In Part 4 of this series on data blending, we reviewed Tableau’s Data Blending Principles. We also reviewed an example of data blending in Jonathan Drummey’s Patient Falls Dashboard. 
Today, in Part 5 of this series, we will peel the onion a bit more and look at Tableau’s Data Blending Architecture.
Again, much of Parts 1 – 5 are based on a research paper written by Kristi Morton from The University of Washington (and others) .
You can learn more about Ms. Morton’s research as well as other resources used to create this blog post by referring to the References at the end of the blog post.
Integrating Data in Tableau
In Part 5, we discuss in greater detail how data blending works. Then we discuss how a user builds visualizations using data blending using several large datasets involving airline statistics.
Data Blending Architecture
The data blending system, shown in Figure 1 above, takes as input the VizQL query workload generated by the user’s GUI actions and data source schemas, and automatically infers how to query the data sources remotely and combine their results on-the-fly. The system features a two-tier mediator-based architecture in which the VizQL query workload is analyzed and partitioned at runtime based on the corresponding data source fields being used. The primary mediator initiates this process by removing the visual encodings from the VizQL query workload to yield an abstract query. The abstract query is partitioned for further processing by the primary mediator and one or more secondary mediators. The primary mediator creates the mediated schema for the given query workload. It then federates the abstract queries to the primary data source as well as the secondary mediators and their respective data sources. The wrappers compile the abstract queries into concrete SQL or MDX queries and instantiate the semantic mappings between the data sources and the mediated schema for each query. The primary mediator joins all the result sets returned from all data sources to produce the mediated result set used by the rendering system. 
A visualization is organized by its discrete fields into pages, partitions, colors, etc., and like a GROUP BY clause in SQL, these grouping fields comprise the primary key of the visualization. In a blended visualization, the grouping fields from the primary data source become the primary key of the mediated schema. In Figure 2 above, these are shown as the dark-green fields in the primary data source, and the light green fields represent the aggregated data. Each secondary data source must contain at least one field that matches a visualization grouping field in order to blend into the mediated schema. The matching fields in a secondary data source comprise its join key, and fields appear in the GROUP BY clause issued by the secondary mediator wrappers. The aggregated data from the secondary data source, shown in light-purple, is then left-joined along its join key into the mediated result set. Morton (et al) refer to this left-join of aggregated result sets as a post-aggregate join. 
Primary Key Cardinality
many mapping between the domain values of the primary key and those of the secondary join key, because the secondary join key is a subset of the primary key and contains only unique values in the aggregated secondary result set. Morton (et al) find that this approach is the most natural for augmenting a visualization with secondary data sources of uncertain value or quality, which is a common scenario for Tableau users.
Data blending supports many-to-one relationships between the primary and each secondary. This can occur when the secondary data source contains coarser-grained data than the mediated result set, as discussed in Part 3 of this series.
Since the join key in a secondary result set may match a subset of the blended result set primary key, portions of the secondary result set may be duplicated across repeated values in the mediated result set. This does not pose risk of double-counting measure values, becaused all aggregation is performed prior to the join. When a blended visualization uses multiple secondary data sources, each secondary join key may match any subset of the primary key. The primary mediator handles duplicating each secondary result set as needed to join with the mediated result set.
Finally, a secondary dimension which is not part of the join key (and thus not a grouping field in the secondary query) can still be used in the visualization. If it is functionally dependent on the join key, a secondary dimension can be used without affecting the result set cardinality. Tableau references this kind of non-grouping dimension using both MIN and MAX aggregations in the query issued to the secondary data source, which allows Tableau to determine if the dimension is functionally dependent on the join key. For each row in the secondary result set, if the two aggregated values are the same then the value is used as-is, reflecting the functional dependence on the grouping fields. If the aggregated values differ, Tableau represents the value using a special form of NULL called ManyValues. This is represented in the visualization as a ‘*’, but retains the behavior of NULL
when used in calculated fields or other computations. The visual feedback allows a user to distinguish this lack of data from the NULLs which occur due to missing or mismatched data.
Inferring Join Keys
Tableau uses very simple rules for automatically detecting candidate join keys:
- The secondary data source field name must match a field with the same name in the primary data source.
- The data types must match
- If they are date/time fields, they must represent the same granularity date bin in the date/time hierarchy, e.g. both are MONTH. A user can intervene to force a match either by providing field captions to rename fields within the Tableau data model, or by explicitly defining a link between fields using a simple user interface.
Another Simple Blending Example
A Tableau data blending scenario is shown in Figure 3 above, which includes multiple views that were composed in minutes by uniquely mashing up four different airline datasets, the largest of which include a 324 million row ticket pricing database and a 140 million row on-time performance database. A user starts by dragging fields from any dataset on to a blank visual canvas, iteratively building a VizQL statement which ultimately produces a visualization. In this example, the user first drags the VizQL fields, YEAR(Flight Date) and AVG(Airfare), from the pricing dataset onto the visual canvas.
Data blending occurs when the user adds fields from a separate dataset to an existing VizQL statement in order to augment their analysis. Tableau assigns the existing dataset to the primary mediator and uses secondary mediators to manage each subsequent dataset added to the VizQL. The mediated schema has a primary key composed of the grouping VizQL fields from the primary dataset (e.g. YEAR(Flight Date)); the remaining fields in the mediated schema are the aggregated VizQL fields from the primary dataset along with the VizQL fields from each secondary dataset.
Continuing our example, the user wishes to drag AVG(Total Cost per Gallon) from the fuel cost dataset to the visualization. The schema matching algorithm examines
the secondary dataset for one or more fields whose name exactly matches a field in the primary key of the mediated schema. While the proposed matches are often sufficient and acceptable, the user can specify an override. Since the fuel cost dataset has a field named Date, the user provides a caption of Flight Date to resolve the schema discrepancy. At this point the mediated schema is created and the VizQL workload is then federated to the wrappers for each dataset. Each wrapper compiles VizQL to SQL or MDX for the given workload, executes the query, and maps the result set into the intermediate form expected by the primary mediator.
The mapping is performed dynamically, since both the VizQL and the data model evolve during a user’s iterative analytical workflow. Finally, the primary mediator
performs a left-join of each secondary result set along the primary key of the mediated schema. In this example, the mediated result set is rendered to produce the visualization shown in Figure 3(a).
Evolved Blending Example
Figure 3(b) above shows further evolution of the analysis of airline datasets, and demonstrates several key points of data blending. First, the user adds a unique ID field named unique carrier from the primary dataset to the VizQL to visualize results for each airline ID over time. The mediated schema adapts by adding this field to its primary key, and the secondary mediator automatically queries the fuel cost dataset at this finer granularity since it too has a field named uniquecarrier. Next, the user decorates the visualization with descriptive airline names for each airline ID by dragging a field named Carrier Name from a lookup table.
This dataset is at a coarser granularity than the existing mediated schema, since it does not represent changes to the carrier name over time. Morton’s (et al) system automatically handles this challenge by allowing the left-join to use a subset of the mediated result set primary key, and replicating the carrier name across the mediated result set. Figure 4 below demonstrates this effect using a tabular view of a portion of the mediated result set, along with portions of the primary and secondary result sets.
The figure also demonstrates how the left-join preserves data for years which have no fuel cost records. Last, the user adds average airline delays from a 140 million row dataset which matches on Flight Date and uniquecarrier. This is a fast operation, since the wrapper performs mapping operations on the relatively small, aggregated result set produced by the remote database. Note that none of these additional analytical tasks required the user to intervene in data integration tasks, allowing their focus to remain on finding insight in the data.
Tableau provides several options for filtering data. Data may be filtered based on aggregate conditions, such as excluding including airlines having a low total count of flights. A user can filter aggregate data from the primary and secondary data sources in this fashion, which results in rows being removed from the mediated result set. In contrast, row level filters are only allowed for the primary data source. To improve performance of queries sent to the secondary data sources, Tableau will filter the join keys to exclude values which are not present in the domain of the primary data source result set, since these values would be discarded by the left-join.
Data Cleaning Capabilities
As mentioned in the Inferring Join Keys section above, Tableau supports user intervention in resolving field names when schema matching fails. And once the schemas match and data is blended, the visualization can help provide feedback regarding the validity of the underlying data values and domains. If there are any data inconsistencies, users can provide aliases for a field’s data values which will override the original values in any query results involving that field. The primary mediator performs a left-join using the aliases of the data values, allowing users to blend data despite discrepancies from data entry errors and spelling variations. Tableau provides a simple user interface for editing field aliases. Calculated fields are another aspect of Tableau’s data model which support data cleaning. Calculated fields support arbitrary transformations of original data values into new data values, such as trimming whitespace from a string or constructing a date from an epoch-based integer timestamp.
As with database fields, calculated fields can be used as primary keys or join keys.
Finally, Tableau allows users to organize a field’s related data values into groups. These ad-hoc groups can be used for entity resolution, such as binding multiple variations of business names to a canonical form. Ad-hoc groups also allow constructing coarser-grained structures, such as grouping states into regions. Data blending supports joins between two ad-hoc groups, as well as joins between an ad-hoc group and a string field.
Next: Data Blending Using MicroStrategy
 Kristi Morton, Ross Bunker, Jock Mackinlay, Robert Morton, and Chris Stolte, Dynamic Workload Driven Data Integration in Tableau, University of Washington and Tableau Software, Seattle, Washington, March 2012, http://homes.cs.washington.edu/~kmorton/modi221-mortonA.pdf.
 Hans Rosling, Wealth & Health of Nations, Gapminder.org, http://www.gapminder.org/world/.
 Jonathan Drummey, Tableau Data Blending, Sparse Data, Multiple Levels of Granularity, and Improvements in Version 8, Drawing with Numbers, March 11, 2013, http://drawingwithnumbers.artisart.org/tableau-data-blending-sparse-data-multiple-levels-of-granularity-and-improvements-in-version-8/.