Search This Blog

Thursday, December 1, 2016

Tableau - Implementation Challenges and Best Practices

Hi All,

I thought of sharing my leanings and experiences with Tableau so far.

This post will describe some of the challenges you could face while implementing Tableau BI in your enterprise from architectural standpoint.

If you are evaluating BI tools or planning to start implementation, you will definitely benefit from this post. I would be highlighting some of the best practices that you can include in your list.

Tableau is flexible when it comes to playing with and analyzing your data. It gives you complete freedom to choose and connect to your data source, and quickly start building those nice Viz (reports or charts or dashboards).
You can do pretty much everything to join the data sources in a SQL, put filters to restrict your data. If you are a data analyst, you can build some really compelling data visualizations or charts in a very short span of time.
Now you show those nice visualizations with your team or department and they too get very exited.

Till here it was all cool stuff. The challenges starts from here.

1. Do I don't need a Datawarehouse star schemas.?
Datawarehouse star schema contains Fact and dimensions that gives you enormous benefits in simplifying your implementation. You won't believe how it can benefit in terms of performance, scalability and maintenance.
Some may argue that Tableau doesn't need any kind of warehouse or these fact and dimensions star schemas.
Well, if you are really a very small enterprise then you may not need it but otherwise if you have good amount of data and have various source systems and applications, then do not build your BI without a datawarehouse. Or sometimes, your organisation has a warehouse but as a data analyst you may be tempted to NOT use it.
Since Tableau does not have any centralized metadata layer, users are free to create their SQL the way they want. This freedom proves costlier in long term strategy.
Developers build their SQL's on top of OLTP or normalized data structures and the result is you have highly complex SQL's with large numbers of joins giving you poor performance.
Very soon you will have hundreds of those complex SQL's with lots of duplicate data/information where one SQL may differ from another SQL slightly. It's not so easy to debug those complex SQL's to make any additions or alterations. Now you understand how difficult it would be to maintain those SQL's.
Star schema reduces those joins and makes your SQL very simple, and of course the performance is way better.
Tableau can extract the data in extract mode and improves the performance to some extent but do not just ignore the other benefits.For some reason in future if you need to make your application in Live mode then you may need to completely redesign it. Such reasons could be more frequent data refresh or implementing row level security for which you need to have Live connection for your Tableau application.

2. Temptation to put ALMOST everything in one Tableau Workbook:
When you start creating an application, you start with small dataset providing answers to very limited or few business questions. This is what tableau is built for.

Slowly when more and more people starts looking at it, they start asking for more and more information. This is when we start adding new data sets, joins, transformations and conditions. And our application starts growing from all angles.
It becomes more complex, performance goes down and it becomes difficult to  scale.
If we take a break here and plan things, we can do it in much better way.
Once we realize that our application is growing, think of going to point no 1 above of creating/extending the dimensional model.
You need recreate your application using a dimensional model. If you think about this early, you will reduce the amount of rework you would have to do.
The ideal design would be to do all the data analysis/discovery using your source systems structures (assuming you do not have a warehouse or the required information is not present in a warehouse at all).
Utilize all the freedom Tableau provides here. But once you start thinking of making it available for mass consumption by enterprise users, design the required subject areas (Facts and dimensions) or extend the existing ones.
Build your application now using these subjects areas. Your application would be simple, fast, scalable and easy to maintain. Since the new SQL would be using less joins, fewer calculations and aggregations, it would be fast and easier to read.
You can now imagine the benefits. If you need more data elements or metrics, simply keep adding them to your subject areas.
This will enable you to extend or scale your application to a greater extent BUT this does not mean you can still put almost everything in one workbook.
Definitely there is some more work here but I am sure you would appreciate the benefits it would bring in the long run.

3. I Still want to put almost everything in one Workbook:
You may be wondering if I am against that. Well I am not.
There are many instances where we need to have information to be displayed on our dashboards side by side that may be coming from different subject areas or Stars but there are certain things we need to consider and remember.
Since Tableau does not have a Semantic layer (aka Common Enterprise Reporting Layer), we need to have all the tables added to that one workbook as Data sources.
Here the grain of the data plays an important role. If the grain of the data is same then all can fit in one data source/SQL.
But if the grain of the two data sources are different and there is a need to have an interaction between these data sources then the real trouble starts.
When I say interaction between these two data sources, I mean to say that we need to pass common filters between them or need to show the data coming from these two data sources into one Viz/report.
When we need to have an interaction, we need to have a join between these two data sources. Tableau allows joins across data sources or perform blending but it may prove to be very costly in terms of performance and stability.
You would be surprised that even if individual queries have sub second response time, after applying the join the response time may be in minutes.
If your individual queries have limited or small data, it may work for you in some cases.
Better always test it out. Even Tableau experts suggest to avoid using the blending.

4. OK. what is the Solution then:
I know its frustrating when we talk about limitations only. Here it is also important to understand why such limitations when Tableau is such a nice tool?.
Well, Tableau is a tool for data discovery. Quickly go grab your data and starts visualizing it. Maps are inbuilt and required no configuration like in many other tools. But once we have built those nice dashboards we need to make it available for the enterprise users. Tableau can do certain things here but its not made for that. Now you are trying to make it do something that some enterprise BI reporting tools such as Oracle OBIEE or Business Objects or Cognos are just made for that. These tools can do some data discovery but not the way Tableau does, similarly Tableau can do some dashboarding but not the way they do it.
Here I am not comparing Tableau with them since they are not comparable and have totally different use case and technologies.

5. What else can I do to?
All right. Here is the solution.
We need to design our Tableau workbooks and dashboards intelligently keeping in mind the limitations.
Think of having a common landing page workbook with hyperlinks to all the other applications. Think of having some very common filters on your landing page. So your first workbook have just dimension data for those filters.
Now you can also think of making one or more of these filters mandatory meaning users need to have a filter value selected in order to go to a specific workbook/dashboard.
This would help in cases when your workbooks/dashboards have tons of data and you want to avoid just showing all of that data and slow down your application.
Now, you can build your simplified workbooks based on individual common subject areas and link them to your landing page.
Since Tableau allows to pass the filters between workbooks, you can pass the common filters from one workbook to another.
There may be certain cases when we want to have a dashboard/report having data from 2 different data sources and in those cases you can consider blending. I know I said Tableau experts suggest to avoid it.
See if blending works fine for you else think of creating a physical table in database combining the two sets of data having different grains.
This table will have data at both the grains and some indicator column will tell the row has data for which grain. you will find any example on the web for such cases since this issue is not specific to Tableau but common to data warehouse.

6. THAT'S IT?
Well I guess So until something comes to my mind. Please post your comments and questions, and share your thoughts and experiences.

Thanks for reading.
Manohar Rana