Lead Architect creates views in the database build script. The ETL developer must stage the source data and then the data warehouse ETL developer populates the dimension table before the view can be created, which takes 3 sprints or six weeks. The Power BI data model developer adds the table to the model in the 4th sprint. After a prototype report is created, the BSA gets feedback from a stakeholder user who tells us that customer names should be in a single column rather than separate first name and last name columns. A task is added in JIRA to modify the view with another task to refresh the data model, so it takes two weeks to add the column.

After the data warehouse is in production, the Power BI report developer gets word that the customer city, state and zip code need to be concatenated into a single column and marked as a geographical location, so they can be used in a map visual. The data warehouse is in production and managed by an offshore DBA group. A support ticket is created to request that the view used to populate the customer table be altered and a CustomerLocation column be added. Three days later, the contracted help desk determines this is not in their area of responsibility and closes the ticket as “completed” while the model developer continues to wait for a call or email. The email goes to the IT Director, who happily dismisses it since it was marked as “completed”. Two weeks later, the issue resurfaces and the Project Manager organizes a meeting with the IT Director, BI Lead Architect, BI Lead Developer, Database Developer, In-house DBA and Help Desk Contractor Liaison to resolve the issue. In the meantime, users have exported their report to Excel and are working around the issue using a copy of the data.

A month after the request, the Power BI Developer spends 2 minutes creating a DAX calculated column and then creates the map report.