watching Jeffrey Wang as a live stream guest with Reid Havens, and one of many dozen great issues that Jeffrey shared with the viewers was the checklist of optimizations that the DAX engine performs when creating an optimum question plan for our measures.
And, the one which caught my consideration was relating to the so-called “Sparse measures”:
To make it easy, when you outline the measure, Formula Engine in VertiPaq will add an implicit NonEmpty filter to the question, which ought to allow the optimizer to keep away from full cross-join of dimension tables and scan solely these rows the place data for the mixture of your dimension attributes actually exist. For people coming from the MDX world, the NonEmpty perform might look acquainted, however let’s see the way it works in DAX.
The factor that almost all resonated with me was when Jeffrey suggested in opposition to changing BLANKs with zeroes (or no matter specific values) in Energy BI calculations. I’ve already written how you can handle BLANKs and replace them with zeroes, however on this article, I wish to concentrate on the doable efficiency implications of this choice.
Setting the stage
Earlier than we begin, one necessary disclaimer: the advice to not change BLANK with 0 is simply that — a advice. If the enterprise request is to show 0 as an alternative of BLANK, it doesn’t essentially imply that it’s best to refuse to do it. In most eventualities, you’ll in all probability not even discover a efficiency lower, however it should rely upon a number of various factors…
Let’s begin by writing our easy DAX measure:
Gross sales Amt 364 Merchandise =
CALCULATE (
[Sales Amt],
FILTER ( ALL ( 'Product'[ProductKey] ), 'Product'[ProductKey] = 364 )
)Utilizing this measure, I wish to calculate the entire gross sales quantity for the product with ProductKey = 364. And, if I put the worth of this measure within the Card visible, and activate Efficiency Analyzer to test the instances for dealing with this question, I get the next outcomes:
DAX question took solely 11ms to execute, and as soon as I switched to DAX Studio, the xmSQL generated by the Method Engine was fairly easy:
And, if I check out the Question plan (bodily), I can see that the Storage Engine discovered just one current mixture of values to return our information:
Including extra substances…
Nevertheless, let’s say that the enterprise request is to research information for Product Key 364 on a each day stage. Let’s go and add dates to our report:
This was once more very quick! I’ll now test the metrics throughout the DAX Studio:
This time, the question was expanded to incorporate a Dates desk, which affected the work Storage Engine wanted to do, as as an alternative of discovering just one row, this time, the quantity is totally different:
In fact, you’ll not discover any distinction in efficiency between these two eventualities, because the distinction is only some milliseconds.
However that is just the start; we’re simply warming up our DAX engine. In each of those instances, as you may even see, we see solely “stuffed” values — that mixture of rows the place each of our necessities are happy — product secret’s 364 and solely these dates the place we had gross sales for this product — in the event you look totally within the illustration above, dates usually are not contiguous and a few are lacking, corresponding to January twelfth, January 14th to January twenty first and so forth.
It is because Method Engine was sensible sufficient to eradicate the dates the place product 364 had no gross sales utilizing the NonEmpty filter, and that’s why the variety of data is 58: we’ve got 58 distinct dates the place gross sales of product 364 weren’t clean:
Now, let’s say that enterprise customers additionally wish to see these dates in-between, the place product 364 hadn’t made any gross sales. So, the thought is to show 0$ quantity for all these dates. As already described within the earlier article, there are a number of other ways to interchange the BLANKs with zeroes, and I’ll use the COALESCE() perform:
Gross sales Amt 364 Merchandise with 0 = COALESCE([Sales Amt 364 Products],0)Principally, the COALESCE perform will test all of the arguments supplied (in my case, there is just one argument) and change the primary BLANK worth with the worth you specified. Merely stated, it should test if the worth of the Gross sales Amt 364 Merchandise is BLANK. If not, it should show the calculated worth; in any other case, it should change BLANK with 0.
Wait, what?! Why am I seeing all of the merchandise, after I filtered every part out, besides product 364? Not to mention that, my desk now took greater than 2 seconds to render! Let’s test what occurred within the background.
As a substitute of producing one single question, now we’ve got 3 of them. The primary one is precisely the identical as within the earlier case (58 rows). Nevertheless, the remaining queries goal the Product and Dates tables, pulling all of the rows from each tables (The product desk comprises 2517 rows, whereas the Dates desk has 1826). Not simply that, check out the question plan:
4.6 million data?! Why on Earth does it occur?! Let me do the mathematics for you: 2.517 * 1.826 = 4.596.042…So, right here we had a full cross-join between Product and Dates tables, forcing each single tuple (mixture of date-product) to be checked! That occurred as a result of we pressured the engine to return 0 for each single tuple that will in any other case return clean (and consequentially be excluded from scanning)!
It is a simplistic overview of what occurred:
Imagine it or not, there’s a sublime resolution to point out clean values out-of-the-box (however, not with 0 as an alternative of BLANK). You’ll be able to simply merely click on on the Date subject and select to Present gadgets with no information:
This may show the clean cells too, however with out performing a full cross-join between the Product and Dates tables:
We are able to now see all of the cells (even blanks) and this question took half the time of the earlier one! Let’s test the question plan generated by the Method Engine:
Not all eventualities are catastrophic!
Reality to be stated, we may’ve rewritten our measure to exclude some undesirable data, however it will nonetheless not be an optimum method for the engine to eradicate empty data.
Moreover, there are specific eventualities by which changing BLANKs with zero won’t trigger a big efficiency lower.
Let’s look at the next scenario: we’re displaying information in regards to the whole gross sales quantity for each single model. And I’ll add my gross sales quantity measure for product 364:
As you would possibly count on, that was fairly quick. However, what’s going to occur after I add my measure that replaces BLANKs with 0, which precipitated havoc within the earlier situation:
Hm, appears to be like like we didn’t need to pay any penalty when it comes to efficiency. Let’s test the question plan for this DAX question:
Conclusion
As Jeffrey Wang recommended, it’s best to avoid changing blanks with zeroes (or with some other specific values), as it will considerably have an effect on the question optimizer’s potential to eradicate pointless information scanning. Nevertheless, if for any purpose you want to substitute a clean with some significant worth, watch out when and find out how to do it.
As traditional, it is determined by many various facets — for columns with low cardinality, or if you’re not displaying information from a number of totally different tables (like in our instance, after we wanted to mix information from Product and Dates tables), or visible varieties that don’t must show a lot of distinct values (i.e. card visible) — you will get away with out paying the efficiency worth. However, in the event you use tables/matrices/bar charts that present plenty of distinct values, ensure that to test the metrics and question plans earlier than you deploy that report back to a manufacturing setting.
Thanks for studying!
