Understanding the Vertipaq Engine

[Datatouille]

Hello

I am doing a deep dive into the Vertipaq engine and I have a few questions regarding a concrete use case.

I have a .pbix file with two Fact Tables:

- Raw (300k rows and 17 columns) with the following structure:

Location	A	B	C	D	…	O	P
London	644 544	417 834	961 806	367 504		83 035	893 560
Vancouver	717 302	153 883	616 675	997 484		779 110	672 422
Copenhagen	643 329	17 978	215 068	935 608		270 478	827 192
London	279 920	224 229	774 350	764 692		770 781	929 427
Vancouvert	158 089	104 762	209 378	22 990		938 295	93 513
Vancouver	436 206	70 139	223 948	482 096		136 600	930 017
Copenhagen	290 725	986 045	260 513	714 680		491 096	293 309
…

- Unpiv which is the unpivoted version of the 'Raw' table above (4.8 Million rows and 3 columns). So it has the following structure:

Location	Type	Sales
London	A	644 544
London	B	417 834
London	C	961 806
London	D	367 504
London	O	83 035
London	P	893 560
Vancouver	A	717 302
Vancouver	B	153 883
Vancouver	C	616 675
…
Copenhagen	A	643 329

Obviously, I used the super Vertipaq analyzer from @marcorusso & @AlbertoFerrari to understand data compression.

Not surprinsingly, because Vertipaq uses a columnar storage, my 'Unpiv' Table is ~45% smaller than 'Raw' table:

Result from the Vertipaq analyser

We clearly see (i) the impact of the cardinality of the columns on their total size (ii) the relevance of unpivoting the table. Fair.

Here are my questions :

1. As we can see above, the size difference between 'Raw' and 'Unpiv' is mainly due to the "Column Hierarchies". What is it about ? I found no real explanation on the web or on Marco Russo's book. I read somewhere that it has to do with MDX structure...?!

2.  How is the size of Column Hierarchy computed? How can it be estimated or perhaps turned off? I understand the logic of value encoding , dictionnary encoding and RLE encoding. Is it something similar ?

3. What is the unit of the measures displayed by the Vertipaq analyser in the pivot table above ? When I save my .pbix file on my disk, its total size is: 74MB. I have no relationships between 'Raw' & 'Unpiv' and 0 graph on my report. so the data model itself is ~ 73MB (extracted by zipping the .pbix file).

--> As a result, I would expect the Total Table size from the Vertipaq analyzer of my two tables above (49 391 934) to be similar to the 73 MB of the data model. Where does this difference come from ?

Thanks in advance for your answers !

Tristan

@Datatouille

[marcorusso · Accepted solution]

1. Columns Hierarchies are internal structures storing a sorted textual version of the content of each column. It is required for MDX, and it is now used also internally for optimizing certain queries (in other words - we cannot get rid of it, even for columns we never use un MDX)
2. See above - even numbers are converted into strings
3. I don't understand the question - but VertiPaq Analyzer reports the memory consumption of the data model, whereas the size of the files could be different (serialization and additional structures).

I hope it helps!

Marco Russo - SQLBI

[Anonymous]

Dear Marco,

I hear you in this video sayng that is possible to disable Hier Size in SSAS, when not used, of course.

And now I can also see the green tickbox just on top, which makes me think that is really possible.

Could you be so kind to address me to some specific article where I can see how to do that?

Also let me thank you for all the availability, support and videos.

You are making a great job among PBI and Dax world! A Big thank you.

[marcorusso]

You should set the "Available In MDX" property to False. The internal property name is "isAvailableInMdx" and it is described here: Chris Webb's BI Blog: How The New IsAvailableInMDX Property For Analysis Services Tabular Can Reduce...

This property is available only in Tabular Editor and Visual Studio, you cannot modify it from Power BI Desktop.

Disabling it in Power BI Desktop is not supported - it could work, I never tried it. It is supported for Analysis Services (Azure and SSAS). It should work on Power BI Service as well through XMLA endpoint.

Don't disable this feature on columns that could be used as filter or exposed as visible columns to Excel. Disable only when a column is only used in an aggregation like SUM (...) and it is never visible to user.

[Anonymous]

I will definitely follow your recommendations. Thanks a lot for replying quickly.

Hopefully I will get the best from VertyPaq analyzer and improve performance of the Tabular Model which we have.

Thanks again. 

[Datatouille]

For the folks on this thread, there is a new 'Available in MDX' feature in Azure Analysis Services, SSAS 2017 (and soon Power BI Premium) which answers my questions n°1 and n°2 above: https://blog.crossjoin.co.uk/2018/07/02/isavailableinmdx-ssas-tabular/