Data Hubs: Purpose and Peak Performance

[#1]

You may have heard about a model called a data hub, but perhaps you aren’t confident that you understand the fundamentals, primary functions, or considerations when architecting one. There are three main advantages to incorporating a data hub:

1. Single source of truth: Stores all transactional data from the source system.
2. Data validations: Ensures all data is correct and valid before the data gets to the spoke model(s).
3. Performance: It is always faster to load data from a model rather than a file.

Additionally, the administrator can ensure the correct granularity of data in the spoke model(s) when using a data hub. For example, the source system may only contain transactional data at the daily level, but the planners may need the data aggregated to the month. The data hub can summarize the data and export only the data needed.

The following information is designed to further define a data hub and support you in your journey of building your own.

Table of Contents

• Definition of the Data Hub
• Anaplan Architecture with a Data Hub
• Factors to Consider When Implementing a Data Hub
• User Stories
• Transactional
• Flat Lists
• Modules
• Loading Data vs. Using Formula’s in SYS Modules
• Exporting to Spoke Models
• Tips and Tricks
• Worked Example
• Additional Resources

Definition of the Data Hub

First, we need to define what a data hub is. This can be split into four sections:

1. Use cases: The data hub should be the first model built, whether you have a single use or multiple use cases. The data should be automatically refreshed on a schedule, whether it is nightly, weekly, monthly, etc., from the source system—often an Enterprise Data Warehouse (EDW). All modules and views that create hierarchies or lists should be stored in the data hub, which enables your models in having one version of truth, as well as reducing the duplication of data.
2. Model connectivity: Anaplan Connect, one of our 3^rd party vendors (Informatica Cloud, Dell Boomi, Mulesoft, or SnapLogic), or our REST API can be used to automate the loading of data to the data hub from the source system, as well as transferring data from the data hub to the spoke model(s). Additionally, transitional data should not be loaded directly into the spoke module, especially if there is a large volume of data.
3. Functions: Often, simple ETL (Extract, Transform, and Load) functions can be utilized within your data hub to transform the data for the spoke model(s). This is helpful when consolidating data from multiple sources where you have different “codes” and need a mapping module to ensure the correct data gets mapped correctly.
4. Team: The management of the data hub should have a designated team of experts who understand what data is stored in the data hub (to ensure duplication doesn’t happen), as well as the how and when the data gets loaded.

Anaplan Architecture with a Data Hub

There are several ways your Anaplan architecture could look, depending on the number of workspaces you currently have and the type of security your company requires. The following are illustrations of common architectures.

Master Hub Model: Across Workspaces

The most common, and recommended, architecture is when the data hub is in its own workspace. Not only does this have the advantage of not interfering with the other models, but it also adds an additional security layer, with a segregation of duties. In this view, the Anaplan Workspace Admin(s) can limit the access to the data hub workspace to only the people who require it.

Master Hub Model: Within a Workspace

The simplest depiction is where your data hub is within the same workspace as your spoke models. While this can be accomplished, it is not best practice as there is no segregation of duties and there is a possibility, upon heavy loads from the source system, of performance issues. Additionally, when adding users, the Anaplan Workspace Administrator (Admin) would need to ensure users don’t have access to the data hub, as well as any users of the data hub not having access to the spoke models

Multiple Data Hubs

Finally, the data hub doesn’t necessarily have to be the only model in the workspace. You can have additional data hubs, if needed.

Factors to Consider When Implementing a Data Hub

There are six main elements to think about when architecting a data hub:

1. User stories.
2. Source systems.
3. Lists.
4. Modules.
5. Data validation.
6. Exporting data to spoke model(s).

User Stories

One of the cornerstones of The Anaplan Way is data (process, model, and deployment being the others), which is critical to all implementations. You will need to know what data is needed for a certain use case. Consider the following, common, data questions that need to be answered in order to be successful:

• What granularity of the data is needed?
• How much history is needed? How much history do you have?
• Does the source system only have transactional data, but the use case needs the data at the month level? Can the source system do the aggregation for you?

After all data questions have been answered, shift your focus to the source system and consider the following:

• Consider the source system. Where is the data coming from? What is the source system, and is it a trusted environment? Is it Excel? Typically, you should stay away from Excel as the “source” because Excel cannot be audited.
• Define the data source owners. Who has access to this data? Who is preparing it? Are they part of the project? These are often-overlooked questions that are critical to success. Ideally, the data source owners need to be part of the project from the start to understand the file specifications and prepare the initial load of the data, as well as towards the end of the project to do a final load of the data.
• Define file specifications. How many files will be needed? Typically, you will need master data, as well as transactional data. Instead of having one file with all of this data, determine if the data can be split between different files (one for transactional, one for the unique members of the master data). It will be better for Anaplan (for performance reasons) to split these to reduce warnings during the data load process.
• Analyze the data. Understand what makes each record unique (date/period and transactional amounts should not be part of this), and make sure the data owners don’t give you everything (Select * From Employee) when you only need five columns. Remember, it is better to ask for additional columns midway through the project than getting all columns in the beginning and only using a select amount.
• Consider custom codes in the source system. Find more on this in the transactional lists section. This is a great trick for transactional data. After you have analyzed the data to understand what makes each record/row unique, concatenate the “codes” of the metadata into one transactional code, but remember, you will need to be under the 60-character threshold.
• Define the schedule. When is the data available? Is the data on a certain schedule? What is the schedule required with this use case?
• Determine the ETL medium to be used. Will Anaplan Connect be sufficient, will one of our 3^rd parties be used, or will a more custom application be needed, such as REST API? Does your company already have this experience inhouse, or will training be required? These will need to be factored into all data stories.

Transactional

Usually, the largest lists are those containing transactional data. There can be millions of transactional ID’s with several list properties defined. First, properties should not be defined on a transactional list (or any list, except for Display Name, as they do get accounted for in the workspace memory). Secondly, instead of loading metadata to list properties (Cost Center and Account as properties), try to figure out a way to incorporate them into the code. If the transactional data is defining a transactional amount at the intersection of Cost Center and Account for a particular month, attempt to use the code of the Cost Center and the code of the Account concatenated together (0100_57000). Not only will this decrease your list size, but it will also create a healthier model.

In the below example, the model builder did not create a custom code, but rather used a combination of properties to make the record unique, which included the date/period, as well as the transactional amount. Notice the original number of records vs. the number of records after a custom code was created.

By incorporating the date/time period, as well as the transactional amount, it inflated the list size exponentially based on the number of years that were loaded. Doing this not only caused the model to be bigger, but also caused poor model opening performance.  See the Appendix for a simple worked example to explain further.

Learn more about sparsity in the two-part series The Truth about Sparsity: Part 1 and The Truth About Sparsity: Part 2.

Flat Lists

Similar to transactional lists, flat lists are not part of a hierarchy and are a series of records grouped in a list, like Products, Companies, Cost Centers, or Employees. These are your “legends” or “anchor” for all metadata about this unique record. Again, the only property that should be defined is a Display Name, if needed. It is best practice, from a model builders’ perspective, to suffix the name with “Flat” or “- Flat”. This helps identify whether the list is part of a hierarchy or flat list (Employee – Flat, Cost Center – Flat, Product – Flat). These lists can be used for data validation, which will be described later in this article.

Modules

Ideally, you should have three types of modules in the data hub:

1. Transactional: A Transactional module will store the transactional data by the time series, whether that be by day, week, month, quarter, or year. The only data, or line items, should be transactional data. No other line items should be defined. Additionally, to keep the size down, make sure the summaries on the line items are turned off, or None, as there is no reason to sum the data within the module.
2. System: System (SYS) modules, or the “S” in DISCO, do not have time associated with them and should only be dimensionalized by the same list (Employee Flat, Cost Center Flat, Product Flat). These modules store the metadata or attributes about the list item that doesn’t change over time, for example the employee’s start date. Another example of a SYS module would be any kind of mapping that is required, whether it be SYS Time Filter module or a mapping from one source system to another.
3. Export modules: If the data from the source system is being loaded at a lower granularity than needed in the spoke model(s), export modules can aggregate the data to the specified need (month, quarter, or year level), which will lead to more efficient data load performance to the spoke model(s). Additionally, it is better to only load the granularity of data needed instead of loading all data to the spoke model, but only using a portion of it.

Loading Data vs. Using Formula’s in SYS Modules

If you can devise a custom code where all of the attributes of the data are accounted for, you can greatly increase the performance of your data load, especially on very large data volumes. It is actually faster to use formulas to derive the data from the custom code than it is to load the data. Why? A couple of reasons. First, when data is loaded, the load is triggering the change log, and every change is being recorded in the model history. Second, loading data to another module is an additional action. If you didn’t need that action, you would save processing time.

In the example below, the exact same data was loaded four different ways:

• Import Properties to a List: A list was created with all attributes, including the transactional data, and was loaded to list properties (not best practice and against DISCO).
• Import to List, Attribute, and Trans: A list was created, the transactional data was loaded to a transactional module, and all of the attributes were loaded to a SYS Attribute module.
• Import to List, Trans, Calculate Attribute: A list was created, the transactional data was loaded to a transactional module, but the SYS Attribute model was calculated using two different methods:
  • One Line Item: Using FINDITEM() with several functions parsing out the information from within the FINDITEM(). For example, FINDITEM(Cost Center, RIGHT(LEFT(Trans Details.Code, '2nd Group’), 3)).
  • Multiple Line Items: Parsing of the member spread across multiple line items and using FINDITEM() with only the list and code as the parameter. First, you do the parsing to get the correct piece of the code (one line item), and then the FINDITEM() of that code (2^nd line item).

Load Performance

Notice, the best performing data load was the last one, Import to List, Trans, Calculate Attribute (multiple line items), where the parsing out of the data was spread over multiple line items. This is due to the fact that the data load was able to take advantage of Anaplan’s multithreading capabilities. The worst performing data load occurred when data was loaded to the Attribute module because, due to the sheer size of the data, a save had to be performed.

Exporting to Spoke Models

One of the most important concepts to remember when exporting data is to use a view from a module. Lists should not be exported because you lose control over what you export. It is either all or nothing. By using views, you can employ a filter (should always be a Boolean) to render exactly which data needs to be exported. If you need more than one filter, combine both into one line item and use that line as the filter. You will have much better performance if you are only using one Boolean line item as a filter vs. having multiple filters defined.

Another important concept to remember is to only export detailed information, as there is no reason to export parent information (quarter, year, etc.). Not only will you get warnings when exporting parent information, but the performance of the export will decrease because the system will have to create a debug log. The goal is to make sure a debug log is not created, all green checks, so if there ever is an issue, you will know it truly is an issue that needs attention.

Line items in the data hub formatted as text should not be exported as text, but actually as list formatted line items in the spoke model (text->list formatted line item). The goal is to reduce the number of text formatted line items in the spoke model.

Some say they need to do validation in the spoke model, therefore they need to import the data as text. Actually, this is false, because the validation should have already been done in the data hub, so there should be no need to do the validation again.

Lastly, you should think about what really needs to be exported. Do you really need to export historical data that hasn’t been changed? Instead, just export the newly loaded data, or delta data. This can be accomplished by using one of two methods:

1. From the source system, request IT to only send the updated information, not the full load every time. Additionally, request IT to create a column in the source file with a hardcoded value of “TRUE.” This will tell Anaplan which row is new or has been updated and can be used as a filter for an export. Just know, before the import of the source data gets loaded, make sure the first action within the process clears out the previous true records (set this up via a view using a filter where the view only shows members with a value of true).
2. Utilize the current period function to only export the current period data. In the SYS Time Filter module, create a line item named Current Period with the formula CURRENTPERIODSTART(). In the export views, filter the data on this line item.

Tips and Tricks

A few of tips and tricks to be aware of include the following:

• Hierarchies should not be in the data hub.
• Analytical modules should not be in the data hub.
• Do not delete and reload lists.
• Data Validations Model

Why should hierarchies not be in the data hub? To answer that question, you need to understand why hierarchies are used in the first place. Essentially, hierarchies are only needed to aggregate data for analytical purposes, and since users will not normally login to the data hub, the lists essentially take up space. With that said, it is perfectly okay to create the hierarchies for testing purposes to ensure your actions from the meta modules are building the hierarchies correctly, but as soon as the actions are working correctly and have been verified, you can remove the list structures from the data hub. A case can be made that certain implementations may need the hierarchies created in the data hub for validation purposes of several sources. If this is the case in your implementation, just be sure to only use the hierarchies for validation purposes.

In addition to the above, there are two more reasons to not have hierarchies built in the data hub—cluttered data, and spoke models that pull data from the lists.  

Data hubs need to be clean and clutter free to ensure optimal performance, which also makes it easier for the administrators to understand exactly what data is stored in the data hub. Additionally, when you have lists—especially hierarchical lists—spoke model builders will sometimes build their lists from the lists within the data hub instead of from a view. It is best practice to always build lists from views from within a module so the action can benefit from filters (there are no filters when importing from lists).

Analytical modules should not be in the data hub since end users don’t normally access the data hub. There really isn’t a reason to have products by versions by time in the data hub, that belongs to the spoke model. Remember, the data hub should only be used to store data from the source system(s).

Within your nightly data load process, do not delete and reload data, including the list structures. If you have a proper code, you shouldn’t need to do this. Additionally, not only does this impact the overall performance of the process (adding an additional action to delete the list, which then deletes all data associated with that list), but the process is essentially filling up the change log with the exact same data that it had before the delete. When a certain threshold is surpassed, the model will require a save, thus taking up even more time. Ultimately, you are forcing the model to re-aggregate all of the data, instead of just the new data.

Lastly, if you know you will have to do a lot of transformations on your data (consolidating multiple source systems or your data is not clean), think about creating a Data Validations model.  This model’s sole purpose would be to clean the data and then feed the data to the data hub, thus keeping the transformations to a minimum in the data hub as well as keeping the data hub clean.

Worked Example

Use Case: Transaction Data is by Store and SKU and Month

Bad Way

• The code for the Transaction list is a three-part code Store_SKU_Month
• Attributes for Store, SKU and Month are imported as Text and matched against the Store list, SKU list and Time period respectively
• An additional line item is needed for the Store and SKU code (for export).

This is the screenshot of the bad way:

Notice the repetition of the attributes. STR07 and SKU031 are repeated each month.

Good Way

• Two data files
  • Unique combinations of Store and SKU (two-part code)
  • Store SKU code by month for the quantity.
• The transaction details are stored in a module dimensioned by Transactions
• The Store and SKU attributes are calculated using the “_” delimiter
• The quantity is stored in a module dimensioned Transactions and by month
• The additional line item is needed for the Store and SKU code (for export). This is a subsidiary view in the module as it is not dimensionalized by Time.

These are the screenshots of the good way:

Below lists out the breakdown of the model in terms of List size, Line items and the associated member usage of the various structures. The main reasons for the improvement are because lists themselves account for approximately 500b for each member and also there is repetition of the attributes per “month” in the transaction data (as mentioned above).

---

Hopefully, this article has shed some light on data hubs, how they should be used, and what you can do to ensure they perform at their peak level. Remember, analyze the data to understand what makes the row unique and use that as the code. Every list should have a code—every list!

Additional Resources

[andrewtye]

Hi @rob_marshall  - question on list hierarchies. For a store network have the Store, Region, Division. So would have three lists and then a module dimensioned on Store with Region / Division as line items that would be updated via import from source and then this would flow into the spoke models as a hierarchied list starting at the top and going down?

[rob_marshall]

Hey @andrewtye,

Yes, ideally you would like for Store, Region, and Division to be flat lists in the data hub and the transactional data coming into a module with the "transactional" key/list to be a concatenated with the members to make the row unique. 

Then, you would have a "properties" module defining all master data for that unique row.

Lastly, you would have views defined on the properties module to create the hierarchy in the spoke model.

Hope this helps,

Rob

[andrewtye]

Thanks @rob_marshall !

[CallumW]

Hi @rob_marshall

This is a fantastic article, thanks for sharing.

When you did your analysis regarding "Loading data vs using Formulas", was file type accounted for?

I know zipped files are loaded into Anaplan much more quickly than normal csv/txt if they are large and just wondered if this was included in processing time.

Also, it would be good to have some clarity over "upload" time and "processing" time, I'm not sure if there is any distinction between them in an upload process but I would assume so, and I presume this article focuses on the processing time?

Best,

Callum

[rob_marshall]

@CallumW ,

Thanks for reading it, I am glad you liked it.  Regarding the file types, this was not part of the article nor the actual loading of a file into Anaplan, more of loading the data to the line items when the file had already been uploaded vs. using a formula to determine the data (from the code).  Does that help and clarify?

Thanks,

Rob

[CallumW]

Appreciate the quick reply @rob_marshall .

I guess my question would then be whether there is any processing benefit in using different file types when loading into line items, or is there a process that converts to "Anaplan language" upon upload and therefore before loading into line items, thus no benefit?

i.e. the only benefit of loading different file types would be seen on the upload stage before Anaplan loads into line items?

Apologies for the confusion!

[rob_marshall]

@CallumW,

In asking around (Ben Speight), there is no difference in uploading a csv file vs. a text file and there is no difference in the reading of the data from the file and importing them to line items.  Now, if you are using Anaplan Connect, I believe the data gets zipped on the way into Anaplan but if you are using a browser, the data does not get compressed.

Hope this helps,

Rob

[CallumW]

Appreciate the follow-up, @rob_marshall !

[JaredDolich]

@rob_marshall I love this post.

I just finished loaded a years worth of transactions and tested both ways, the bad way and the good way.

The good way was 82% smaller and 90%+ faster. AMAZING!

Hope you include this best practice in the level 3 cert training.

[rob_marshall]

@JaredDolich  - That is fantastic.  It is amazing what the differences are and how easy it is to get them.

[nicole.johnson]

This was such a helpful article. Thank you!

[nicole.johnson]

Any opinions on best practices when the source system is an Anaplan spoke, rather than an external system? For example, the annual budget is developed in Anaplan using a Budget spoke. While there are transaction modules in the Budget spoke for user data entry, there are multiple calculations performed on that data to arrive at the final budget (ex: headcount based expenses). The end result of the Budget spoke is a dimensionalized module, not a transactional module. There are other reporting spokes that need summarized, dimensionalized budget data (ex: for variance analysis).

Since the data in the Budget spoke is already dimensionalized, it doesn't make sense to go to great lengths to flatten it to get it into the HUB, only to need it dimensionalized in other reporting spokes. Is this an example in which it makes sense to have a module with more than just the time dimension in the Hub? Or maybe I'm missing something big here. I'm not looking for a perfect answer, just more of a general guidance around when a spoke serves as a data source. Perhaps it doesn't flow through the hub at all?

[rob_marshall]

@nicole.johnson ,

Great question...There is no need to flatten in the spoke (Budget), just bring in the data that is needed (aggregated at the correct level) to the target spoke model.  Also, I would not advise taking the data back to the HUB from the Budget model to the spoke model (Budget -> Hub -> Spoke), just get the latest and greatest data directly from the Budget model.  If you do go to the HUB first, there is a chance the spoke misses an important update in the Budget model that hasn't made it to the Hub yet.

Hope this helps,

Rob

[piotr.weremczuk]

@rob_marshall , @nicole.johnson 

I do agree that there is no need to flatten the data. Less data you move around is better for performance.

On the point about pushing the data through the HUB I just want to bring one more thing - it's always good to have single point of thruth for you data. Possibly you will be using the data in the future to feed other model (that is possibly not yet there in the environment) and then it will be easier to use HUB simply rather than redesign data flow, I guess. Moreover if you schedule your imports between models carefully you won't end up with an issue of missing updates.

Piotr

[rob_marshall]

@piotr.weremczuk ,

If there is a pre-defined schedule, you are correct.  With that said, often times folks need to pull the data on an ad hoc basis using the same actions/process for a quick update.  If you are having to waiting on the data going to the Hub, it may be missed or not there yet.

Rob 

[obriegr]

@piotr.weremczuk @rob_marshall 

Wouldn't it be a good idea to allow both? setting up an action that pulls meta data from both the HUB (automated processes, and Single Source of Truth (SSoT)) and what is currently live/most up to date according to the related Spoke model?

Eg:

If you want to make it so the user can't edit, you just set up a hidden value (they can edit/update) and point the end point and you are good to go.

Greg

[piotr.weremczuk]

@obriegr 

It is an idea but from the other hand this way you generate more traffic between models what impacts perfromance. I wouldn't go that way unless users really pushing to. Having the processes scheduled during night prevents performance downgrades since noone is usually using the models that time.

Also if you have the setup you've proposed you should also take into account that it may cause chaos if they will try to reconcilie data between models but someone did not run the process yet. From supporting perspective it may be quite messy.

[obriegr]

@piotr.weremczuk 

I was referring to Meta data only (top level summaries, or limited lists, eg. instead of the million+ item list, the next level up or 1 cell) and giving the end user the ability to run the light process.

Alternatively, the user may not be able to update from the Hub (as it is meta data that is pulled automatically in lock step with the rest of the model), and only pull the meta data from the Spoke model so they know how many changes are coming. This should be significantly lighter load as it is as small as 1 line item with 1 cell or at least under 100 cells total.

[Asslam]

@rob_marshall , Great article with good examples. Have you come across a scenario, where as per your example loading the details in DAT01 Transnational Details and then load data in DAT02 Transnational Data. However, source system drops the one of the unique combination all together and respective data is removed from the database (hard delete). Basically, the data is create on day one of the month and day two the record is removed. Therefore, now the data hub and spoke model has the value for the unique combination and the values does not gets updated from source as the unique combination does not exist anymore.  

[rob_marshall]

@Asslam ,

If I am understanding you correctly, I have not come across that scenario where a piece of the data (unique row) is gone/dropped from one day to the next.  Just to be clear, in the above, DAT01 and DAT02 are using the same list, but the modules are dimensionalized differently as DAT02 uses Time and DAT01 does not.

Are you saying in your example, SKU or Product is no longer needed from one day to the next?

Rob

[obriegr]

@Asslam 

Are you using ALM? If you push an update and don't have production flagged correctly that could be the cause.

[iamwilliam_nz]

@rob_marshall , what a great article! We've already tested a few of your recommendations and are already seeing significant performance gains - so thank you!  

Currently, we have about 15 flat transactional modules in our Data Hub which we load monthly however they don't all share the same dimensions.  So, in order to implement your approach above we will need to create 15 unique lists (one for each load) and 15 Properties modules? I was wondering if there is a more efficient way to go about this?

Thanks

[rob_marshall]

@iamwilliam_nz 

Actually, having 15 different unique lists is not a bad thing as long as you set then up correctly as well as the corresponding "Properties" modules.  By doing this, this can and will help with potential validation issues as now you can figure out which source needs fixing.

If you would like to get on a call, DM me and we can talk further about this.

Hope this helps,

Rob

[Asslam]

@rob_marshall , thanks for your message,  Yes, SKU or Product not needed from day 1 to next , might be needed for future load but not for a particular month. We had come across a scenario where the record was cleared out on day 2. Basically, the whole record is dropped from the source application. Which means on Day 1 we have a value for Jan for combination of  SKU or Product  and on Day 2 the value is not zeroed but rather no records comes through for the  SKU or Product combination. Therefore, now the Jan values does not gets updated to zero. We figured out a solution using a data variation list ( two list items current load / previous load) in the data module to load two version of the data and clearing out all item in the previous load (second load) to manage this scenario, Still following the good way with couple of extra load actions.

Thanks again this article really helped. 

[rob_marshall]

@Asslam ,

So, again, if I am understanding you correctly, this is one of the main tenants of the article - load the data into a transcational module (dimensionalized by Time) instead of deleting the list and reloading it and adding Time as part of the code/key.  In your case, if you have data for Jan 1, but not Jan 2, then nothing should be loaded into Jan 2 and thus it would be zero.  Please correct me if I am wrong, but I think the issue above is about the deletion and/or clearing of a module.  If so, you can change the clearing of data within the action definition.

Please let me know if this makes sense,

Rob

[jake.fletcher]

@rob_marshall - This article is very insightful. Thank you for putting this out here on the community. I'm curious if you're doing any exception reporting and data validation in your models using the Worked Method above? I'm assuming that any of the data exception reporting would be occurring in DAT01 Transactional Details?

Thanks,
Jake

[rob_marshall]

@jake.fletcher 

Absolutely, and that is a great point.  In an ideal world, you should be doing all validations in the Data Hub and within those modules that actually warrant it, so definitely in DATA01.  In theory, you only want validated data to be imported into your models, so doing all of the data validation in the data hub is the correct method.  If the data has issues or needs a further review, then that data should not be imported to your downstream model until it has been verified.

Hope this helps,

Rob