watching Jeffrey Wang as a stay stream visitor with Reid Havens, and one of many dozen fantastic issues that Jeffrey shared with the viewers was the listing of optimizations that the DAX engine performs when creating an optimum question plan for our measures.
And, the one which caught my consideration was concerning the so-called “Sparse measures”:
To make it easy, when you outline the measure, Method 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 information for the mixture of your dimension attributes actually exist. For people coming from the MDX world, the NonEmpty operate might look acquainted, however let’s see the way it works in DAX.
The factor that the majority resonated with me was when Jeffrey suggested towards changing BLANKs with zeroes (or no matter express values) in Energy BI calculations. I’ve already written how one can deal with BLANKs and change them with zeroes, however on this article, I need to concentrate on the attainable efficiency implications of this choice.
Setting the stage
Earlier than we begin, one essential disclaimer: the advice to not change BLANK with 0 is simply that — a advice. If the enterprise request is to show 0 as a substitute of BLANK, it doesn’t essentially imply that you need to refuse to do it. In most eventualities, you’ll most likely not even discover a efficiency lower, however it’ll rely on 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 need to calculate the overall 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 occasions 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 present mixture of values to return our information:
Including extra substances…
Nevertheless, let’s say that the enterprise request is to investigate information for Product Key 364 on a every day stage. Let’s go and add dates to our report:
This was once more very quick! I’ll now test the metrics inside 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 a substitute of discovering only one row, this time, the quantity is totally different:
After all, you’ll not discover any distinction in efficiency between these two eventualities, because the distinction is only some milliseconds.
However that is only the start; we’re simply warming up our DAX engine. In each of those circumstances, as you might even see, we see solely “crammed” values — that mixture of rows the place each of our necessities are glad — product key’s 364 and solely these dates the place we had gross sales for this product — in case you look totally within the illustration above, dates should not contiguous and a few are lacking, akin to January twelfth, January 14th to January twenty first and so forth.
It is because Method Engine was sensible sufficient to remove the dates the place product 364 had no gross sales utilizing the NonEmpty filter, and that’s why the variety of information is 58: now we have 58 distinct dates the place gross sales of product 364 weren’t clean:
Now, let’s say that enterprise customers additionally need to see these dates in-between, the place product 364 hadn’t made any gross sales. So, the concept is to show 0$ quantity for all these dates. As already described within the earlier article, there are a number of alternative ways to switch the BLANKs with zeroes, and I’ll use the COALESCE() operate:
Gross sales Amt 364 Merchandise with 0 = COALESCE([Sales Amt 364 Products],0)Mainly, the COALESCE operate will test all of the arguments offered (in my case, there is just one argument) and change the primary BLANK worth with the worth you specified. Merely stated, it’ll test if the worth of the Gross sales Amt 364 Merchandise is BLANK. If not, it’ll show the calculated worth; in any other case, it’ll change BLANK with 0.
Wait, what?! Why am I seeing all of the merchandise, after I filtered every thing 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 an alternative of producing one single question, now now we have 3 of them. The primary one is strictly 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 accommodates 2517 rows, whereas the Dates desk has 1826). Not simply that, check out the question plan:
4.6 million information?! 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 compelled the engine to return 0 for each single tuple that might in any other case return clean (and consequentially be excluded from scanning)!
This can be a simplistic overview of what occurred:
Consider it or not, there’s a chic answer to point out clean values out-of-the-box (however, not with 0 as a substitute of BLANK). You may simply merely click on on the Date discipline and select to Present objects with no information:
This can 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!
Fact to be stated, we might’ve rewritten our measure to exclude some undesirable information, however it could nonetheless not be an optimum approach for the engine to remove empty information.
Moreover, there are particular eventualities wherein changing BLANKs with zero is not going to trigger a major efficiency lower.
Let’s study the next state of affairs: we’re displaying information concerning the whole gross sales quantity for each single model. And I’ll add my gross sales quantity measure for product 364:
As you may count on, that was fairly quick. However, what’s going to occur after I add my measure that replaces BLANKs with 0, which triggered havoc within the earlier situation:
Hm, appears like we didn’t must pay any penalty by way of efficiency. Let’s test the question plan for this DAX question:
Conclusion
As Jeffrey Wang instructed, you need to keep away from changing blanks with zeroes (or with another express values), as this may considerably have an effect on the question optimizer’s potential to remove pointless information scanning. Nevertheless, if for any cause it’s worthwhile to substitute a clean with some significant worth, watch out when and learn how to do it.
As typical, it is dependent upon many various elements — for columns with low cardinality, or once 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 may get away with out paying the efficiency value. Then again, in case you use tables/matrices/bar charts that present loads of distinct values, make sure that to test the metrics and question plans earlier than you deploy that report back to a manufacturing setting.
Thanks for studying!
