How-to: Bridge Tables and Many to Many Relationships Demystified in OBIEE 11g

Bridge tables - entire books have been devoted to this concept, countless blogs write about it, and organizations offer entire classes dedicated to demystifying this idea. Ralph Kimball, creator of Kimball Dimensional Modeling and founder of the Kimball Group has written quite a few great articles discussing the theory of bridge tables.

Yet when researching for comprehensive guides on how to actually implement a bridge table in OBIEE 11g, the documentation available is either:

• Out of date
• Contains implementation steps for OBIEE 10g which has since been deprecated
• Does not contain adequate detail 
• e.g. missing key steps

This guide is going to outline the basic use case of a many to many relationship, how OBIEE 11g resolves this dilemma and how to successfully implement a bridge table model within the 11g platform.

First thing's first - what is a bridge table and why do we need it?

At its core, bridge table solve the many to many relationship we encounter in many datasets. Many to many relationships in itself are not "bad", but when attempting to conform a data set to a star schema - many to many relationships just do not work. Star schemas assume a one to many (1:N) cardinality from the dimension to the fact. This means "one attribute of a dimension row can be found in many rows of the fact table".

For Example:

• One job (job dimension) can be performed by many people
• You would see the same JOB_WID repeating in the fact table
• One employee (employee dimension) can have many jobs
• You would see the same EMPLOYEE_WID  repeating in the fact table
• One call at a call center(ticket dimension) can have many ticket types
• You would see the same CALL_WID repeating in the fact table
• One patient (patient dimension) can have many diagnosis
• You would see the same PATIENT_WID repeating in the fact table

This 1:N cardinality is represented in OBIEE as (using call center/employee example) :

"Cardinality of '1' applied to the dimension and cardinality of 'N' applied to the fact'

But what happens when in the above examples, the cardinality is actually N:N? 

For Example:

• Many employees can have multiple jobs and each job can be performed by multiple employees
• Many patients can have multiple diagnosis and each diagnosis can be 'assigned' to many patients
• Many calls can have multiple call ticket types and each ticket type can belong to multiple calls

This many to many relationship is initially (and incorrectly) represented in OBIEE 11g as:

'Cardinality of '1' is applied to the two dimension tables and cardinality of 'N' is applied to the fact'

Any BI Architect should recognize the above model - it's a traditional star schema! If you stop here and decided to ignore the issue with your dataset and 'hope' OBIEE aggregates the model correctly, you're about to be disappointed.

Why star schemas dont work for N:N cardinality

Consider the following scenario: You're a call center manager and you want to capture the number of calls with positive feedback per employee. You also want to capture the type of calls an employee answers in any given day.

Upon analysis of the requirements you conclude that "each call received by an employee can have many call types and each call type can be answered by multiple employees".

For example:

• I answer a take a call that is classified as a 'new call', 'urgent', and 'out of state transfer' (three different call types) - this is the "each call received by an employee can have many call types".
• A colleague also received a phone call that is classified as 'out of state transfer' - this is the 'each call type can be answered by multiple employees"

Now let's put this data in a traditional star schema fact table as modeled below:

ID	EMPLOYEE_WID	CALL_TYPE_WID	NUMBER_OF_GOOD_CALLS
1	1	1	300
2	1	2	300
3	1	3	300
4	2	2	500
5	2	3	500
6	3	1	200

You can see in the above data set that:

• EMPLOYEE 1:
• Has 3 different call types
• Has 300 positive reviews (NUMBER_OF_GOOD_CALLS) 
• This metric is at the EMPLOYEE level and not the call type level!
• EMPLOYEE 2:
• Has 2 different call types
• Has 500 positive reviews (NUMBER_OF_GOOD_CALLS)
• This metric is at the EMPLOYEE level and not the call type level
• EMPLOYEE 3:
• Has 1 different call type
• Has 200 positive reviews (NUMBER_OF_GOOD_CALLS)

Now you receive a requirement to create a KPI that displays the Number of Good Calls as a stand alone widget.

PROBLEM 1 - Aggregation :

The number of good calls you received based on the above fact table is not 2100 - it's 300 + 500 + 200 = 1000

• Employee 1 received 300 good calls
• Employee 2 received 500 good calls
• Employee 3 received 200 good calls

but due to the many to many cardinality of the data, the star schema duplicates the measures because each employee can take calls of many call types and each call type can be assigned to many employees!

PROBLEM 2 - Grand Totaling:

What if you don't care about aggregates? What if you just want a report that contains the employee, call type and a summation/grand total?

Notice how NUMBER_OF_GOOD_CALLS is repeated across multiple call types and the grand total is still incorrect. It's being duplicated due to the many to many relationship that exists between call type and employee. Furthermore, it paints an incorrect picture that NUMBER_OF_GOOD_CALLS is some how related to CALL_TYPE

How do we resolve this many to many cardinality with a bridge table?

When all is said and done, the incorrectly built star schema:

should be modified to:

Let's break this down:

The bridge table:

This the purpose of the bridge table is to resolve the many to many relationship between the call type and employee. It will contain, at a minimum, the following four columns:

1. The primary key of the table
2. The EMPLOYEE_WID
3. The CALLTYPE_WID
4. The weight factor

The weight factor is what's going to resolve the issue of double counting. 

• If an employee has 3 call types, there would be 3 rows and the weight factor of each row would be .33
• If an employee has 10 call types, there would be 10 rows and the weight factor of each row would be .1

In our bridge table data set, we're going to use the same 3 EMPLOYEE_WIDs and create the following:

ID	CALL_TYPE_WID	EMPLOYEE_WID	WEIGHT
11	1	1	0.33
12	2	1	0.33
13	3	1	0.33
23	2	2	0.5
24	3	2	0.5
31	1	3	1

You can see from this example that we've taken the N:N dataset in the fact table and moved it into this bridge.

The dimension that is joined to both the fact and bridge

This is a generic dimension that contains the unique EMPLOYEE IDs in your organization's dataset.

For example:

ID	EMPLOYEE_ID
1	1
2	2
3	3
4	4
5	5
6	6
7	7
8	8
9	9
10	10

The dimension that is joined to only the bridge table

This dimension contains all of the possible call types. Note how this table is not physically joined to the fact. This is because this specific dimension (CALL_TYPE) is what's causing the N:N cardinality

For example:

ID	DESC
1	Call Type 1
2	Call Type 2
3	Call Type 3
4	Call Type 4
5	Call Type 5
6	Call Type 6
7	Call Type 7
8	Call Type 8
9	Call Type 9
10	Call Type 10

The Fact Table

We've moved the N:N cardinality from the original fact table to the bridge table so the new fact table now contains exactly one row per employee and does not have the CALL_TYPE_WID.

ID	EMPLOYEE_WID	NUMBER_OF_GOOD_CALLS
1	1	300
2	2	500
3	3	200

How do we implement this model in OBIEE 11g?

Step 1: Import Tables into Physical Layer

This is always the first step performed when creating a model regardless of its type. In the above example i'm importing four tables:

Step 2: Create the Physical Data Model

Based on our data set above the join conditions would be implemented as follows:

• 1:N relationship from employee dimension to fact table
• 1:N relationship from employee dimension to bridge
• 1:N relationship from call type dimension to bridge

Notice how employee_demo_d is the only dimension that is joined to the fact. w_call_type_d is not joined to the fact because that is the dimension that is causing the many to many relationship issue.

Step 3:  Create the Logical Data Model
The creation of the BMM is where we deviate from our standard build steps of a traditional star schema:

1. All associated dimension tables referencing the bridge table will be stored in a single BMM table
2. The single BMM table will have two logical table source

Step 3.1 : Drag the fact table and dimension table that is connected to the fact table into the BMM. 

In our example, we are dragging w_calls_f and w_employee_demo_d into the BMM:

Step 3.2: Create a 2nd LTS in the existing dimension table

1. Right click W_EMPLOYEE_DEMO_D -> New Object -> New Logical Table Source
2. Name it 'Bridge'
3. Add W_BRIDGE_D and W_CALLTYPE_DEMO_D (the two dimensions not directly joined to the fact table) under the 'Map to these tables' section

1. Next add the remaining dimension columns from W_CALLTYPE_DEMO_D and W_BRIDGE_DEMO_D to the Dimension table in the BMM

Step 3.3: Create a level-based dimension hierarchy for the dimension BMM

1. This step should be completed whether or not the schema is a star or bridge

Step 3.4: Confirm the BMM model has a 1:N relationship from the dimension to fact

Step 3.5: Set aggregation rule of NUMBER_OF_GOOD_CALLS to sum 

All measures in the BMM must have a mathematical operation applied to the column

Step 3.5: Set the Content level of the dimension table to 'detail' in within the LTS of the fact table

Again, this is something that should always take place regardless of the type of model

Step 4: Create the Presentation Layer

This part is straight forward, just drag the folders from the BMM into the new subject area:

The moment of truth

So why did we go through this elaborate exercise again? To fix the aggregation issues we were having with NUMBER_OF_GOOD_CALLS due to the N:N cardinality of the data set. Let's create that 'standalone KPI' Number of Good Calls:

Notice how the metric correctly sums to 1000. Let's check the back end physical query to confirm:

Notice how it's hitting the fact table and not the bridge or the call type dimension. 

But what about the weight factor?

Let's go back to the scenario where we want to compare across dimensions joined via the bridge table (EMPLOYEE and CALL_TYPE):

• When creating a report that uses a measure from the fact table, a dimension value from the the employee table, and a dimension value from the table that causes the N:N cardinality - you need to use the weight factor to make sure your measure isn't getting double or triple counted:

• Notice column is using the the NUMBER_OF_GOOD_CALLS multiplied by the WEIGHT factor in column 2
• Each row in column 1 correctly represents the NUMBER_OF_GOOD_CALLS in the fact table despite having the repeated values of multiple call types
• Note the aggregation of grand total sums to 997. This is because the weight factor is rounded to the 2nd decimal for EMPLOYEE_WID = 1 (.33%)

In order for grand totaling to work correctly with bridge table measures that use weight facts you must set the aggregation rule of the column (in this case column 1) to sum within Answers:

So what did we accomplish in this guide?

• A basic understanding of many to many (N:N) cardinality
• A basic understanding of why the star schema won't work for N:N cardinality
• How to resolve the cardinality issue with a bridge table
• How to implement a bridge table in OBIEE 11g

keywords: bridge table, cardinality, many-to-many, OBIEE 11g, helper table, answers, analytics, aggregation, LTS, measures, kimball

FYI: Kimball Fact Table Primer (Star Schemas)

One fact table does not fit all, depending on your requirements you may have a need to identify all invoices transactions for a given point in time.  Maybe the end user wants to identify sales revenue for a regular, predictable date interval (such as quarter).  Also consider the order fulfillment pipeline which creates a scenario where a supply chain specialist must identify the span of time it takes a product to be processed through the pipeline.

Should 1 denormalized fact table design be used for the above scenarios? Of course not . Take, for example, Oracle Business Intelligence Applications HR Analytics 7.9.6.3 . Oracle designers created work force fact tables of various designs to capture the business process being measured: Transactional fact tables for employee salary, periodic fact tables for headcount, and accumulating fact tables for the HR recruitment process.

Below is a primer of Kimball fact tables and the characteristics of each one. Before selecting the fact table type to use for your design, always make sure you:

1) Identify the business process you are trying to measure
2) Determine the grain of the fact table
3) Identify the attributes of the fact table (dimensions)
4) Identify the measures to be captured in the fact table.

Also highly recommended: Read Kimball's Data Warehouse Tool kit. This is a must read for any OBIEE Architect. The table below is referenced from the above book.

CHARACTERISTIC	TRANSACTION GRAIN	PERIODIC SNAPSHOT GRAIN	ACCUMULATING SNAPSHOT GRAIN
Time period represented	Point in time	Regular, predictable intervals	Indeterminate time span, typical short-lived
Grain	One row per transaction event	One row per period	One row per life
Fact table loads	Insert	Insert	Insert and update
Fact row updates	Not revisited	Not revisited	Revisited whenever activity
Date dimension	Transaction date	End-of-period date	Multiple dates for standard milestones
Facts	Transaction activity	Performance for predefined time interval	Performance over finite lifetime

These three fact table variations are not totally dissimilar because they share conformed dimensions, which are the keys to building separate fact tables that can be used together with common, consistent filters and labels. While the dimensions are shared, the administration and rhythm of the three fact tables are quite different.

How to: Default Date Prompt to NULL in OBIEE Answers

Consider the following scenario:

• You have an accumulating periodic fact table that tracks an item as it moves through its distribution life cycle from warehouse to the point of sale at a store.

If applying Kimball Design Principles (which you should always follow in a denormalized dimensional model), your accumulating periodic fact table would be sparsely populated and contain NULL values for stages the item has not yet encountered. Take the fact table below as an example.

This fact table is what Kimball classifies as a 'periodic accumulating snapshot' because:

1) each record represents the life cycle of 1 item
2) each record is updated as it reaches a key date, as opposed to a new record being added
3) each record has a definite beginning and end

Assume you had a requirement to create a report that identified all of the products that arrived at the store but have yet to be sold.

To translate this technically, you're looking for Sales Dates that are NULL.

You cannot simply create a filter that says 'Sales Date <> NULL' because NULL is not a valid DATE type. NULL in DATE is represented as a '01-01-1900'. You could then be tempted to create a filter as 'Sales Date <> '01-01-1900' but that is not a valid DATE either, since by default Dates are stored as DATETIME stamps.

An easy way to exclude Dates that are stored as NULL, is to do it in answers as follows:


Step 1: Create a Dashboard Prompt for your Date that Defaults to SQL Values:

and populate it with:

SELECT

case when 1=0 then Time."Fiscal Date" else Cast(NULL as Date)  END

FROM "ENTER LOGICAL FACT TABLE HERE"

The 'Enter logical fact table here' should be replaced with the name of the logical fact table of that subject area in the BMM.

Step 2: Edit Column Formula to Cast as Date

Date dimensions default to DATETIME in the repository so you will need to cast your column as a Date.

This will generate the following result:

Note that although the default value looks 'blank', a filter of NULL is being applied in the physical query: