Dimensional modeling – architecture and terminology
by Joakim Dalby (danish Dimensionel modellering, Datamodellering, Datavarehus)
Information has become a major asset for any organization. The book Das Kapital by Karl Marx could today get the title Die Information. A data warehouse is the way to provide information to the business users and add value to the business and reuse data in a new way to increase the revenue. A data warehouse is a separated system so a user query and analysis will not slow down and not reduce the workload on the operational systems.
Gartner formulates that the platform for data and analytics is the heart of the digital business. Here you will find data management and the company's governance for data and analysis. Data is the raw material in our digital age. Data is collected, cleaned, controlled, optimized, shared and used, and our ability to work flexibly and intelligently at this »assembly line« is growing fast. Data management is crucial to tomorrow's business success. They must shape the technology and the data resource both short-term and long-term for the sake of both the existing business and the future's business. Now, the actual data platform is, with good reason, a focus area for the companies. In a data-centric world, tools, processes and competencies to handle data are infinitely crucial. All of today's greatest technological trends ranging from Internet of Things IoT, robotics, artificial intelligence and virtual reality are ultimately dependent on data managers, architects and company data strategies.
»That's all data is. A gift from yesterday that you receive today to make tomorrow better.« Jon Acuff: Finish: Give Yourself the Gift of Done, page 126.
A data warehouse tells you what has happened - not what is going to happen or why it has happened. There are many gaps that need to be filled in by a different kind of insight than the one that can be delivered in a data flow task from source systems to an enterprise data warehouse, but let's start with what is a data warehouse.
Data warehouse is
· Subject-oriented because data from the business is merged into areas of subjects and not organized around the functional applications of the business.
· Integrated because multiple operational source systems is using common and consistent business names, definitions, formats and units.
· Time-variant because historical data is kept as current data at any given time.
· Non-volatile because data do not change or historical data will never be altered.
· Unite, interconnect, compile, conform and consolidate data into common format.
· Enrichment and processing of data for providing new knowledge.
· Collection of data in support of management's decision making process and analyzing the business with business performance measurement qualifiers and KPI Key Performance Indicator used as a measure of how a company is doing.
· Utilizing dimensional modeling, end users and analysts can easily understand and navigate the data structure and fully exploit the data for self-service BI.
· Goal to support slice and dice of data for business analysis and insight.
· Data-driven to become enjoyable and pleasurable to work with for the users.
· Compliance requirement for provide proof that the reported numbers are accurate and complete by having an audit trail for reconciliation check etc.
· Driven by business requirements specification and user story from stakeholders and users. A data warehouse must have focus on the needs of the business.
Data from multiple operational source legacy systems are stored in tables in a relational database called a data warehouse. Data will be merged and integrated inside a data warehouse to get a consistent format and use same naming. Uniformity means that within the data warehouse the common data elements of multiple source legacy systems are referred to in a consistent manner. Conforming means that within the data warehouse creating a conform column with a single and consistently labeled name that has the same common meaning with identical contents all over the data warehouse e.g. same spelling and unit of measurement. For example, product data can be found in multiple source legacy systems with different names of columns, different spelling of product names and different segmentations of products, therefore a data warehouse will unite, interconnect, compile, conform and consolidate the data into one product table and one product name column to make it easy for the business users to make a list of all products. A conformed dimension supports integration of multiple source legacy systems into a data warehouse.
It is important to design a data warehouse to support reporting and data analyses by a single and common data model that is easy to navigate in and a business user does not think of data across of source legacy systems. Data integration becomes seamlessly. Many operational source legacy systems update its data to reflect the most current and recent state while a data warehouse also maintain history. For example, an ambassador who has lived in Washington, London and New Delhi for the last 25 years and always bought her perfume from a Lancôme store in Paris with a IT system that only stores current shipping address, all the sales over the last 25 years will be placed in India, while a data warehouse will remember all cities and give a correct picture of the sales shipped around the world.
Data mart is a specific subject area that contains the data and information that is relevant for a business user. A data mart has one purpose to customize and summarized data for tailored support of a specific analytical requirement from a business unit. It utilizes a common enterprise view of strategic data and provides business units more flexibility, control and responsibility. For example, Sales mart, Customer mart, CRM mart, Churn prediction mart, Market mart, Production mart, Inventory mart, Shipment mart, HR mart, Tax mart, Fraud detection mart etc. A data warehouse can have multiple data marts bound together with conformed and shared dimensions. A conformed dimension is shared by multiple facts that has conformed measures with same calculation methods and common unit of measure e.g. sales revenue and support revenue in two facts are both pre-VAT (value added tax) in dollar and adding them will result in a total revenue. A data mart contains the data to calculate Key Performance Indicators (KPI) e.g. MRR Monthly Recurring Revenue.
Dimensional modeling is a technique approach that seeks to present data in a standard, intuitive framework of dimension (descriptive context and hierarchy) and fact (measurement) and focuses on ease of end user accessibility and provides a high level of performance access to the data warehouse. Ralph Kimball recommends in the design of a data warehouse to decide what business process(es) to model by combining an understanding of the business requirements with an understanding of the available data. Kimball page 37: »The BI team needs to understand the needs of the business as well as the realities of the underlying source data.« A process-centric fact table supports integration of multiple operational source legacy systems via a set of conformed dimensions, and the entire data warehouse is build process-by-process.
Business Intelligence (BI) system provides the information that management needs to make good business decisions.
Data warehousing is made up of three components: the data model, the architecture, and the methodology. It is a method to help database designer to build a comprehensive and reliable data warehouse system, e.g. based on Dimensional modeling principles. A OLAP (OnLine Analytical Processing) cube can be at the top of the Dimensional modeling to present data in tools like Excel, Power BI, QlikView, Tableau or Targit.
BI is going from data and information to knowledge and wisdom to the users.
Previously called Decision Support System. BI could also stand for Business Insight.
In the article I will focus on the basic concepts, terminology and architecture in Dimensional modeling. My homepage has other articles of how to design and implement a data warehouse system with levels of data in the ETL process (Extracting, Transformation, Loading). I see Dimensional modeling as a subset of the Entity Relationship (ER) data modeling design method of a relational database system OLTP (OnLine Transaction Processing). Peter Chen, the father of ER modeling said in his seminal paper in 1976: »The entity-relationship model adopts the more natural view that the real world consists of entities and relationships. It incorporates some of the important semantic information about the real world.« Where a ER data model have little or no redundant data, dimensional models typically have a large amount of redundant data. Where a ER data model have to assemble data from numerous tables to produce anything of great use, dimensional models store the bulk of their data in the single fact table or a small number of them and divide data into multiple databases called data mart. Read more.
Back in the 1970's Bill Inmon began to define the term data warehouse as a one part of the overall business intelligence system. An enterprise has one data warehouse with entities in a relational model as a centralized repository, and multiple data marts which source their information from the data warehouse. In the data warehouse, data is stored in 3rd normal form. Data warehouse is at the center of the Corporate Information Factory (CIF), which provides a logical framework for delivering business intelligence. A data mart is a simple form of a data warehouse that is focused on a single subject. Denormalization and redundancy is the norm for data modeling techniques in a data mart. Ralph Kimball: »A data warehouse is a union of all its data marts.« Kimball’s data warehousing architecture is also known as Enterprise Data Warehouse Bus Architecture matrix (BUS matrix) as a collection of conformed dimensions that has the same meaning to every fact (e.g.).
A table contains columns also known as attributes or fields. Important tip by naming columns, do not use possessive nouns such as Recipient's birth date, better use CamelCase like RecipientBirthDate in a column name, and don’t use values in column name like EmployeeDay/NightCode better with EmployeeShiftTypeCode. With conforming or conformance we avoid synonym columns with same content but different names, and we avoid homonyms columns with same name but different content, e.g. Type is a general term better give a column a name as ProductType, CustomerType, TransactionType etc. Names of tables and columns must be business-friendly and must make sense to the business users.
A column contains an atomic value like a First name or a composite value like a Full name that can be further divided into three columns for Firstname, Middlename and Surname. A null represents a void value in a column that is missing at the present time or is not defined for the row. A column has a data type e.g. integer, string or text, date, amount, boolean or bit for true (1) and false (0). A column can contain a null value or a column have a constraint of enforcing it to always contain a value.
A derived column represents a value that is derivable from the value of a related column or set of columns and not necessarily in the same table. For example, the age of a person is derivable from the date of birth column and date of today, or a calculation of an amount or a glue column to bind rows together for fast search.
A single-valued column holds a single value for an entity occurrence.
A multi-valued column holds multiple values, MovieCategory »Children, Comedy«.
A primary key is a unique identifier for all rows in a table so each row contains unique values and can’t contains null value (is not null, not allow null, non null).
A composite primary key is composed of multiple columns, also called compound.
An entity is a set of objects with the same properties and is implemented as a table with rows where a primary key is unique and for use in identification of rows.
A relationship is a meaningful association among entities between primary key and foreign key in a one-to-one, one-to-many or many-to-many relationship to make sure of the consistency in database. When a database is consistent we can navigate with a join among tables and relationships become navigation paths. Example of a many-to-many relationship is between Employee and Course, that one employee can take many courses, and that one course can be taken by many employees.
SQL stands for Structured Query Language e.g. Select From Join Where Union.
A view is a stored sql query criteria statement used to fetch data from a database and the data set has columns as a table and is called a result set or a recordset. Also called a virtual table representing the result of a database query.
A materialized view is a database object that contains the results of a sql query e.g. data is being persisted into a table, a file or placed in a server memory cache. Materialization is for performance reasons as a form of optimization of the ETL process and for fetching data to business users e.g. five tables that are joined and aggregations are precomputed. For a client-server database application, the result can be placed in a local copy on user computer PC.
Data is a plural of datum e.g. 5.5 or 38 and after labelling them shoe size in US and in EU they become information. When I know many of my friends' shoe sizes together with other information I have knowledge that I can use for better birthday gifts or sell to a marketing campaign.
Conforming - what to call it?
Humans all over the world is using many terms for the same thing, for example a transport system: Subway in New York City, Tube or Underground in London, Metro in Paris or Copenhagen, U-Bahn in Berlin, Vancouver and Bangkok has Skytrain as an elevated train system above the streets. It can be a little tuff job to find and determine a good and common conformed column name and conforming the data.
When a filipino wants to buy toothpaste in a store, he will ask for a colgate. A sign on the road says xerox instead of photocopying. In Denmark I will ask for a kleenex instead of a facial tissue. Some names of brand becoming words in our language.
If Shakespeare had been a data modeler, I wonder if he had made the question:
»A transaction has relationships or a transaction is a relationship?«
· If a transaction can be referenced and has a TransactionID as primary key and it is a thing (its own entity) then it has relationships.
· If a transaction only contains foreign keys and with no need for a unique identifier and nothing else will reference it, then it is a relationship,
e.g. many-to-many data-bearing relationship.
In Peter Chen's Entity Relationship Diagram ERD there are many kind of transactions e.g. an order, an invoice or deposit money into a bank account that is a many-to-many data-bearing relationship in the conceptual level design with the caption like "deposit transaction". Moving on to a physical level design a many-to-many relationship becomes a table in the database.
A counterpart to an invoice is called a credit memo, credit memorandum or a credit nota to reduce the amount of the invoice for several reasons, for example: the item is returned, the billed unit quantity is larger than the one delivered, the invoiced unit price is higher than agreed or there are errors in the delivery that justify price reductions. If a transaction in a source legacy system is changed, corrected or cancelled then a data warehouse can calculate a counterpart transaction with a negative amount. The calculation of the final amount will be a summary of transactions whereby the final total amount will be less than the original amount.
1.1. Data layers in a dimensional modeling architecture
A description of the content of data layers of the throughput from source legacy systems to business users PC screen for improved analytic decision making. I am using the term area to point out that an area may contain several databases to split data from multiple source legacy systems to make the entire solution scalable and can be distributed over multiple servers. Each area is labelled with data warehouse acronyms as an abbreviation formed from the initial letters.
Source legacy system area – SLS
A company or an organization is using several operational systems to handle daily data of many different kinds. The systems is often based on a database with primary keys and foreign keys, and there is three types of candidate keys to identity an entity object:
· Natural key exists in the real world e.g. a fingerprint, a hotel room number, a medical diagnosis, a phone number, a currency code and other kind of codes that a user understand e.g. JobCode MNG is manager, PRG is programmer and CLN is cleaner. A website address/url. An address with country + state/providence/ region/district + postal code + street name + house number + floor number (stairwell, door number, PO box). A general ledger account number. Country name. Names of my friends combined with their birthday. The key value is not immutable but is meaningful for a human being. Back in 1990s we could remember many phone numbers, but the cell mobile phone contact list took over.
· Business key exists in a system and is build as a artificial number e.g. ISBN International Standard Book Number, Social Security Number (SSN) a civil registration number like each danish person by birth is given a CPR number for Central Person Register (PersonCivilRegistrationIdentifier) but it is GDPR sensitive, two-letter ISO country code, a vehicle identification number (VIN), a vehicle registration number plate, Customer number, Purchase order number, Invoice number with Product numbers, Insurance policy number, Flight number, Login name etc. Some of the numbers do exists in a natural way in the real world because they are printed on papers and things. The key value is not immutable but is meaningful for a human being that a business user prefer to use to identify a thing and as a search lookup value giving in a phone call to a company, a bank, a hospital or the government.
· Surrogate key exists in a system is a artificial auto-generated unique sequence number or identity column (id, uid) that is used as an unchangeable column instead of the natural/business key and a composite primary key. It is represented as an integer, a guid or a hashbyte of natural/business key. The key value is immutable and is meaningless for a human being and will normally not be exposed to outside users, but is useful for join operations inside the system.
A natural key or a business key is what the business uses to uniquely describe data. If a natural key or a business key is immutable and unchangeable in the source legacy system then call it a durable natural key or a durable business key.
A table called Salestype has a business key as an one-letter code and has a natural key as a textual value, e.g. R = Retail sales, D = Direct sales, O = Online sales.
A table called Status has a column StatusId that is a surrogate key auto-generated unique sequence number, a column StatusCode that is a durable business key and a column StatusName that is a durable natural key as a textual value.
A table called Customer has a column CustomerId that is a surrogate key auto-generated unique sequence number, a column CustomerNumber that is a business key and a column CustomerName that is almost a natural key. It is common in a customer system that the same person exists several times because contains in duplicate rows with same name and address and with different CustomerNumber, therefore I will not call it a durable business key.
From a data warehouse point of view, a natural key is also a business key, and a surrogate key is sometimes called a business key e.g. a table called LeadState has a surrogate key Id and a column Text with values as »New, Contacted, Qualified, Lost«, then a data warehouse lets Id become the business key because a text value can be changed over time, like »Lost« is changed to »Not interested«. A beloved child has many names.
Data profiling of a source legacy system is very important and the choice of column to become a data warehouse business key requires a good data analysis and criteria for the choice.
Be aware of a value of a natural key or a business key can be changed in a system because the original value was entered incorrectly and need to be corrected. A composite natural key in a Member table with Name+Address will be updated when a person change place of residence. I saw in an airport, that the flight number was selected as the business key for the data warehouse and used in the ETL process, but no one had taken into account, that when a plane was delayed, the letter D was added at the end of the flight number, therefore the data warehouse thought it was a new flight.
A surrogate key as an integer or guid won't ever need to be changed, but if a customer is deleted by a mistake and the customer later will be inserted again with the same CustomerNumber as business key, then it will get a new sequence number in CustomerId. When a company has several operational systems handling customer data, each system will use its own surrogate key for CustomerId, therefore business key CustomerNumber is the glue between the systems.
It is important for a data warehouse to receive the value of the business key (and/or the natural key) from a source legacy system, and in case of a value has been changed, that data warehouse also receive the before-value. A database has many tables for example, a Customer table has columns of CustomerId as surrogate key, CustomerNumber as business key and a CustomerName and an Invoice table has columns of InvoiceId, InvoiceNumber, InvoiceDate and CustomerId as foreign key, the invoice data must be extended with the business key CustomerNumber either by source legacy system or inside a data warehouse to know the customer of an invoice. I prefer that a data warehouse also store the value of surrogate key to be helpful when a value of a business key is changed and to obtain traceability. Merging and integrating data can be hard to set up in a ETL process and to test afterwards.
Input data area - IDA
From source legacy systems to Input data area for storage of raw data. Also called Raw data, Source, Legacy, Capture Operational Data layer, Data Acquisition or Extract. Empty this area in the beginning of the ETL process. Data types will be adjusted to fit with the receiving database system. Do reconciling between data warehouse and source legacy systems with reconciliation of row count and sum of values and mark as reconciled and do auditing report. (to reconcile in danish at afstemme, stemmer overens). Maybe a receipt system where IDA tells the source legacy systems that data has been received. Data might be wrong in input data area or in archive and later source legacy system will resend a new revision of data, therefore important to create output views that filter wrong data away and making correct data as an extraction to the next layer of the data warehouse. Each table has a column IdaInsertTime so all rows in all tables will use the same capture time as a received date that can be used later for keep history. IdaInsertTime represents a snapshot of the total amount of data as it look likes in the source legacy systems. IdaInsertTime can be set in a System database when the fetching is starting and ETL jobs/packages/transformations/sql will reuse the same time for all data rows. Sometimes a source legacy system has a Changed column with data/time of first inserted or data/time of latest updated data in the row/record that can be used for later history.
Exchanging data from source legacy system to IDA can be in different formats e.g. JSON data-interchange format where I prefer storing data for IDA in a relational database with tables with rows, columns and data types, I must parse a json format.
Kimball recommends that source legacy systems express data at the lowest detail possible for maximum flexibility and extensibility to provide the data warehouse for simplicity and accessibility and it will be the data warehouse to make summary data, not the operational source data system. IDA is a mirror of the source.
Archive area - ARA
From Input data area to Archive area. Also called Operational Data Store (ODS), Persistent Staging Area (PSA), Persistence Layer, Data Repository or History. Never empty this area because it is archiving of time variant source data and it will retain historical value changes in the source legacy system. Simple data adjustment can be done to gain same date and amount format and same data representation of a social security number etc. but it will be in new columns so the original values are unchanged. Each table has columns e.g. IdaInsertTime, ArcInsertTime (datetime2(7) default sysdatetime(), not unique per row), ArcTs (timestamp, rowversion, unique per row), ArcRecordId (a sequence number per row per table for uniqueness) to fetch delta data out of the archive that have not yet been loaded, and maybe an ArcGlobalId (unique sequence number per row across all tables). Each table can also have a source legacy system indicator. Reasons for having an archive can come from issues of data quality and data governance and external regulations e.g. Sarbanes-Oxley, Basel I, Basel II, HIPAA. Archive is a versioning of the source. Read more about how to load data from an archive in chapter 8.
Archive area can also be called ADA for Archive data area or Active data archive not to be confused with ADA for Advanced data analysis or Automated data analysis. Microsoft using the term Landing Zone as the place source data is landing and is never deleted.
Data staging area - DSA
From Archive area to Data staging area by extraction to prepare data cleansing, data cleaning and data scrubbing (e.g. trim string, unicode string, max length of string, null gets a value as 0 or empty string '', replace illegal characters, replace value, verification of data type, data syntax, date format (e.g. mm/dd/yy convert to UTC yyyy-mm-dd), correcting misspelling, fix impossible values, punctuation and spelling differences), to achieve common format, notation, representation, validate data, correct data and data deduplication by selecting one of the duplicate rows and keep the others in a variety table. The data cleansing and data integration process with multiple source legacy systems is to make data cleaned and conformed, e.g. a gender code from one source as »Man, Women«, another as »M, F« will be conformed to »Male, Female« through a mapping of source data. Data enrichment according to business rules, identify dimension data and fact data with derived values e.g. total cost and revenue, elapsed time and overdue time. This is the transformation of data. Empty this area in the beginning of the ETL process because data is only transient in this layer. When an archive is stored on an another server it is common to load data from the archive into a data staging area e.g. data to a dimension of type 2 or 7 and from the stage table do the merge or insert/update to the dimension table in the data mart area. I like to design a stage table to suit the destination or target table structure in a data mart rather than the source e.g. translate names of columns and data types, merge or divide columns like a Name to become two columns of Firstname and Surname. I like to use computed columns in a stage table for calculation, string manipulation and hashbyte value for a comparison column to compare data with a dimension table in the data mart area, but a hashed column is not good for a index because new values never comes in the end so it takes extra performance to update the index with hashbyte values. DSA database will contain different auxiliary tables with names as: stg, stage, tmp, temp, temporary and with multiple stage tables to the target they can have an additional number e.g. Stg_Customer_Order_1, Stg_Customer_Order_2, and I have seen that the last table before the target table is using 0, Stg_Customer_Order_0.
DSA will perform a data quality and filter wrong data and invalid data into a quality assurance and quality control (QA/QC) database. DSA continues reconciling data from the archive for referential integrity i.e. foreign key value exists as primary key value, relationship cardinality rules for dependencies as mandatory, optional or contingent to make sure there is a foreign key or not, it can be null or it needs a value. Many other value checks and summation and so on for a true validation of data. When all tables in this area is loaded with correct data, the DSA is successfully completed with ValidFrom and ValidTo based on the date of today, or the source legacy system new date or changed date or the IdaInsertTime that represent a batch of insertion and mark all data to that batch time no matter which source legacy system data is coming from. All ValidFrom will be using the same batch time which is useful for join between tables and between source legacy systems to fetch the right data shapshot at a particular time back in history.
Furthermore is derived columns and calculations ready for loading to the next layer of the data warehouse. In case of an error in the ETL process or in important data there will be raised a halt condition to stop the ETL process. Other times data will pass by with a flag in an audit dimension. Kimball has called it data wrangling to lasso the data and get it under control (data munging). DSA can also be called Work area or just Staging or STG. (In danish this layer is called data forberedelsesområde eller data rangeringsområde til data vask, data udsøgning, data berigelse, data behandling, data beregning, data sammenlægning, sammenstilling, samkøring, data er blevet vasket og strøget.) Kimball calls this layer ETL-system or the first system that is off limits to all final data warehouse clients and he use an analogous: »The staging area is exactly like the kitchen in a restaurant. The kitchen is a busy, even dangerous, place filled with sharp knives and hot liquids. The cooks are busy, focused on the task of preparing the food. It just isn’t appropriate to allow diners into a professional kitchen or allow the cooks to be distracted with the very separate issues of the fine dining experience.«
Kimball's ETL is using DSA as a temporary space to create the data warehouse as a collection of data marts. Inmon's ETL is using a relational 3NF ODS Operational Data Store with validate referential integrity for operational reporting and as a source of data for the Enterprise Data Warehouse (EDW) which feed the Corporate Information Factory.
Data mart area - DMA
From Data staging area to Data mart area with dimensional modeling, conformed and shared dimension tables, star schema around each fact table, assigned surrogate key (artificial key, identity column, a unique sequence number) for each dimension and use it in fact as foreign key. When multiple fact tables share dimensions it is called constallation schema or multi-star schema. Keep historical data in dimension and fact with timeline columns ValidFrom and ValidTo columns. Make it easy for users to search for current data through a view and to give a historical date to a table-valued function to fetch dimension data and fact data at any point of time. A data mart can have multiple fact tables with different granularities. One or multiple fact tables can create an extra derived fact table with special calculations and search filter criteria to enrich data to match the business requirements specification. Make a conform name of a column with a conform data type, for example when you merge customer addresses from multiple source legacy systems make a string extra long to fit all address characters and avoid a value to be mandatory or set a default as the empty string to ensure a robust ETL process. Never empty this area and always backup before the ETL process. Data mart is a front-room for publishing the organization’s data assets to effectively supports improved business decision making. A data mart contains one data area for one purpose and is subject-oriented, therefore a data warehouse will consist many data marts and they will have some common dimensions. To avoid to maintain the same dimension in multiple data marts may it be considered to have a dimension data mart that share its tables through views inside the other data marts. (In danish this layer is called tilrettelagt lag eller behandlet lag.) Kimball calls this layer dimensional presentation area and he is using the term first-level for a data mart contains data from only one source legacy system and second-level or consolidated for a data mart with multiple sources to cross business processes.
Presentation interface area - PIA
From Data mart area to Presentation interface area through data access tools like Reporting Services (SSRS) and Microsoft Access or data visualization tools like Tableau, QlikView, Qlik Sense and Power BI can import data from a dimensional schema and handle in-memory calculated KPI as quick measure using filter. A OLAP cube or a Tabular model is loading data from a data mart and the processing is doing calculation of KPI and display a pivot in Excel, Targit or Power BI. A fixed report with criteria parameter do search, select and calculate data from a data mart based on an ad hoc query of a user or the report is send out as pdf file og excel file to business users every day, week, month and so on. Data mining and machine learning will also use data from a data mart. Percentage (fraction) and ratio based on a dimension slice will be calculated »on the fly« in a BI tool when a fact table contains a numerator and a denominator, e.g. in DAX a query language for tabular model in Power BI and best to use a star schema.
Average price = DIVIDE(Sum(UnitPrice x Quantity), Sum(Quantity), 0)
Active customers =
VAR measurevalue =
FILTER ('Fact_Sales';'Fact_Sales'[EndDate] = BLANK() &&
'Fact_Sales'[Amount] > 0))
RETURN IF (ISBLANK(measurevalue); 0; measurevalue)
Kimball calls this layer Business Intelligence Application.
A date range lookup in a Kimball type 2 dimension when fact only has a business key (_bkey):
CampaignName = CALCULATE(VALUES('Dim_Campaign'[CampaignName]),FILTER(
'Dim_Campaign', 'Fact_Product'[Campaign_bkey]='Campaign'[Campaign_bkey] &&
Supporting databases (not part of data layer and is introduced by Joakim Dalby)
When custom data in the Usage database is going to be changed it can be done in a UAT user acceptance testing environment. After the ETL process has been executed we can test the data warehouse. Later we can take a backup of the Usage database and restore it in the production environment.
A data integration mapping example from the Usage database where an entity sales device has two kinds of codes (here named _bkey for business key) from two operational systems. I have invented an immutable surrogate business key _sbkey (also called a retained key or a persistent durable supernatural key in Kimball page 101), that is a simple letter code to be used in a dimension in a data mart for easy search criteria for a user query and a drop-down box. It is the _sbkey that stands for the integration to a conformed dimension. Let's have an example.
Table SalesDevice_Map with a surrogate business key _sbkey as the primary key merging all devices from retailer shops and from online users plus a good name description and a sort order for a SalesDevice dimension and a report, all entered into a Usage database by a data steward:
Table SalesDevice_Map_Retailer with business key _bkey from source system mapped to the relevant surrogate business key _sbkey and timeline columns ValidFrom and ValidTo:
Table SalesDevice_Map_Online with business key from source system mapped to the relevant surrogate business key. We see that sales devices PC and Mac has different business key values in this source system compared to the above mapping table, and over time they have changed the business keys in the source system from a letter code to a number value. Business key 10 represent two different sales devices first a Tablet computer and later a Personal computer.
For example, business key value 1 has two mappings, one for a retailer as Electronic cash register and another for an online user as iPad. The ETL process will for transaction sales data do a lookup in the relevant map-table to fetch the _sbkey value and another lookup in the SalesDevice dimension to fetch the key value to be saved within the fact row, more about this later in the article.
Twelve horoscope star signs as a dimension which values will never change. Name of horoscope is a natural key, abbreviation is a business key and I am using the planet/house cusp number as surrogate key. Element becomes a grouping of the twelve signs and the first level of the hierarchy Element→Horoscope. (icons)
1.2. Data capture or data ingestion
Data capture (in danish modtagelse, hjemtagelse) from source legacy systems is based on two data flow directions:
Granularity of data capture integration strategy is to consider when the amount of data is huge in the source legacy systems. A data warehouse prefer to receive at the lowest granularity level of detail in case of specific analysis usage or data validation, but sometimes it is necessary to aggregate and summarize source data to a higher granularity like per day and per customer segment for transaction data or transactional data from an OLTP system e.g. orders, invoices, billings, payments, site hits. Non-transactional data e.g. Customer, Location, Contact, Supplier, Part, Product can be stored in a Master Data Management (MDM) database as master tables that is shared with multiple operational applications in the organization or in the company. MDM is an extension of the Usage supporting database.
Data capture from source legacy systems is based on multiple data delivery forms:
Data capture from source legacy systems is based on two data delivery methods:
It is important to do a log of number of rows/records from the source legacy systems to the Input data area for later reconciliation check count auditing (in danish kontroloptælling, afstemning) in case a source legacy system does not deliver the expected number of rows and to monitor over time the number of rows to see if it increases as expected and be used in a graph over the amount of data in the data warehouse. Other data values e.g. amount, quantity and volume can also be logged and monitored. I recommend that source legacy system tells the number of rows per table that will be compared to the saved number of rows in Input data area in an audit trail. It can be implemented as a view in the source legacy system like this:
CREATE VIEW BI_Audit_NumberOfRows AS
SELECT 'Northwind' AS Sourcename, 'Customers' AS Tablename, COUNT(*) AS NumberOfRows
SELECT 'Northwind' AS Sourcename, 'Orders' AS Tablename, COUNT(*) AS NumberOfRows
SELECT 'Northwind' AS Sourcename, 'Order Details' AS Tablename, COUNT(*) AS NumberOfRows
FROM dbo.[Order Details]
The data warehouse can do its own delta data detection between Input data area and Archive area to identify new data, changed data and deleted data to maintain historical data with ValidFrom and ValidTo datetimestamps is handled by following actions:
Sometimes the source legacy system provides a flag information like N for new data, C for changed data and D for deleted data.
I seen pair as: New – Changed, Created – Modified or Insert – Update.
Sometimes the source legacy system update a data row multiple times but only the most recent version goes to the data warehouse. If source legacy system insert and update a row before a new load to data warehouse, only the updated version goes to data warehouse. If source legacy system insert a new row and delete the same row before a new load to data warehouse, the data will never goes to data warehouse. It is seldome that a source legacy system has a log and is logging all changes with a revision nummer or a ValidFrom and ValidTo datetimestamps, therefore a data warehouse does not contains 100% full historical updates of data in an organization. Sometimes a source legacy system has for each (or selected) tables in the database an extra historical log table (shadow table) that contains all inserted, updated and deleted data rows from the original table together with an Action column (values for Insert, Update, Delete) and an EntryDate datetimestamp column that mark when the action was occurred and a SequenceId that is used for delta data detection to pull data from Source log to Input data area. An example of a source historical log table with rows every time a customer change:
Implemented as after-insert/update/delete triggers in source legacy system.
For example, Jane Macquarie was sales manager for the eastern region until December 31, 2018, and then took responsibility for the western region from January 1, 2019. How is that registered in the source legacy system? I have seen systems without a date to register the shift, because an operator will update the responsibility at beginning of January, which means the data warehouse can let Jane get a ValidFrom date as 2019-01-05 for western region but it is a wrong picture and sales fact data for few days in 2019 will belong to eastern region. Please, always ask a source legacy system for multiple date columns and use them for ValidFrom and ValidTo to get a real picture of the business in archive area and data mart area.
Data latency describes how quickly source data has to be ready in the data warehouse for the business users to do their reporting. Deadline is normally in the morning based on a periodic batch of source data from yesterday. For a real-time load to a data warehouse with constantly new presentation in a dashboard, the ETL process must be streaming oriented where source data continuously flows into the data warehouse do to transformation and make conforming data.
1.3. Enterprise Data Warehouse modeling architecture
A data warehouse can be modeled in different ways and I will briefly present a few of them. In a data warehouse solution the data from the source legacy systems can go into a database called Enterprise Data Warehouse (EDW) that is a common data area before the Data mart area (DMA). Notice, to modelling EDW on the business, not on the source legacy systems.
1.3.1. Bill Inmon modeling - Corporate Information Factory model
When a EDW is modeled after Bill Inmon it is based on Peter Chen’s Entity Relationship data modeling with super-sub entity, associative entity and with 80% normalized data. The EDW offers integrated, granular, historical and stable data that has not yet been modified for a concrete usage and can therefore be seen as neutral. It can keep historical data meaning all the changes to the data or only the days-end status e.g. End Of The Day for each data revision from Archive area or Input data area. An entity Person can have a one-to-many relationship to the addresses of the person’s living places in another PersonAddress entity with ValidFrom and ValidTo columns and these entities or tables will be merged into one type 2 dimension table according to Kimball in a data mart. The modeling is subject-oriented meaning all business processes for each subject e.g. client needs to be modelled. A EDW common data model is not technical it must be based in (logical) business processes. Since all source legacy system data is integrated and collected into one common database, the EDW contains and represents a »single version of truth« for the enterprise.
Let the design of the data model for a EDW start doing a »helicopter view« for identification of entities, e.g. salesperson or engineer will be stored in an Employee entity with a one-to-many to a Jobfunction entity, or maybe do more abstraction in a bigger picture making a Person entity where employees and customers users can be stored together. Later in the design process move on to a »weeds view« for all the details for columns, conformed names and data types etc. In the EDW keep the source data with time span columns like Start date and Stop date, and sometimes also convert the source data to row wise entity as a kind of transactional data that is a sequence of information exchange like financial, logistical or work-related data, involving everything from a purchase order that becomes an invoice with multiple invoice line items, to shipping status, to employee hours worked, plan and activity records, for subscription period and to insurance costs and claims. The EDW will be the source to one or multiple data marts using Dimensional modeling with denormalized data controlled by the ETL process. Or maybe the data mart is using a different modeling that fit better for use of data. An enterprise data warehouse should be accountable and auditable which by default means pushing business rules of changing/altering data downstream to »between the data warehouse and the data marts«. EDW tables and views can be data source for a tabular modeling with a DAX formula to do filter and calculate in-memory KPI to be visual in Power BI, and a data mart area is skipped or been replaced by EDW views, therefore it is called a ELT process because the transformations is executed on the way to the user.
1.3.2. Anchor modeling
When a EDW is modeled after Anchor modeling it is based on four elements that holds on historical and time-variant raw data in entities/tables with labels:
Anchor is a surrogate key generator of a key from a source legacy system that we are called a business key from a data warehouse point of view.
Attribute has descriptive values from a source legacy system connected to the business key therefore attribute has a foreign key reference back to the anchor. The attribute table is based on type 2 historic, therefore it has also a surrogate key. A view will mix anchor and attribute and it becomes a type 7 dimension because surrogate key of anchor becomes a durable key and surrogate key of attribute becomes a type 2 surrogate key. Both keys will be added to a fact, and the fact can join to a most recent dimension to show the current value or join to a historical dimension to show the registered value when the fact data occurred with its transaction date. The model in Power BI can choose the content for its presentation interface because all data is on a silver platter.
Knot is a lookup table with basic data.
Tie is a relationship table between anchor tables to handle one-to-one, one-to-many and many-to-many relationships, therefore no foreign key in anchor or attribute table except to a knot table (»tie the knot« means getting married).
Anchor modeling has only extension of new tables and none modification of existing tables. This ensures that existing data warehouse applications will remain unaffected by the evolution of the data warehouse. When a business key exists in multiple source legacy systems, there will be one common anchor table and several attribute tables, because each source legacy system has its own attribute table.
Anchor modeling allows to build a real-time ETL process to a data warehouse where some attribute tables needs a more frequent update than others to provide fresh data to the users in a report or for a online dashboard. Other attribute tables only needs updating per hour or at end-of-day.
1.3.3. Data vault modeling
When a EDW is modeled after Data vault modeling it is based on component parts of a Core Business Concept (CBC) as an ensemble consisting of three components that holds on historical and time-variant raw data in entities/tables with labels:
Hub (blue) separates the business keys from the rest of the model and will translate business key to a unique hashbyte key value. A composite business key of multiple columns will also turn into one hashbyte key column. E.g. a HubProduct has ProductHashKey together with business key ProductNumber. A hub is an integration point of a business key or a unique identifier and will never change. A hub exists together with at least one satellite. A data vault table has a RecordSource column as a reference back to the source legacy system e.g. "Dynamics365.Sales.Product" and can be a multi-valued column with multiple sources. Avoid to composite business key and RecordSource, use instead a Same-as-link table (mapping). (Hub in danish er et samlingspunkt for én forretningsnøgle der har data fordelt over flere satellitter som kredser om en hub eller et link).
Satellite (yellow) [sat] stores the context, descriptive data and measure values in columns of either a hub or a link. A satellite is connected to one hub or one link, and a hub or a link can have multiple satellites. E.g. a SatProduct with foreign key ProductHashKey from HubProduct and data values for Name, Category, Target group etc. Satellites has all the relevant data for the data warehouse. Satellite is history-tracking and to handle historical data in a satellite, the primary key is composite of HashKey+LoadDate (Load_dts, LoadDateTime, ValidFrom) column so the same value in e.g. ProductHashKey can be repeated several times. A hash difference HashDiff column is a checksum of all data value columns for making an easy comparison for Kimball type 2 history, see more in section 4.3.
To be 100% insert compliant there is no columns for EndDateTime, ValidTo and IsCurrent flag, therefore no updates of rows in a satellite. If a source legacy system can tell that a product has been expired or deleted then a new row is inserted into SatProduct with filled DeletedDate.
One hub or one link will usually have several satellites associated because we will regroup data into multiple satellites by classifications and types of data and information and by rate of change, so each of them can have its own granularity and timeline. Split logical groups of data into multiple satellites. With multiple source legacy systems the RecordSource column is helpful when data sets is similar, else let each source legacy system has its own satellite and share the HashKey from the hub to all its satellites. For example, EmployeeHashKey is in all the sat tables:
SatEmployeeBasic(SSN, BirthDate, MaidenName) constants/correction overwrites,
SatEmployeeName(FirstName, MiddleName, LastName, Gender),
SatEmployeeContact(CellPhoneNumber, Email, Skype,Twitter,Facetime,Whatsapp),
SatEmployeeAddress with columns for address or a reference to an Address table,
SatEmployeeJob(FromDate, ToDate, Department, Title),
SatEmployeeHoliday(Date, Occasion) a weak entity without a business key in SLS,
SatEmployeeSalary(DateOfMonth, Amount) and
SatEmployeeChristmasGift(Year, NameOfGift, AmountOfGift, LevelOfSatisfaction).
The above data comes from multiple source legacy systems and data is added independently of each other to the individual satellites (divide and conquer). There will be outer join from a hub to its satellites and inner join from a satellite to its hub.
Conformed and calculated data is placed in its own Computed Satellite. Other satellites are: Overloaded satellite, Multi-active satellite, Status tracking satellite, Effectivity satellite and Record tracking satellite.
Link (green) integrate and capture relationship between hubs and links. E.g. a product is placed in a store, therefore we have a HubProduct and a HubStore with data values in a SatProduct and a SatStore, and a LinkProductStore represents the placement by combining ProductHashKey and StoreHashKey as foreign keys from the hubs to capture in which stores a product is placed and for sale. A link creates its own hashkey as a unique combination of the involved hub business keys e.g. ProductStoreHashKey from ProductNumber and StoreCode. Data values connected to a link is placed in its own satellite, e.g. SatProductStore with primary key/foreign key ProductStoreHashKey and data values for PlacementInStore and QuantityInStore. When a product no longer exists in a store, the ETL process will insert a new row in SatProductStore with the expiration date in the column DeletedDate.
A link handles one-to-one, one-to-many and many-to-many relationships because data vault has only optional many-to-many. A link is a dataless and timeless connection among hub business keys and is not a data-bearing relationship from ERD. A satellite on a link represents the history of the connection or the relationship.
An example of a Product table with foreign keys SupplierID and CategoryID will often be put together into a LinkProductSupplierCategory with ProductHashKey, SupplierHashKey and CategoryHashKey from HubProduct, HubSupplier and HubCategory and a primary key of ProductSupplierCategoryHashKey+LoadDate. Over time a product can change supplier and category which is being handled in the SatProductSupplierCategory with a new row inserted with filled DeletedDate and a new row is inserted into link and sat for the new data combination with a new LoadDate. If the business one day allows a product to have multiple suppliers then the link is already prepared for it. In the future, the business will like to place products in stocks all over the world, and it will be registered in a new inventory management system, that creates a HubStock and a SatStock, and creates a LinkProductStock and a SatProductStock to handle all the new data and relationships.
A unary/recursive relationship (self join) e.g. LinkMarriage represents a connection between two persons with Person1HashKey and Person2HashKey from HubPerson with business key SSN and it gives primary key MarriageHashKey. All data are placed in SatMarriage(MarriageHashKey, MarriedDate, DivorcedDate, WidowedDate) that is labeled Effectivity satellite because of the columns of dates. A validation rule in SatMarriage ensures that one person can only be married to one person at a time else it will be marked as an error for an audit trail to catch. When a person enters a marriage and changes last name, it becomes a new row in SatPerson. When a marriage ends either by divorce💔 or by death† then a new row is inserted into SatMarriage with filled DivorcedDate or WidowedDate. Couples who are divorced and later remarry each other will reuse the same MarriageHashKey value from LinkMarriage for the new row inserted into SatMarriage. The LoadDate will be in order of the events. The DeletedDate will be used in connection with an error detection in the source system or an annulment of the marriage.
Another modeling of marriage: SatMarriage(MarriageHashKey, Date, MaritalStatus) with values of marital status: Married, Divorced and Widowed, makes it easy to extend with extra values e.g. Separated, Registered partnership and Abolition of registered partnership. When persons as partners are living together we will modeling it in a LinkCohabitation and a SatCohabitation. When persons only dates we make a LinkRelationship and a SatRelationship. Therefore each link and satellite is a flexible and agile way to divide data (agile in danish som smidig for forandringer, let og hurtig at udvide). In a generic perspective »helicopter view« we could have started with a general link to represent a connection between two persons with Person1HashKey and Person2HashKey from HubPerson, and label the link a general name like LinkCouple with a CoupleHashKey, and let the types of data create multiple satellites for the link as SatRelationship, SatCohabitation and SatMarriage. It is normal for a couple to start dating with data in the SatRelationship, living together with data in the SatCohabitation. When the couple gets married, the cohabitation timeline stops in the SatCohabitation, and the data recording continues in the Sat-Marriage. It is agile to think and data modeling in a general way and divide data.
An invoice becomes a HubInvoice where business key InvoiceNumber becomes InvoiceHashKey used in a SatInvoice with data values IssueDate, DueDate, DeliveryDate, PaymentMethod and PaymentDate. When a payment happens e.g. a week later, it will become a new row in the SatInvoice, or it will be added as a row to a SatInvoicePaid with InvoiceHashKey, PaymentDate, PaymentMethod, PaymentAmount to divide data in seperated satellites because of the time difference in data capture and allows for payment in installments in the future. If one day the sales department replaces the billing system then create a new SatInvoice. An invoice involves a customer in a store, therefore a LinkInvoiceCustomer includes InvoiceHashKey, CustomerHashKey and StoreHashKey and combines business keys to InvoiceCustomerHashKey used in a SatInvoiceCustomer with data values like Reference no, Requisition number and Voucher number. An invoice line item involves a product, therefore a LinkInvoiceProduct includes InvoiceHashKey and ProductHashKey and combines business keys to InvoiceProductHashKey used in a SatInvoiceProduct with data values LineItemNumber, Quantity, CouponCode, Discount (%), UnitPrice, VAT and Amount. If the quantity in an existing invoice is corrected then a new row is inserted into SatInvoiceProduct. If a product is replaced then for the old product a new row is inserted into SatInvoiceProduct with filled DeletedDate and the new product is inserted into LinkInvoiceProduct and into SatInvoiceProduct.
A link does not have its own unique identifier from a source legacy system. Other links are: Nondescriptive link, Computed aggregate link, Exploration link and Link-to-link or Link-on-link.
Dan Linstedt comments my post at LinkedIn that he is using color: green for hub, blue for link and white for satellite.
Address will not be a Hub since the address itself is not a CBC but a description of a CBC e.g. Customer, Employee or Store. A CBC could be a CustomerDeliveryLocation as a hub with a satellite and a link to Customer hub because a customer could have more than one active location where goods can be delivered. The business key for the hub is using a concatenation of zip code, street name, house number to get a unique identifier for the CustomerDeliveryLocationHashKey.
There is several other types of entity/part/table, for example.
Transactional-link (light green) integrate and capture relationship between hubs and links to capture multiple transactions that involve the same business keys e.g. many sales to same customer of same product from same store, meaning multiple rows with the same set of keys. A transactional-link has its own unique identifier for a transaction (TransactionId, EventId) from a source legacy system. Notices that a transactional-link keeps its business key, it is not placed in an associated hub. If a revision of a transaction is received from a source legacy system or a counterpart then insert the data values into the associated satellite. Since DV raw data is a structured archive, it will be wrong to calculate a counterpart row, but when data is extract-transform-load to a Transactional fact it is fine to calculate a counterpart for the fact.
Nonhistorized-link is for immutable data that has no edit history, in other words, constant data that will never be changed or deleted in a source legacy system. Each transaction (data row) has a unique identifier (TransactionId) and there is no revision of a transaction in the source legacy system. When there is a counterpart to a transaction it will have its own unique identifier and a reference identifier back to the original transaction.
Reference (gray) is referenced from a satellite and is the same as a lookup table with basic data e.g. RefDate or RefCalendar with many columns like Week, Month, Quarter and Year, RefMaritalStatus and RefMarriageType with Code and Text, and RefZipCode with Zipcode, City, State, State abbreviation for US. Date table is called a nonhistory reference table. A satellite is normalization to 3NF and contains only non-foreign key columns except simple reference values. If a reference (cross-reference or lookup table) contains history then move the old data values to a History-based reference satellite with primary key Code+LoadDate and Text contains the previous texts. I prefer to have a RefAddress with extra columns like State, Region, Municipality and Latitude and Longitude (gps coordinate). Kimball page 235.
Same-as-link (turquoise) [sal] to map different business keys from multiple source system where each source has its own satellite. In a sal we can »merge together« differently named business keys to a single master key that is really the same-as the other business keys. Same-as-link is connected to a hub and is used to identify when the same business objects are identified by multiple business key values, I shown an example for a SalesDevice in section 1.1.
Hierarchical-link (silverblue) [hal] to capture hierarchical relationship between the business concept records in same hub e.g. parent-child hierarchy.
Point-in-time (purple) [pit] is a helper table to a hub or to a link and is calculated by the ETL process based on all the satellites of the same hub or of the same link, because these satellites share the same HashKey together with different values in LoadDate. A pit table has its own LoadDate column (SnapshotDate), a column for the shared HashKey and columns for each involved satellite’s LoadDate, therefore all the values of LoadDate from the satellites is represented in the pit table to be used to derive all valid versions for ease of joins of satellites and to improve the performance and includes to find the newest or most recent rows and to calculate EndDate or ValidTo. By adding a surrogate key sequence number to a pit table it can through a view create a virtual Kimball type 2 dimension.
Bridge (taupe) combines (or bridge) hashkeys and optional business keys from multiple links and their related hubs. A bridge can represent an aggregated level and can include derived and calculated columns.
Data vault 2.0 recommends for a Hub to hash a business key value instead of using an identity column as a unique sequence number because:
· HASHBYTES('MD5',CONCAT(SSN,';',LOWER(TRIM(Area)))) binary(16)/char(32).
SHA1 use binary(20), SHA2_256 use binary(32) and 512 bits takes 64 bytes.
· Before hashing a business key we do data operations as lower case, left and right alignment or trimming and save it in a column called »business key comparison« together with the original value in a Hub table.
· For Hash diff to make an easy comparison for all the other columns, we do the same data operations when it is agreed with the business and the data profiling.
· Hashing two values could create same hashkey value which is called a collision and is very unlikely but can’t be left to chance or avoidance. If a data batch contains two different business key values which give same HashKey then bulk insert will result in primary key violation in a Hub. If a new business key value gives the same HashKey as an already existing one in a Hub then the loading can assume the new business key value is already there. Therefore first to check in Hub if new business key value exists and the HashKey exists then increase HashKey until no collision or using a collision sequence number, but there is no guarantee of same hash key values when data is reloaded. Maybe a need for a hash key detector detection job and detecting an undefined hash key. What to do when business key value is null or empty string?
· It is ease of copying data from environment to another like dev to test to prod without worrying about sequences being out of sync.
· When two source legacy systems share same business key but in one source it is an integer data type and in the other source it is a string, by hashing they become same data type in the hub table and sat table.
· When a source legacy system is replaced, the new source comes with different data types e.g. from an integer business key to a string business key, it will still be hashed to same data type in the data vault, therefore no maintenance.
· A composite business key of multiple columns become one hashkey column which is easy to use for join and merge.
· Easy to create a hashkey of a business key value when needed by a ETL tool or in a sql select statement instead of doing a lookup to find the sequence number stored in a table. But of course still need to make sure that the same business key is not inserted twice in same hub table and avoid collision.
· Loading to tables can be done parallelly e.g. at same time to a hub table and to a satellite table from the same source legacy system table (parallel loading, enables parallelization). It sounds awesome! What if a ProductNumber is used in a Store relationship and will be inserted into the LinkProductStore with ProductHashKey and StoreHashKey but the ProductNumber is an orphane and does not exists in the Product table from a source legacy system and therefore is not in the HubProduct then it gives a referential integrity violation in the data vault, read about inferred members in section 4.5.
I find the Data vault model as an archive of structure where data from source legacy systems has been divided, and where the business rules and transformations is first enforced when data is going to data marts.
Linstedt says: »The Data Vault is for back-end data warehousing, not for production release to business users directly.« DV becomes a persistent staging area or an archive area (ARA).
Querying a data vault required many more joins than Inmon's model. Be aware it can be a performance issue to have columns of data type binary(16) or char(32) for storing millions of rows and for join many tables together compared by using an integer data type. Joining large links and hubs together is expensive for performance and it is often debated by Linstedt. As far as performance goes, normalised structures are no where near as performant as a dimensional model. Nothing beats dimensional modeling for performance of queries.
Data vault divides data in a Raw area and a Business area, where the last area contains business calculations and to look at data on a different level of granularity by making new link with more or fewer hubs (and business keys) involved. Pit and Bridge is part of the Business area to provide table to give a better query performance. For a ETL process from a Data vault to a Dimensional modeling, think that a Hub and its Satellites become a Dimension, and a Link and its Satellites become a Fact. Pit and Bridge is a big help and sometimes the ETL is done by views only.
When to use Data Vault? When a data warehouse has many source legacy systems!
1.3.4. Agile modeling
Bill Inmon's modeling is based on a normalized data model and the new data models for EDW wants to break things out into »parts« for agility, extensibility, flexibility, generally, scalability and productivity to facilitate the capture of things that are either interpreted in different ways or changing independently of each other. Normalization to sixth normal form 6NF is intended to decompose table columns to irreducible components, a hypernormalized database where essentially every column was its own table.
Agile modeling has only extension of new tables and none modification of existing tables. Changes in an agile data warehouse environment only require extensions, not modifications and no impact to the data warehouse and it becomes quick-moving to act quickly to easily customize the data warehouse when business changes. An ensemble modeling pattern gives the ability to build incrementally and future changes should not impact the existing design. Agile development is performed in an iterative, incremental and collaborative manner. I think we will always meet an issue that address us to do a model refactoring or re-engineering. Read more. A generic data model does not need any changes when there is a change in the source legacy systems. There is a Party modeling, a 2G modeling and a Focal Point modeling and more than dozens of data warehouse data modeling patterns that have been introduced over the past decade.
EDM = Enterprise Data Model is a term from IBM back in 1998 for a data modeling technique for data warehousing, where ETL was called Capture, Transform, Apply.
A simple Guidedance for EDW: Do not lose data (auditing/compliance). Model it, so you can understand it. Make it physical, so you can query it.
Data warehouse is using a surrogate key instead of a business key to remove dependence from the source legacy system, see more in section 4.2.
1.3.5. Dimensional modeling
Ralph Kimball does not like to store data in a EDW, he only store data in data marts that is using Dimensional modeling, therefore EDW becomes a union of all data marts.
Dimensional modeling ends up in a star schema or constallation schema (multi-star schema) with fact tables (analysis variable, measures, events) surrounded by dimension tables (context), where dimensions explain the facts. Dimension and fact conformance is a must in a successfull data warehouse implementation to meets the requirements of legislation, accepted practices, prescribed rules and regulations, specified standards and terms of a contract.
1.3.6. Data warehouse architecture and ETL modeling
Three business intelligence enterprise data warehouse modeling architectures:
Sources represent either Input data area (IDA) or Archive area (ARA).
It is common for all models to have a Data staging area (DSA).
Enterprise Data Warehouse (EDW) is sometimes a term for the whole solution and the different models each use their own terms e.g. Data warehouse or Raw data.
EDW is a common database layer before the Data mart area (DMA).
A data mart has a Presentation interface (PIA) in front, above it is called BI-Apps.
I have used the general word data mart for a database with tables of dimensions and facts, but there is a more narrow word Multidimensional database (MDB) that is the source database for a OnLine Analytical Processing (OLAP) application like a cube also known as a Multidimensional online analytical processing (MOLAP) application. For me, it is a principle for a cube that all dimensions and all measures can be combined freely else divide data into multiple cubes.
ETL process stands for Extracting, Transformation, Loading and exists between the data areas of a data warehouse solution, e.g.:
From each data area (layer) there will be extract criteria to limit data, and there will be transforms to make data conform and calculate new data like KPI, and data will be load into the next area. It is a common way of handling data in a multitier architecture or layered Architecture. (In danish lagdelt arkitektur.)
1.3.7. Data mart modeling
A data mart area will consist of several data marts database. Each data mart database has a specific subject area data for tailored support and can be characterized as one of the categories below:
with common dimensions to be reused in the other data marts. Examples of dimensions is Date, Time, Employee, Organization, Product, Retailer, TransactionType. Can also include role-playing dimension e.g. Manager based upon Employee and different date dimensions with unique names of columns.
For example, Sales mart, Customer mart, CRM mart, Churn prediction mart, Market mart, Production mart, Inventory mart, Shipment mart, HR mart, Tax mart, Fraud detection mart.
with views and/or tables making conformed columns in a Kimball dimensions and facts model. Fetching data from EDW in a ETL process to provide data to OLAP or Tabular cubes or to visualization tools like Tableau, QlikView/Qlik Sense and Power BI. The data mart can be used for analytics by business users and data scientist to do ad hoc sql query.
with views making conformed columns and fetching data from EDW and could display data in a Kimball dimensions and facts model. Stored procedures with criteria parameters e.g. FromDate, ToDate, Customer, Branche, Product, SalesDevice, TransactionType will fetch the relevant data from the views and present data to a reporting tool used by business users like Reporting Services (SSRS) or Microsoft Access with pass-through query to call a parameterized stored procedure.
with views and/or tables fetching data from EDW to create data in a format that is customized to the individual data recipient as another IT system internally or externally.
with tables of raw data from source legacy systems, views on an archive area or views on a data warehouse area layer (edw) for some business users to do data discovery by sql statements. It is a sandbox where users also has write access and can create their own tables for staging data and for the final result dataset to Excel or ad hoc Power BI. The purpose is to prepare a good business requirements specification and a user story for extension to the data warehouse solution.
Data mart area (DMA) has several databases for each mart and sometimes there is no need to store data in a mart database instead the mart database contains virtual star schema non-materialized views that look like dimensions and facts.
Remember to set a process for user management, permissions and authentication etc. to manage and monitor and reduce the risk of cross contamination of data to ensure that a user only has access to view relevant data for his analyzes and reports.
One part of a data warehousing project is to provide compliance, accountability, and auditability. After data capture, please remember to implement a Reconciliation Summary Report with the results from your recon e.g. Uncategorized assets, Invalid country codes, Derivatives transactions with missing currency code.
Audit trail (in danish kontrolspor) becomes important for the credibility of a data warehouse. An example from a source legacy system that has a value 0.003589 and export it to a txt file where the value becomes 3.589E-3 and by a mistake in the ETL process the data warehouse saved and displayed the value as 3.589. A contract number 700002848572 becomes 7.00003E+11 and the last part of the value got lost. When reconciliation is built-in the data model and the ETL process, this mistake would be reported and the programmer can fix the import and update his data profiling documentation.
1.5. Data quality
Some data quality skills and data cleansing examples:
1.6. Big data
Some properties of big data:
A data lake can be used for massive quantities of unstructured data and big data with tools that can easily interface with them for analysis for business insights. A datum in a lake has tags to give it a characteristic and by the tags we can fetch data from the lake without knowing the physical location like a server url with a folder path. A data lake can contain files on multiple servers on premise in different folders and in the cloud (many nodes), and we only using a tag to find and fetch data. For example, I like to find photos of smiling employees in all albums, I can search for a tag FacialExpression = smiling. A data lake is using ELT (extract, load, and then transform). A tool for a data lake can be like Apache Hadoop or Microsoft Azure. Data Discovery Area is an end-user sandbox. Can use U-SQL to dive in the data lake and fetch the wanted data and do the wanted transformations.
A relational database is characteristic by ACID (Atomicity, Consistency, Isolation, Durability). This means that a transaction is either carried out completely or not at all (Atomic), that only valid data is added to the database (Consistent), that transactions never affect each other (Isolated), and that transactions are never lost (Durable). Read more.
A NoSQL = Not Only SQL database means it can use SQL type query language, but usually do not do so. NoSQL database often designed to run on clusters, made by open source and the database does not operate with a fixed schema structure but allow the addition of data without a pre-defined structure. A NoSQL database is characteristic by BASE (Basic Availability, Soft state, Eventually consistent). A ACID system guarantees data consistency after each transaction; a BASE system guarantees data consistency within a reasonable period of time after each transaction. In other words, there is data consistency in the system, just not immediately. This leads on to the Soft State principle. If the data is not consistent at all times, the system must take a temporary data state into account. The sum of both these principles means that data accessibility is given very high priority, even if coincident errors occur in the database system, operating system or hardware. If parts of the database does not work, other parts of the database take over, so that data can always be accessed.
For EU GDPR General Data Protection Regulation (Persondataforordning) a business key like a social security number needs to be anonymous after some year by adding a calculated cross sum of the ssn and hash it. It will be a nice scramble or encryption by a scrambling algorithm. To be sure of total anonymity only save Birthday and Gender of a person; no name, no address, no zipcode and no city. A Master Data Management (MDM) database can contain social security number, name, address together with CustomerId that is used in emails to customers. Therefore we have a surrogate key CustomerInternalId that is used in all other operational systems and data warehouse together with Birthday and Gender which is not personally identifiable or enforceable. A web service will provide all systems the data of personally from MDM database. When a customer is going to be deleted because it is his/her wish or the data limitation period, we only need to anonymization (the way of delete data) in the MDM database and we can keep CustomerInternalId, Birthday and Gender in the other systems to make sure statistics remain unchanged back in time. When a system is calling the web service it will give back unknown for ssn, name and address when a customer no longer exists in MDM database. If we don’t have a MDM, we must do anonymization in all data layers and in relevant system legacy systems as well, and flag data with an IsAnonymized column. Personal Information (PI), Personally Identifying Information (PII), Sensitive Personal Information (SPI) is data relating to identifying a person, read more.
2. Grain (granularity) of data warehouse
The grain is the level of detail of data in the table, both in dimension tables by the hierarchy and in fact tables by the definition of the measurement event and of interest, and the grain can later be expressed in terms of the dimensions. The lowest grain keep all relevant data from source legacy systems and is therefore the most flexible approach but also take most storage space and easy can cost high query performance. To improve query performances the grain can be lifted up to higher level while data will be aggregated and summarized and therefore take less storage space. Data can also be divided such as current year data is in grain daily (dately), previous year in weekly and older data in monthly, because very detailed information is normally not relevant for analysis years back in time. Granularity of fact table can be divided into four types:
Grain yearly to monthly to weekly to daily we say that each level e.g. daily increase the granularity and the number of rows in the fact table.
Aggregation is the process of calculating summary data from detail base level table rows (records) and is a powerful tool for increasing query processing speed in data marts. For example, a sales is a fact with analysis variable and measures like quantity sold and amount, and dimensions like product, customer and date of purchase brings a sales in a context. The grain of the sales is limit to a date (like December 23) and a time (like 4:30 pm) and therefore the fact is on Transaction grain. In case we drop the time, the measures would be called TotalQuantitySold and TotalAmount because they are the result of a summation of sales times to a sales date and therefore the fact is on Periodic grain. Also if we decide to summarize the sales date to a weekly or montly level. In case we decide to grain the customers by aggregate them into segments and don’t keep names and addresses, then the fact becomes Accumulating/Aggregated grain. A Product dimension has a three level hierarchy of category name, brand name and product name, and therefore we can say that product has the highest level of detail. When a source legacy system only gives a sales transaction with a brand name and no product name, a Product dimension must fit that grain of fact and take out the product name level so brand name becomes the lowest level that match the fact grain.
Instead of summation of sales times or sales date and loose some important information of time of purchase, we can summarize data into two other fact tables for a weekly level for data from the last year to this year, and a monthly level for older data, because when we go back in time, we don’t need to analyze on a daily or weekly level, and by aggregation we save harddisk space and improve the performance of the query because fewer rows in month level need to be summarized to fetch the year level data.
A business key has the same granularity and the same semantic meaning across a business organization company.
Granularity is a description of a level of detail of data, e.g. the combination of customer and product is a fine level to tell me about sales, and when I add store to the combination the granularity is driven one level lower, and when I add salesperson to the combination the granularity is driven one level lower and I know much more about sales, and when I add a date or better a datetime it will bring the granularity to its lowest level or to the highest level of detail of sales and a sale.
The lowest level of aggregation or the highest level of detail is referred as the grain of the fact table.
A fact table can contain fact’s data on detail or aggregated level depends of the grain approach. A fact table can have five types of columns:
The ETL process has to secure the key, so fact rows are distinct. Fact tables in a dimensional model express the many-to-many relationships between dimensions and is implemented as one-to-many relationships between dimension tables and fact tables. I never have foreign key constraint on a fact table because it decrease inserting performance and I trust the ETL process and date range lookup, and no human being will be doing a update or delete of a dimension table or a fact table.
3.1. Types of facts
Let us characterize the various facts into different types of facts or more exactly different types of the columns in a fact table. Conforming facts means making agreements on common business metrics such as key performance indicators (KPI) across separated source legacy systems so that these numbers can be compared mathematically for calculating differences and ratios.
Fully-Additive measure - summable across any dimension
A fact table has numerical measures that can be summed up for all of the dimensions in the fact table, so the measure columns data type is a number. A sales fact is a good example for additive fact with measures like Quantity sold and Amount. In case of a transaction dataset to a fact table refer to a measure column which value is empty, null or nullable, use the default value 0 because this won't bother aggregation like summation. Each measure must have its metrics. If it is a monetary measure, it may have a currency column and if it is a unit measure it may have a column to explain the kind of units used like centimeters, litres, cubic metres etc. Fact can have a calculated measure or a derived measure based on existing measures and constants e.g. Profit or Surplus = Revenue – Costs.
Semi-Additive measure - summable across some dimensions
A fact table has measures that can be summed up for some of the dimensions in the fact table and not for other dimensions. For example, a daily balance measure can be summed up through the customers dimension but not through the date dimension. Inventory levels cannot be summed across time periods.
Non-Additive measure - not summable for any dimension
A fact table has measures that cannot be summed up for any of the dimensions in the fact table. For example, a room temperature fact is non-additive and summing the temperature across different times of the day produces a totally non-meaningful number. However, if we do an average of several temperatures during the day, we can produce the average temperature for the day, which is a meaningful number. Sum of a measure called DaysForOrderToCompletion in a FactOrder is meaningless but finding a minimum, maximum and average values is meaningful to planning of production. Other examples is a percentage and a ratio are non-additive measures. A fact table that only contains column with transaction number such as order number, invoice number or a voucher number that can’t be summing up. A order fact with measures like Unit price for a single product makes no sence to summarize, but the derived column Amount = Unit price x Quantity is to be summarized and becomes an additive column, called a calculated measure. Trend, Stock and Ranking can't be added and in general all calculations on one specific intersection of the dimension. Year-to-Date ytd measure can’t be summed up. Count of rows is normally used. Unit price can be placed in a Product dimension as a current standard list unit price and still keep Unit price in the fact together with a Discount (%) when the purchase happened, occurred or took place. Back in 2016 I had a customer that used two kinds of discounts when a product had a bundle option and was showned in an invoice line with description »Discounted Bundle«:
Unit price = List price - Unit discount amount.
Line amount = Unit price x Quantity – (Unit price x Quantity x Line discount percentage / 100).
Some business users used the term List price other users said Catalog price.
When there was a promotion going on, the unit price had a different calculation:
Unit price = List price - Proration amount (also called deduction amount).
There was also a Line discount amount and an Onder discount amount. I added all the columns into the Sales fact.
A conformed measure in multiple facts must use the same common business rule and definition so multiple facts can be united in a report or a cube. When several data marts are using fact data with same name of fact tables or name of columns for measures and they have compatible calculation methods and units of measure and support additivity across business processes. If a measure e.g. Revenue is used in multiple fact tables with different calculations and meanings, it is best to use different column names because theses facts are not conformed.
A fact table that contains no measures is called factless or measureless fact. For example, a fact table which has only columns for employee, date, time and type, where type is like »workstart«, »workstop« and »workstopofsickness« and there are no columns for measure. You can get the number of employees working over a period by a »select count(distinct EmployeeId)« or by a distinct row-count calculated measure in your OLAP cube. For a factless fact you will normally count the number of rows, row count or counting rows and call it »Number of <a name>«. Sometimes a factless fact has a value column called Count with only one value as 1 used in a data access tool to sum over and get the number of rows. In case the fact grain is weekly and a week is missing, it can be inserted to have all weeks complete and here will Count gets the value 0. Factless fact is for registration of event or assignment e.g. attendence take place in a school class with dimensions for student, class, room and professor. If we add a measure column for Attendence with 1 or 0 per date per student it is not a factless fact anymore. Factless fact can be used to represent a many-to-many relationship among dimensions. Factless fact can have a ValidFrom and a ValidTo column to tell the timeline/period when the fact was valid.
Capture a relationship in the fact
To be a column in a dimension or to be its own dimension and used in a fact is a good question. Kimball’s financial services example starts with an Account dimension including data of products and branches but he choose to remove these descriptive columns to form independent dimensions of Product and Branch and use them in a fact together with the Account. Therefore the fact capture a relationship among accounts, products and branches. Another example is that an account can belong to two customers and a customer can have several accounts. This many-to-many relationship can be expressed in a factless fact or in a bridge table, see later. A bank transaction is done by one customer from an account and it is natural to have a Customer dimension in the fact.
Kimball says: »Demoting the correlations between dimensions into a fact table«, and I like to add: »With capture a relationship in the fact table we also keep the registered relationship at the time the event or transaction occurred and was entered into the fact table«.
An example is a retailer SuperChemi belongs to a chain Ethane at date of sale e.g. 2013-05-17 and the fact table has two columns for dimensions to Retailer and Chain. In 2015 the retailer SuperChemi changes to another chain Propane but we still keep the registered relationship back in 2013 and 2014 in the fact. When a chain is not a part of the fact table and we in year 2016 like to find sales of retailers for a specific chain e.g. Propane, we will use the current relationship between dimension Retailer and dimension Chain as a snowflake dimension, meaning that SuperChemi belongs to Propane, and when we summarize sales per year since 2010, chain Propane will include the sales of SuperChemi in 2013 even though the retailer belonged to Ethane at that time. Kimball also says: »Combine correlated dimensions into a single dimension«, e.g. Retailer and Chain into one dimension with a hierarchy Chain→Retailer including history to handle a retailer changes chain or a retailer changes name or changes data in other columns.
It is always a consideration to have one dimension or to have several separate dimensions e.g. a many-to-many relationship between addresses for ship-to and bill-to or between name of sales rep and customer, it is best to handle as separate dimensions and keep both of them in a fact table row, read more in Kimball page 175-76:
Transactional fact or Transaction fact
A fact table that describes an event or operation that occurred at a point in time in a source legacy system e.g. an invoice line item. The row has a date e.g. a transaction date, an entry date or a post date or a datetime time stamping to represent the point in time with other lowest-level data. Key values for dimensions is found at transaction datetime. If data needs to be changed, corrected or cancelled in source legacy system, then a data warehouse needs to make a counterpart row in the fact with a new time stamping or a time span. Source data examples, booking a hotel room, order a car in a special color, buy a lottery ticket with my lucky numbers, receive an invoice, put and pick items in stock (in danish lægge varer på lager og plukke varer fra lager). In a data warehouse modeling process I try to interpret some data as transactional data or transactions. See more in section 6.4. Some data modeler does not like separate facts for each transaction type but build a single blended fact with a transaction type dimension and a mix of other dimensions can make N/A dimensions. I have seen a blended fact called FactEvent which I think is a poor and non-signing name of a fact table, and the date dimension gets a generalized name. I prefer multiple facts where the date dimension is role-playing for e.g. order date, purchase date, shipping date and receipt date.
Periodic snapshot fact
A fact table that describes the state of things in a particular instance of time or a one point in time, and usually includes more semi-additive and non-additive measures. It is a table with frozen data, meaning a row will never be changed/modified/deleted (is unchanged) because the row can have been used in a report like a monthly or annual report and later it is a must to be able to create the same raport with the exact same data. The periodic snapshot fact table will never be done empty or updated, therefore it is a true incremental load and the table has a column like a Month (yyyymm) for a monthly basis or a Year to fetch the wanted data as a month slice or a year slice. When a month or a year is over and data is ready, data will be loaded. Sometimes data for the current month is also loaded every day for a current-month-to-date, therefore the current month will be updated until it is over, finish and can be closed. Measures can be a balance in account or inventory level of products in stock and so on. Key values for dimensions is found at the end of the period. Fact table can be monthly partitioning for making a faster query performance when searching for a month. In case there is no data for a specific month it will be inserted as an artificial row with »missing« value of the dimensions and 0 in the measures. (Periodic snapshot in danish »en skive data«, fastfrysning.)
Accumulating snapshot fact
A fact table that describes a process with milestones of multiple dates and values columns with different names which will be filled out gradually. Therefore over time the same fact table row will be revisited and updated multiple times where a default date key value is -1 for »hasn’t happened yet« or »be available later«. Each stage of the lifecycle has its own columns e.g. milestones of a hospitalization or steps of the manufacturing of a product. In a retail store a product has three movements as ordered, received and sold that would be three date dimension columns in an accumulating snapshot fact.
A fact table where a row is a summarize of measurement events occurring at predictable steps between the beginning and the end of a process. For example, a source legacy system for current payments from customers where some pay several times over a month, and the first payment becomes a new row in fact table with date of payment in columns BeginDate and EndDate and the amount in Paid column. The next ETL process will do a summarize of payment per customer from the Begindate to current date or end-of-month date, and then update the fact row with same BeginDate with the new summarized payment and new EndDate, so a fact row will be revisited and updated multiple times.
Derived fact or Additional fact
Kimball recommands fact data at the lowest detail grain as possible for ensures maximum flexibility and extensibility, he call it a base-level fact. A derived fact table is created for performing an advanced mathematical calculation and complex transformations on a fact table like for a specific KPI (Key Performance Indicator), or an aggregate fact with a summation of measures to a higher grain like from date level to month level and from product level to brand level and using shrunken dimensions for Month and Category as a dimension lifted to a higher grain from the base-level dimensions as Date and Product. A derived fact can be based on multiple fact tables for making faster ad hoc query performance and simplify queries for analysts and for providing a dataset to the Presentation interface area.
Aggregate fact or Summarized fact
A derived fact table that is created to referred to as a pre-calculated fact with computed summarized measures at a higher grain level of one or more dimensions to reduce storage space and query time greatly and eliminate incorrect queries. An aggregate fact is derived from a base-level fact and measures in an aggregate fact is a computed summary of measures in the base-level fact. Dimensions in an aggregate fact can be derived from base-level dimensions and is called shrunken dimensions because the values is rolled up to create less fact rows e.g. Date dimension becomes a Month dimension, Product dimension becomes a Category dimension and an Address dimension becomes a Region dimension, and therefore the measures can be summarized to less fact rows for better query performance.
Year-to-Date ytd fact where month February is a summing up or roll up of January and February and so forth. Last-year-this-year fact with calculation of index compared to last year as a new column and easy to display in a report.
Aggregate fact table is simple numeric roll up of atomic fact table data built solely to accelerate query performance. It is called incremental aggregation when a ETL process do a dynamically update of a table by applying only new or changed data without the need to empty the table and rebuild aggregates.
A fact table used to combine fact data from multiple source legacy systems and multiple business processes together into a single consolidated fact table when they expressed at the same grain. E.g. sales actuals can be consolidated with sales forecasts in a single fact table to make the task of analyzing actuals versus forecasts simple and fast to compare how the year is going. A consolidated fact is a derived fact table that combine data from other facts.
A fact table contents several measures but only one or few of them has a value in each fact row. For the fact rows with same dimension member repeated in multiple contiguous rows with identical values, they will be smashed or collapsed into one fact row using operation as sum, min or max to limit the number of rows in the fact.
Time span fact
A fact table used for a source legacy system that is regularly updatable meaning that the source change and overwrite its values. To capture a continuous time span when the fact row was effective, the fact table will act as SCD type 2 dimension with BeginAt and EndAt columns to keep historical data and to represents the span of time when the fact row was the »current truth«. It is called Slowly Changing Facts. See more in section 6.4.
Counterpart fact (negating fact) and Transactional fact
A fact table used for Slowly Changing Facts because the source legacy system is changing fact value without keeping the old value as a historical transactions. See more in section 6.4.
Column wise fact and Row wise fact
A column wise pivoted fact table is useful to be columns in a report e.g.
Revenue Jan, Revenue Feb, Cost Jan, Cost Feb, Sales Jan, Sales Feb.
For a cube a row wise is much better because it gives good dimensions e.g.
Period, Entry, Amount.
Therefore a data warehouse needs to convert from columns to rows or vice versa.
A fact table contents huge number of rows where a period e.g. from a »date of employment« to a »termination date« as columns in one row, will be turned around to many rows with one row per day of the period, or per week or per month depends of the wanted grain of the period. When an employee has 10 year anniversary it will make more than 3650 rows on day grain, and if the employee dimension keep history of different names, departments and job functions etc. for each staff, then the fact rows will have different key references to the dimension. Another grain would be 15 minute grain which gives 96 rows per day for a windturbine power production. When the exploded fact is a source for a olap cube it can sometimes be implemented as a view in the database, and when it is for ad hoc reporting it will be used several times per day then it must be a materialized view stored in a table or sometimes as a indexed view.
3.2. Other fact classifications
Transaction has one row per transaction when they occur together with a datetime.
Periodic has one row for a group of transactions made over a period of time through summation like from daily grain to monthly grain so the Date dimension is represented by the month level with the first day of the each month. Notice that certain dimensions are not defined when compared to a transaction fact such as dimension TransactionType.
Accumulating/Aggregated has one row for the entire lifetime of an event and therefore constantly updated over time. For example, an application for a bank loan until it is accepted or rejected or a customer or working relationship. These fact tables are typically used for short-lived processes and not constant event-based processes, such as bank transactions.
Mostly a fact table describes what has happened over a period of time and is therefore an additive or cumulative facts. An example of a status column in a fact table that receive data from a school system where a student follow a course and later finish it, but sometimes a student skip the course and are delete in the system. Before reload the fact it can be a good idea to have a CourseStatus column with values like: Active, Completed or Dropped.
A dimension table containing business element contexts and the columns contain element descriptions and a dimension is referenced by multiple fact tables so the containing measurements make sense.
4.1. Purpose of a dimension
Some purposes as I seen it:
The dimension values can be placed in a hierarchy like a Location with levels Country→Region→City, and in a group or band interval like a Age with intervals of Child (0-12), Teeanage (13-19), Adult working (20-66) and Senior citizen (67-130). A dimension normally contains one or multiple hierarchies and/or groups to fulfill requirements from the users. E.g. a Product dimension has two hierarchies and one group: (a) Product→Product group, (b) Product→Brand→Brand group,
(c) Product→Price level.
A dimension can of course be non-hierarchical and non-grouping.
Different dimensionality covers the issue that not all combination of multiple dimension values are allowed in a fact and the data warehouse needs to make sure of the data quality.
4.2. Dimension keys
A dimension table has mininum three types of columns:
Primary key is a surrogate key identity column a unique sequence number to remove dependence from the source legacy system and for using in fact table foreign key. It can be called Entity key EK because it represent an entity from source or Surrogate key SK, but for me »surrogate« is a characteristic or a property not a name of a column. Primary key must not be data-bearing (in danish data bærende), it must be meaningless, but for a date dimension and a time dimension I like to use a smart valued primary key, e.g. a date 2013-12-31 as an integer value 20131231 and a time 08:30 am as an integer value 830 or 08:30 pm as 2030.
Business key is from a source legacy system and can be a primary key or another column that has unique values. In section 1.1 I used the terms natural key and business key and from a data warehouse point of view they are merged into one term business key. For example, Social Security number (SSN), StatusCode and CustomerNumber. The key value is not immutable but is meaningful for a human being that a business user prefer to use to identify a thing and as a search lookup value giving in a phone call to a company, a bank, a hospital or to the government. A business key has an embedded meaning and represent a unique object in the business. An alias is an Enterprise wide business key because the same value is used in multiple source legacy systems. If a business key is immutable and unchangeable in the source legacy system then call it a durable business key.
Surrogate key is from a source legacy system and is most often a primary key as an auto-generated unique sequence number or identity column (id, uid), that is immutable and meaningless for a human being.
Textual data is representing the dimension value context description and is saved in columns of the dimension table and will be shown to the users as descriptive columns to explain fact data.
Multiple source legacy systems have a business key for a Social security number per person with a primary key as a surrogate sequence number, therefore we can’t use the primary keys to interconnector (join, drill-across) the systems data, instead we use the business keys. The value of a business key can change over time e.g. a person obtains witness protection and gets a new Social security number, therefore it is up to the ETL process to make a solid mapping. When a business key is an employee number, and the employee resign, and some years later is rehired, there could be two employee numbers, and a data warehouse must find a way to glue or map them together, so the employee only occurs once in a dimension.
4.3. Changing dimensions
Source data is volatile data because they will change over time e.g. a customer change his name or address and a product change place in a hierarchical structure as a result of a reorganization. Therefore dimension values will normally be changed over time because of the volatility in source legacy systems. The rate of changes can be divided into two kinds of classifications and afterwards we will be looking into techniques to handle and tracking changes and to capture its history and preserve the life cycle of source data also called Change Data Capture CDC.
Slowly Changing Dimensions SCD
Columns of a dimension that would undergo changes over time. It depends on the business requirement whether particular column history of changes should be preserved in the data warehouse or data mart.
Rapidly Changing Dimensions RCD
A dimension column that changes frequently. If you do need to track the changes, using a standard Slowly Changing Dimensions technique can result in a huge inflation of the size of the dimension. One solution is to move the column to its own dimension with a separate foreign key in the fact table. Rapidly changing data usually indicate the presence of a business process that should be tracked as a separate dimension in a separate historical data table.
Techniques or methods to handle dimension values that is changing over time from the source legacy systems, called Ralph Kimball’s eight types or approaches of SCD:
Type 0: Original value, where the dimension value never change (method is passive) meaning keeping the original value from the source legacy system, and when the value is changed in source legacy system, we don’t change the value in the dimension. A single fixed value does not change over time but can be corrected in case of an error in a source legacy system, e.g. a correction of a misspelling of a name or a date of birth, date of launching something or date of first purchase. »Retain original«.
Type 1: Current value, where the old dimension value will be changed and forgotten when value is changed in source legacy system. The history of data values is lost forever. The active, actual, newest or latest of the value or most recent indicator. A fact table refers to a dimension value most recent, as-is. (In danish aktuelle, nuværende, gældende, seneste værdi). »Current by overwrite«. There is a one-to-one relationship between the business key and the surrogate key identity column as primary key of the dimension.
Type 2: Keep all values, where new dimension value is inserted into a new row to have a unlimited history of dimension values over time marked by timeline columns Effective date and Expiration date (Active date, Expired date, Expiry date) (or StartDate and StopDate or BeginDate and EndDate or ValidFrom and ValidTo), a pair of data type »date« or »datetime« that represents the span of time when a value was the »current truth«. A fact table refers to a dimension value in effect when fact data occurred, as-was, often by a date column in fact table based on source data or by a current load insert date when the fact data was entered into the fact table or was created, born or occurred. (In danish oprindelige værdi). The value of a business key will be repeated every time the textual data is changing, therefore the primary key is a surrogate key identity column a unique sequence number. A view upon the dimension will provide the current values. A view upon the fact will provide the current keys to join to dimension view. A column called IsCurrent has two values: 0 for historical and 1 for current to mark each data row of a type 2 dimension. This is the technique for Slowly Changing Dimension. »Keep history in rows«. There is a one-to-many relationship between the business key and the surrogate key identity column as primary key of the dimension.
Type 3: Keep the last value, where the previous dimension value is stored in a Previous column (or Historical column) and current dimension value stored in a Current column. An extended example of a type 3 dimension for a customer contains three columns for postal codes. The column names is Current Postal Code, Previous Postal Code and Oldest Postal Code. When a customer address has changed to a new postal code, the ETL process will move the value from the Previous column into the Oldest column and move the value from Current column into the Previous column and add the new value into Current column. This is the technique for Slowly Changing Dimension. »Keep history in column«.
Type 4: When a group of columns in a dimension is going to change often, we can split the dimension into a base dimension that has slowly changed columns and into one or more separate Mini Dimensions that has often changed columns (volatile data values) to keep the fast changing values in its own table. This is the technique for Rapidly Changing Dimension to store all historical changes in separate historical data tables. The base dimension can be either a type 1 or type 2 and the mini dimension becomes a type 2. A fact table contains foreign key to the base dimension and an extra foreign key to a mini dimension. We get a »capture a relationship in the fact«. »Keep history in tables«.
Type 5: Builds on the type 4 where a mini dimension gets a view to show the current rows of the mini dimension and the base dimension is extended with a key value that point to the current view and the key becomes a type 1 column outrigger in base. We get a »capture a relationship in the dimensions« because we only need to join base dimension and mini dimension without include the fact to save query performance. The join can be implemented in a view that is used in PIA. The ETL process must update/overwrite the type 1 column in base whenever the current mini-dimension changes over time. Therefore 4 + 1 = 5 type. »Outrigger current dimension«.
Type 6: Mixture of type 1 and type 2 columns therefore a good idea to suffix columns as _t1 and _t2 to know which columns can be overwritten in the current row. Can also have column of type 3, therefore 3 + 2 + 1 = 6 type. Type 6 act as type 2 of tracking changes by adding a new row for each new version but type 6 also updates _t1 columns on the previous row versions to reflect the current state of data by using the business key to join the new row with the previous rows. »Hybrid«.
Type 7: All rows follow type 2 to keep track of historical values with a key column and an extra key column called a durable key follow type 1 for the current value. The durable key is an integer representation of the business key. The fact table contains dual foreign keys for a given dimension (its key and its durable key) to show the historical value of the dimension at the time when the fact data was entered into the fact table or was created, born or occurred (in danish oprindelige værdi), and to show the current value of the dimension (in danish aktuelle, nuværende, gældende, seneste værdi). A view upon the dimension will provide the current values with the durable key to join to the fact durable key. »Dual Type 1 and Type 2 Dimensions«.
Correction of data together with the types
For type 0 I mentioned »correction of a misspelling« e.g. a first name of a customer Kelli is corrected to Kelly after few days in the operational source system, for type 2 or type 7 it will per automatic mean an extra row in the dimension, but I think it needs some consideration especially if source system can inform about the correction, maybe the dimension can do an update of the row as type 1 and mark the row with a housekeeping column/support column/administrative column IsCorrected.
Section 6.2 will show an example of type 1 and type 2. Section 6.3 will show an example of type 7 that has historical rows with a current mirror. Here is a type 6 example: In a dimension table the columns can be a mix of type 1 and type 2 e.g. a Customer dimension where columns for customer name, street name, house or apartment number is type 1 because we only need the recent value for shipment, and columns for postal code (zip code) and city is type 2 because we like to tracking these changes and keep data for the city where a customer was living in when purchase a product. A »sales in cities« report for the last ten years will use the right city of the customer at the time the purchase happened, occurred or took place.
4.4. Types of dimensions
Let us characterize the various dimensions into different types of dimensions.
Conformed dimension or Shared dimension or Common dimension
A conformed dimension has the same meaning to every fact in multiple data marts and measures will be categorized and described in the same way and ensuring consistent reporting across the data warehouse. A conformed dimension is a consistent interface to make sure that data can be combined in a data warehouse and be used all over the business because values of a dimension means the same thing in each fact. With a conformed dimension we can combine and drill across from fact to fact in one data mart or over several data marts, and analyze common columns and values. Seperate fact tables can be used together with shared, common and conformed dimensions. Conforming of several source legacy system data is part of the integration to achieve a conformed dimension where data is integrated of different meanings and different columns must be compared against each other, rules must be set, and data must be cleansed to create a single version of the entity. Conformed dimensions will unite and integrate data values among multiple source legacy systems so it is easy to search across different types of data and sync them in a common report. Shared dimension is utilized in multiple fact tables in a data mart or across multiple data marts. Dimension values comes from either the source legacy systems or is built by business rules in Usage supporting database. Non-conformed dimension can only be used within one fact. It is part of the ETL process to do conforming by merge, unite and consolidate multiple source legacy system data across the enterprise for making a conformed dimension e.g. a Customer dimension from different business areas as order, sale, invoice, delivery and support service (for B2B and B2C) with both different customers and same customers and with different business key values and with different addresses like a shipping address and a billing address. Sometimes data from a conformed dimension is send back to the source legacy systems as master data to be reusable in the organization.
The dimension values can be placed in a hierarchy like a Location with three levels Country→Region→City. A dimension can have several separate and independent hierarchies with different numbers of levels.
The dimension values can be placed in a group that is grouping values in band intervals like person ages in custom buckets like a Age group column with intervals of Child (0-9), Tween (10-12), Teeanage (13-19), Young adult (20-29), Adult (30-66) and Senior citizen (67-130).
Data classification is the process of organizing data into categories, group or category is part of a categorization or grouping of data to make a dimension more user-friendly to see data of a dimension on an aggregated and summed level.
A dimension normally contains one or multiple hierarchies and/or groups to fulfill requirements from the users. A dimension can of course be non-hierarchical and non-grouping.
Date dimension or Calendar dimension
A very common dimension with the granularity of a single day with hierarchies as:
Year→Half year→Quarter→Month→Date (five levels) or
Year→Week→Date (three levels) because a week does not always belong to one month.
When a fact date is connected to the Date dimension it is easy to make a report based on month, quarter or yearly level or a search filter criteria as Q3 2013.
Dates or days can be grouped to a Season or TimeOfYear: Winter (December January February), Spring (March April May), Summer (June July August) and Fall (September October November) (in danish Årstider).
Date dimension has normally many supporting columns like Weekday name and Month name and an IsHoliday column has a data type of boolean/bit with two values True (1) and False (0) as a flag indicator, but to be used in a presentation tool it is better to have a textual column DayOnOff with values »Holiday« and »Working day«. Of course the rule of when a specific date is a holiday is depending on country, collective agreement, company policy etc. is programmed into the stored procedure that builds the date dimension.
A fact table has often at least one column that represent a date e.g. an entry date, order data, invoice date, shipping date. When an event is happening it has normally a date or multiple dates like an injury with an InjuryDate, a ReceivedDate at insurance company, a RulingDate, a TreatmentDate, a BillingDate and a PaymentDate. Date dimension is a role-playing dimension and therefore the data mart contains multiple views upon the date dimension where each view can join to each date column in the fact table. Each view has a good name like the column in the fact table and the columns of the Date dimension is renamed to unique names for each view.
Using an integer value as surrogate key identity column in format yyyymmdd e.g. 20131224 for Christmas day, where the fact table will contain useful information that can be used in a query like:
»between 20131201 and 20131231« for month of December 2013.
Using integer key values to handle these four cases:
»Missing« (-1) when a source legacy system provide a date value that is null, meaning it is not present in some data, not registered, not reported, »hasn’t happened yet« or »be available later« because the date to be determined is expected to be available later and fact table will be updated thereby. Therefore the fact table column for the date dimension gets the value -1 for »Missing date«. When the date dimension is role-playing for a Birthday value -1 might stands for »No answer«.
»Not available« (-2) when a source legacy system provide a date value that is not known and not found in the data warehouse e.g. date 1234-05-06. Therefore the fact table column for the date dimension gets the value -2 for »Unknown date«. (In danish data værdien er ikke tilgængelig og ikke til rådighed i dimensionen).
»Not applicable« (-3) when the dimension is not relevant for the fact row (N/A, N.A.). Therefore the fact table column for the date dimension gets the value -3 for »No date«. (In danish ikke anvendelig, ikke relevant, vil ikke svare på spørgsmålet).
»Bad«, »Corrupt« or »Dirty« (-4) when a source legacy system provide bad date e.g. year of birth as 2099 or 2017-02-29 but year 2017 is not a leap year. Therefore the fact table column for the date dimension gets the value -4 for »Wrong date«. Please read more about it in section 4.5.
When a fact table gets many millions of rows it can be a good idea to use partitioning that is a behind technique to segment a fact table into smaller tables and in most cases it is based on the column of the date dimension. Therefore the date key is using data type date instead of an integer key to make the partitioning of a fact table works best, but that means we have to use specific dates for »Missing«, »Not available«, »Not applicable« and »Bad« like 0001-01-01, 1900-01-01, 1800-01-01, 1700-01-01. Let 9999-12-31 identify a future undetermined date.
Time dimension or Time-of-day dimension
A very common dimension with hierarchies such as:
Hour→Minute→Second (three levels) or
Hour→Minute (two levels)
The time dimension has the granularity of a single second with 86400 rows or of a single minute with 1440 rows.
Time dimension has an aggregation hierarchy that roll up of time periods into more summarized business-specific time grouping e.g. Time division or Time interval (»døgn inddeling«):
Morning (6 am to 8 am)
Rush hour (8 am to 11.30 am and 4 pm to 6 pm)
Lunch hour (11.30 am to 1 pm)
Afternoon hour (1 pm to 4 pm)
Dinner hour (6 pm to 8 pm)
Evening hour (8 pm to 11 pm)
Night hour (11 pm to 6 am)
I like to use an integer value as surrogate key identity column in format hhmmss e.g. 0 for 00:00:00, 63000 for 6:30:00 am and 202530 for 8:25:30 pm because 8 pm is also called 20 o’clock in the evening. When Time dimension has granularity of minute an integer value 10 stands for o'clock 00:10 and 2359 stands for 23:59.
A missing time is represented with the value -1 (»Missing time«), not available is -2 (»Unknown time«), not applicable is -3 (»No time«) and bad is -4 (»Wrong time«) if source provide a time as 24:12:06.
In a ETL process for making fact data rows, developed in a SSIS package, the column Time_key with data type smallint will be created as a derived column to the pipeline where null value becomes -1 and time 10:34:45 becomes 1034, expression:
ISNULL(TransactionDatetime) ? (DT_I2)-1 : (DT_I2)DATEPART("hh",TransactionDatetime) * 100 + DATEPART("mi",TransactionDatetime).
All dimension members or values (or branches in the hierarchy) have the same number of levels which makes the dimension symmetrical or balanced such as the Date dimension or fixed-depth hierarchy Country→Region→City where the three levels is in separate columns. Several columns in the dimension table is not at third normal form (3NF), therefore the dimension table contains redundant data or duplicate data by nature. A regular dimension has a flattened denormalized structure. A Product dimension table can have columns like ProductId, Code, Name and other descriptive columns, a Category because products are divided into categories and each category has a Target group. When a category belongs to multiple products, the target group will be repeated, meaning it does not satisfy third normal form (3NF), but it is okay for a regular dimension.
Example of a Product dimension with hierarchy Target group→Category→Product
There is three levels in the dimension hierarchy and it is balanced meaning that each Product has a Category and each Category has a Target group, so the dimension hierarchy maximum dimensionality is 3 (level of dimensionality).
For the Product dimension the product is the lowest level of granularity, and there is two many-to-one relationships because many products roll up to a single category and many categories roll up to a single target group.
A ragged dimension has leaf members (the last level of a hierarchy) that appears at different levels of the hierarchy and therefore contains branches with varying depths and number of levels. Europe→Denmark→Copenhagen with three levels and
North America→United States→California→Sacramento with four levels representing continent→country→state→capital. A ragged dimension is implement as a Parent-child structure or with a bridge.
To keep a hierarchy structure as a fixed level balanced regular dimension I can make a dummy level for a »not applicable« state of Denmark and get this:
Denmark and France is divided into regions, Philippines is divided into provinces and Germany has states/lands, therefore the name of level »State« could be changed to a broader word like »Division« which covers the different countries' areas.
Used to model flexible hierarchical structure where some dimension values have different levels of hierarchies called unbalanced or variable-depth hierarchy. Every value in the dimension have a related parent (mother) value, except the top value. The hierarchy is asymmetrical or unbalanced because values are placed in different levels within the same hierarchy. For example, an Employee dimension where the parent is the manager and the children is the employees under the manager, and some employees are both a manager and have another manager above, and the chain of command can be different from department to department. Another example is an Organization dimension where some departments have sub-departments and some teams have sub-teams, but there are also teams that don’t have sub-teams. This is the strongest side of Parent-child dimension to modeling.
Dimensional model dimensions is comply with second normal form (2NF) and snowflaking is a normalization of a dimension.
a) Splitting columns of a dimension table into smaller dimension tables with one-to-many relationships so data values fulfill and comply with 3NF and BCNF to avoid redundant data and denormalized structure. Snowflaking is normalization to 3NF.
b) Splitting a dimension hierarchy into two or more dimension tables is called »snowflake a hierarchy«. For example, a Customer dimension with the hierarchy:
Country→Region→City will be split into three dimension tables so only column City remains in the Customer dimension table that is connected to the Sales fact table, and two Snowflake dimension tables for Region and for Country with one-to-many from Country to Region and one-to-many from Region to Customer. Region and Country can be reused in other dimensions for Supplier and Store or in a FactSales to quick divide in regions and countries. Of course depending of business requirements specification.
c) Splitting a dimension with columns from multiple source legacy systems or when some columns will be user type-in from an application to enrich data and to avoid redundant data. See an example in section 6.1.
d) A Product dimension with a Category hierarchy is used in a Sales fact and Inventory fact. For forecasting (budgeting) the data is generated at Category level. Kimball will not do snowflaking of Product dimension instead he roll up to Category dimension as a strict subset of Product that the ETL process must take care of.
e) Splitting dimensions and move common columns to a general dimension is called an Outrigger dimension. For example, an Address dimension with street name, house or apartment number, postal code, city, county, area, region, country, gps coordinates and so forth to be used in other dimensions e.g. a Customer dimension with a shipping address and a billing address, a Building dimension with location address and an Employee dimension with home address. The Address dimension contains one row per unique address. A Sales fact can have a ShippingAddress_key and a BillingAddress_key as role-playing dimensions. Or a factless fact table with Customer_key, ShippingAddress_key, BillingAddress_key and effective and expiration dates for reporting track addresses of a customer. Kimball page 235. See more later.
f) A Shrunken dimension is not a snowflake dimension, see more later.
g) Splitting a dimension with rapidly changing columns is not snowflaking, see Mini.
h) Kimball’s financial services example starts with an Account dimension including data of products and branches but he choose to remove these descriptive columns to form independent dimensions of Product and Branch and not snowflake the Account instead add them to the fact because that’s the way business users think of them too. The fact will capture the relationship among data entities.
Some data modeler does not like Snowflake dimension because of the query performance with more tables in the join. I don’t use the terms snowflake schema and starflake schema because for me snowflake belongs to a dimension and sometimes it is great to have a denormalized structure and other times it is good to think of the above points. Most times a data mart will be a constallation schema (multi-star schema) where some dimensions is shared among different facts.
Now and then a data warehouse architecture has an extra layer after the data mart that is called presentation area where the ETL process or a view has joined and merged snowflaked dimensions together to only one dimension with all the columns with a lot of data redundancy, and that is fine since this layer will be refilled in every ETL process and it is easier for the users to do their query and for the tool to load data into its data model. Can use a materialized view. The layer is also called data consumption layer or data delivery layer for reporting, analysis and self-service BI.
Outrigger dimension or Reference dimension
When many columns belong logically together in a cluster or group it is fine to do a snowflaking to avoid a repeating large set of data and therefore making a dimension smaller and stable. Sometimes a canoe or a sailboat is using a rig to achieve balance because they are very narrow and a cluster of columns is placed in a Outrigger dimension or a reference dimension because it’s primary key will be a foreign key in the main dimension, but there is no reference found in any fact table. For example, a Product dimension has a column called LaunchDate that represent the date when the product will be applicable for the customers, and from that date the product can appear in the Sales fact table. The LaunchDate column is a foreign key to a Date dimension that becomes Outrigger dimension because Date dimension has many columns about dates, weeks, months, years and maybe fiscal columns too. Another example is demographic data for each country which is providing with 50 different columns. When we using outrigger dimensions we let each dimension has its own core columns.
Another example is from a bank with two kind of customers for Person and Company with common data as name and address and with specific data where a Person has social security number, gender and marital status and a Company has VAT identification number, industry code and turnover amount. A Customer dimension handle the connections to the fact tables and becomes a hub, an anchor or a party. Every time the bank gets a new customer, it will be set up in Customer dimension and the surrogate key value will be reused in either Person outrigger dimension or in Company outrigger dimension where both of them have a one-to-one relationship to the Customer dimension.
Shrunken dimension or Shrunken Rollup dimension
A Shrunken dimension is a subset of another dimension columns that apply to a higher level of summary of an aggregated fact and the shrunken dimension key will appear in the fact.
a) The Month dimension is a shrunken dimension of the Date dimension. The Month dimension would be connected to a forecast fact table whose grain is at the monthly level, while Date dimension is connected to the realized fact table.
b) A base-level Sales fact has a grain per date and product and is connected to a Date dimension with columns of date, month, year and a Product dimension with columns of names of product, brand and category. The Sales fact is derived to a aggregate fact with a grain per month and category and is connected to a shunken Month dimension with columns of month, year and is connected to a shunken Category dimension with a name column. The aggregate fact has columns for Month_key and Category_key and a summary amount of the Sales fact where the ETL process is using the Product dimension to roll up a product to a category and match the category to the Category dimension. Therefore both dimensions Product and Category has a category name column that becomes redundant data in the dimensional modeling. Therefore shrunken is not snowflaking because a snowflake dimension is on 3NF.
c) Shrunken roll up dimensions are required when constructing aggregate fact table. When a Sales fact has a daily grain the number of rows can become very large over time, therefore an aggregate fact summarized to monthly level will have less rows. Since the aggregate fact don’t need detailed customer information, the Customer dimension can make new shrunken dimensions for Region out of address, for Gender and for Age groups, and the summarized fact data become even less rows.
d) Sometimes users want few dimension values e.g. Red, Yellow and Green and they want Pink to become Red and Orange to become Yellow and so forth and the rest of the colours gets a residual value called Others. I will make a ColourGroup shrunken dimension with values: Red, Yellow, Green and Others. I will make a mapping table that will translate the colours e.g. Red to Red, Pink to Red, Yellow to Yellow and Orange to Yellow and the rest of the colours to Others. In the loading to the fact table I will let the colours pass by the mapping table and fetch the ColourGroup dimension to the fact table to obtain good performance for various statistics for the users.
e) Store data of an aggregated fact in a column of a dimension for easy searching e.g. a year-to-date SalesSegment in a customer dimension with descriptive values as Excellent, Super, Good, Average, Less, Nothing. The ETL process must ensure the column is accurate and up-to-date with the sales fact rows.
Dimensions that represent data at different levels of granularity to give higher performance. Can also refer to hierarchies inside a dimension with a higher grain.
Like a Month dimension that is derived from a Calendar dimension or we can say that Calendar has been reduced to Month, Year and Week with the start date of the week together with a week number and which year the week belongs to. A derived dimension can also be created by aggregating two existing dimensions.
In a hospital we can from a Patient dimension and an Employee dimension derive a Person dimension. A person can over time be both an employee and a patient or at the same time when the employee becomes sick and will be hospitalized.
Fact data can derive dimension data and it is called a degenerate dimension.
Previously, I showed a Date dimension and a Time dimension and with a combination of them I can create a new dimension to handle date and hour to be used in Power BI Impact Bubble Chart e.g. from 2016-08-22 10:00 to 2016-08-26 22:00.
A fact table or a view can have a derived column like DateHourInterval:
DateHourInterval = FORMAT(TransactionDatetime,'dd-MM-yyyy HH:00','en-US')
A view can make the data rows for the derived dimension:
CREATE VIEW DimDateHourInterval AS
SELECT DateHourInterval = FORMAT(CAST([Date] AS datetime) +
CAST([Time] AS datetime),'dd-MM-yyyy HH:00','en-US')
FROM DimDate CROSS JOIN DimTime
WHERE Minute = 0 AND [Date] BETWEEN '2010-01-01' AND '2029-12-31'
Mini dimension or Historical dimension
For Rapidly Changing Dimensions for managing high frequency and low cardinality changes in a dimension of fast changing volatile columns they are placed in a mini dimension or historical dimension with its own surrogate key identity column which will be included in the fact table. This approach is called type 4. A dimension table will be split into two tables, one with type 1 columns and the other with type 2 columns. An example is a customer with columns for Name, Gender, DateOfBirth, Address and Country→Region→City is placed in a Customer dimension table (can be type 1 or type 2), and the fast changing columns BodyWeightAtPurchaseTime and MonthlyIncome interval e.g. $ 0-10000, 10000-25000, 25000-50000, 50000-99999 is placed in a mini dimension table called CustomerBodyWeightIncome with its own surrogate key identity column and a foreign key back to the main dimension as a type 5. The sales fact table will have two columns to provide data for a customer, one key for Customer dimension and another key for CustomerBodyWeightIncome dimension. Sometimes it is necessary to have two or more mini dimensions if the columns is changing rapidly at different times. Normally there is no hierarchy in a mini dimension.
A bridge is to combine data that have a many-to-many relationship between a fact and a dimension or between a dimension and a dimension to take care of multivalued column at the conceptual level, and break it down at the logical level to a bridge dimension table with two one-to-many relationships and a composite primary key from the involved data. Instead of the word Bridge table a term Combine or Helper could be used. Peter Chen calls it an associative entity. A bridge can also be used for ragged hierarchies. (A song can have a Verse 1, a Chorus, a Verse 2, a Chorus, a Bridge and ends with the Chorus again.)
Multivalued dimension or Many-valued dimension
a) To handle when one fact row have two or more dimension values from same dimension. For example, a Sales fact can have up to four different sales staff employees and a sales staff has many sales. Therefore we say there is a many-to-many relationship between Sales fact and Employee dimension, and sales staff employee becomes multivalued in the Sales fact. It can be implemented by an extra table called SalesEmployeeBridge that contains the Id of a fact row and the Employee_key from the Employee dimension, or the fact table will get a dimension key column for a SalesEmployeeGroup dimension meaning a group of sales staff is connected to one fact row; and since there is a many-to-many relationship between a SalesEmployeeGroup and a Employee dimension, a SalesEmployeeGroupBridge table will express that by combining the keys from SalesEmployeeGroup and Employee.
b) When buying a paint mixture the different colors are mixed with a ratio or weight (sum up to 100%) the amount of paint. One sales fact row contains many colors and one color is included in many paint mixtures. Color becomes multivalued and a bridge gets a weight value to tell how much of that color is used.
c) To handle when one dimension row has two or more dimension values from another dimension. For example, an employee has a group of skills and one skill can belong to several employees. Therefore we say there is a many-to-many relationship between Employee dimension and Skill dimension, and skills of an employee becomes multivalued in the Employee dimension. It can be implemented by an extra table called EmployeeSkillGroupBridge that has a composite primary key of SkillGroup_key and Skill_key, and SkillGroup_key is a foreign key in Employee dimension and Skill_key is a foreign key in Skill dimension, so the many-to-many relationship comes two one-to-many relationships. The data experience level would be placed in the EmployeeSkillGroupBridge. Sometimes we create a fact table e.g. EmployeeSkillFact, but for me a skill is a description and a characteristics of an employee in a company and facts would be a Sales fact or a Production fact. In a School mart I would place the students' courses in a fact table because it is a result of study and passed an exam. Another example is a t-shirt can have up to three sizes »small«, »medium« and »large« and they can become three columns in the T-shirt dimension, or to make a Size dimension that has a many-to-many relationship to T-shirt dimension and the relationship becomes a T-shirtSizeBridge with QuantityInStock. Another example is a bank account can two bank customers like wife and husband, and of course each bank customer can have several accounts like budget, saving and pension. Therefore we say there is a many-to-many relationship between Customer dimension and Account dimension. We create a BankAccount fact table that will refer to the Account dimension, and the Account dimension refer to an AccountCustomerBridge table that again refer to the Customer dimension, so the BankAccount fact table will not refer directly to the Customer dimension. The AccountCustomerBridge table contains two columns Account_key and Customer_key as a composite primary key so an account can have several customers.
Role playing dimension or Repeated dimension
A role-playing dimension is repeated two or several times in same fact e.g. a Date dimension which key column is repeated in three foreign key columns in a fact table for three roles labeled SaleDate, ShipmentDate and DeliveryDate. For each role I create a view with distinguishable, unambiguously and unique names of columns as renaming of the columns from the Date dimension, and in this example it becomes three views upon the Date dimension called SaleDate, ShipmentDate and DeliveryDate with columns like Date of sale, Year of sale, Date of shipment, Year of shipment, Date of delivery and Year of delivery.
A City dimension can be repeated in multiple roles in a fact table of persons like these columns: BirthplaceCity, ChildhoodCity, ResidensCity, WorkingplaceCity, SeniorCity and DeathCity. It will become six views upon the City dimension.
A Manager dimension can be repeated as Sales clerk and Store manager.
Another example is a boolean dimension with key values 1/True and 0/False and in different views for several role-playing statuses the two values is translated to good texts as »Yes« and »No«, »In stock« and »Delivery made«, or »Available« and »Utilized« for a Utilization status.
When a dimension has an outrigger dimension e.g. Customer dimension has a column for FirstPurchageDate I create a view upon Date dimension.
Kimball says: »Create the illusion of independent date dimensions by using views or aliases and uniquely label the columns.« It will be easy for a user in Power BI or Tableau to drag into a data model several views for fact and dimensions without thinking of a dimension is playing multiple roles. In a olap cube data model the fact can be joined multiple times to the same dimension and at Dimension Usage can each role be labeled, but since we are using the same dimension the column names will be reused.
Junk dimension, Garbage dimension, Abstract or Hybrid dimension
A single table with a combination of different and unrelated columns to avoid having a large number of foreign keys in the fact table and therefore have decreased the number of dimensions (dimensionality) in the fact table. Kimball recommand up to 25 dimensions in a fact. The content in the junk dimension table is the combination of all possible values of the individual columns called the cartesian product. For example, four different values that can be cross-joined into a Junk dimension:
Payment method: Cash or Credit card.
Coupon used: Yes or No or Not applicable.
Bag type: Fabric or Paper or Plastic or Unspecified.
Customer feedback: Good, Bad or None.
with 2 x 3 x 4 x 3 = 72 values or rows in a Junk dimension table, but can contain only the combination of values that actually occur in the source data.
The fact table only needs one key to the Junk dimension for getting the descriptive values of fact data for reporting. The pitfall of a Junk dimension is the filtering because a value (e.g. Credit card) exists as duplicate in multiple rows and therefore gives multiple key values to be joined to the fact table. To display unique content of a column from a Junk dimension in a dropdown or listbox I need to create a view for that column e.g. create view [Dim Bag type] as select distinct [Bag type] from [Dim Junk]. The view will handle a one-to-many relationship to the Junk dimension in the same way we handle a snowflake dimension. I can in Power BI create a calculated table upon the Junk dimension with a DAX like:
Dim Bag type = distinct(Dim Junk[Bag type]) and I build a relationship from the calculated table Dim Bag type back to the Junk dimension in the model where Junk dimension already has a one-to-many relationship back to the fact. I hide the Dim Junk because a user does not need it after we have calculated tables for each of the columns in the Junk dimension and therefore the Junk dimension becomes a bridge or a helper. I hope it shows how to use a Junk dimension in practice.
Degenerate dimension values exist in the fact table, but they are not foreign keys, and they do not join to a real dimension table. When the dimension value is stored as part of fact table, and is not in a separate dimension table, it is typically used for lowest grain or high cardinality dimensions such as voucher number, transaction number, order number, invoice number or ticket number. These are essentially dimension key for which there are no other columns, so a degenerate dimension is a dimension without columns or hierarchy. For example, the OrderNumber column can be in the Order fact table with several rows using the same order number, because one order can contain several products. Therefore the OrderNumber column is important to group together all the products in one order. Later for searching for an order number in a OLAP cube, a Order number dimension is very usefull, but it is not a dimension table, it is generated by the Order fact table and there is no additional data like a name or text.
Degenerate dimension is also used as an implemented Bridge dimension table in making a Multivalued dimension. In a data mart these are often used as the result of a drill through query to analyze the source of an aggregated number in a report. You can use these values to trace back to transactions in the OLTP system.
Normally a degenerate dimension is not a table in a database, it is a view with distinct values based on the fact table.
Dimensionalised dimension is a replacement of Degenerate dimension where the view becomes a materialized view meaning it becomes a table in the data mart and where the text column will be transformed to an integer key, like this:
SELECT QuantityPerUnit_key = CONVERT(BIGINT, HASHBYTES('MD5', t.QuantityPerUnit)),
FROM (SELECT DISTINCT QuantityPerUnit
FROM Products WHERE QuantityPerUnit IS NOT NULL) t
Static dimension or Constant dimension
Static dimensions are not extracted from the source legacy system, but are created within the context of the data warehouse or data mart. A static dimension can be loaded manually with Status codes or it can be generated by a procedure such as a Date dimension and Time dimension. The opposite would be called Dynamic dimension.
Several different kinds of entry with different columns for each fact (like sub classes). For example, heterogeneous products have separate unique columns and it is therefore not possible to make a single product table to handle these heterogeneous products.
Measurement type dimension or Entry type dimension or Fact dimension
Used to identify different facts that is populated in the same measure column in a fact table because the fact rows represent different entry typies. An measurement type dimension describes what the fact row represents and how measures must be understand and used. The alternative is to have multiple measure columns for each entry type in the fact table where only one column has a value for each row.
Time machine dimension
A combination of two dimensions for Entry date (or event date) and for Year and where fact data rows is based on either »counterpart« as in a financial accounting or »transactions« as in a financial transactions. Section 6.4 will show an example.
A very large dimension that has a huge number of rows or many columns. For a real estate agent, I implemented a 132-columns dimension through a merge of five source legacy systems. The column names was made together with the users. The created dimension table got the column names in an alphabetical order so it is easy to find a specific column.
Dimensional columns are allowed to become complex objects rather than simple text like unstructured text, gps tracking, graphic images and in time series and in NoSQL bases like Hadoop and MongoDB, and become much more malleable and extensible from one analysis to another. Extensible design in software engineering is to accept that not everything can be designed in advance and extensibility is a software design principle defined as a system’s ability to have new functionality extended.
Audit dimension or Data quality dimension
A description of each fact table row would be »Normal value«, »Out-of-bounds value«, »Unlikely value«, »Verified value«, »Unverified value« and »Uncertain value«. Kimball recommands to load all rows to the fact table and use an Audit dimension to do a tagging of data because of an error condition and thereby to tell the state of each row in an audit report so a user can look at the data and make a data change or a counterpart in source legacy system and do a new load to the data warehouse to fix the data in the fact table. Data quality is monitored during the ETL process and it can procedure an audit statistics.
In a SSIS package a derived column can have an expression to validate data and give value to an Audit dimension where an amount from the source is less than 0 or more than 5000 gets audit key value 12 for out of bounds amount else value 0 for okay: Amount < 0 || Amount > 5000 ? (DT_I4)12 : (DT_I4)0
All rows are checked for compliance with the constraints.
An audit dimension can also have data for name and version of source legacy system, name of data table in source, time of extract from source, time of insert into the fact table etc.
Textual data as a numerical fact measurement
It is a matter of confusion whether a numerical value should belong to a dimension or a fact. For example, a price of a product could be a column of the product dimension, but since the price of a product often varies over time and maybe also over location, it should be a non-additive measure column in a fact together with a Date dimension.
Another example, a national lottery has a fixed pool size amount for each weekly drawing and since the amount is fixed and thus does not change over time, the pool size amount belongs as a column of the draw dimension together with a DrawDate column that becomes a outrigger or reference dimension which provide a week number and a year. The pool size amount can be used for filtering (like to see pool size above $1 billion) and for grouping value band. When a person buys a lottery coupon, the deposit and the date of purchase will be recorded in a fact, and all the deposit amounts for a week can be summarized to a weekly deposit kpi e.g. in a view or in a tabular cube. A Power BI chart can combine the fact weekly deposit with the dimension weekly pool size and it can be seen that the deposit is significantly larger in the weeks that have very big pools.
4.5. Inferred members or Inferred dimensions
A dimension has a member value of »Missing«, »Unknown«, »N/A«, »-« etc. to handle source data that is going to be loaded into a fact table:
Approach: Handling a null/empty business key as a missing member because a row in the fact table for column of business key is not registered.
Approach: »Early arriving fact« for handling of orphaned data where the fact data has a unknown member, meaning the fact value is an orphan child because there is no parent value in the corresponding dimension table, the fact value does not exist in the dimension.
In a relational database it is called a referential integrity constraint violation in a table when a foreign key contains a value that does not exists as a primary key in a different (or the same) table. There is a no referred member and there is a need to have an inferred member. A forthcoming fact row has a member that will infer a new dimension member, therefore it is called inferred member of the dimension (in danish udledt), and it is to improve data quality and audit trail/control track for reconciliation in the data warehouse.
Handling a null/empty business key as a missing member
When a dataset is transformed to a fact table, there can be a business key column which value is empty or null, meaning it does not yet exist for the fact data. To keep all the fact data rows in the fact table, the related dimension table is already born with a »Missing« value (an inferred member) with surrogate key identity value -1, which can be used as a default value in the fact table foreign key column to the dimension table primary key column. Later the fact table with a -1 value can be updated with a real business key value that either exists in dimension table or will be inserted first and get a new surrogate key identity that the fact data row can refer to. Sometimes I have seen a dimension with an inferred member value of -1 for »Unknown«, but I prefer using -1 for »Missing«, and I am using -2 for »Unknown« to handle the situation of early arriving fact in general way.
Early arriving fact as a unknown member
When a dataset is transformed to a fact table, there can be a business key column which value has not yet been received to the related dimension table and therefore does not yet exist. To keep all the fact data rows in the fact table, the related dimension table will first have to insert the new business key value with a »Unknown« value (called an inferred member) which later will be updated in the dimension with the correct text value then it is known. The »Unknown« value gets the next surrogate key identity as a unique sequence number and will be used in fact table like any other dimension member value. A dimension table can at the same time have several »Unknown« member values with their own business key, surrogate key identity and the text value can include the value of the business key like »Unknown 886«, »Unknown 887«, »Unknown 888« and so on to distinct them for the users of the dimension. When a regular dimension has a hierarchy these levels can have text value »Unknown« as a special branch in the hierarchy.
Late arriving dimension
When a dimension data has been delayed:
a) If business key not exists then insert a new row.
b) If business key exists as an inferred member »Unknown« for type 1 update the row by overwriting its values, and for type 2 by inserting a new row to keep history.
c) If business key exists and the delayed data comes with a date of valid then type 2 becomes more complex because validfrom and validto has to be adjusted and fact rows has to be revisited to update the dimension key column to point at the right dimension value at that date. You have to consider if it is allowed to change old data and old reporting result.
Late arriving fact
When a fact data has been delayed maybe it is including a date that can be used to search and fetch the current dimension member at that time if dimension keeps history. You have to consider if it is allowed to add a late arriving fact row because it will change an old report. For example, a boss already got the report of sales for the first quarter and at June 5 a late sale fact for March 31 is arriving and when it is added to the Sales fact, the report for first quarter will change so it does not match the old reporting result.
When building the ETL process for a fact that is using type 2 or 7 dimensions sometimes we can assume that new fact data rows is at current date and therefore we only need to do a lookup for the dimension key value with this criteria:
dim.Businesskey = fact.Businesskey AND dim.IsCurrent = 1
But if a fact data row can be late arriving with an old date stamp in column TransactionDate we need to do a date range lookup for the dimension key value with a criteria to found the fact business key in the dimension at the time when fact date was valid in the dimension:
dim.Businesskey = fact.Businesskey AND
dim.EffectiveDate <= fact.TransactionDate AND dim.ExpirationDate > fact. TransactionDate
Range lookup is an in-memory lookup to assign a surrogate key to each coming fact row in a streaming ETL process of rows from a source legacy system to translate and replace a business key value with a dimension surrogate key value to be saved into a fact data row in a fact table.
Other terms for TransactionDate could be EventDate, HappeningDate or OccurredDate in the business sense e.g. OrderDate, PurchaseDate, SalesDate, InsuranceCoverageStartDate, PlanEffectiveDate or ValueDate etc. and of course a time (o'clock) can also be added.
Date columns in a fact table to make sure of match the old reporting result:
· TransactionDate when something occured in the business sense.
· RegisteredDate when it was known in the operational system.
· InsertTime when it was known in the data warehouse fact table.
When the data warehouse receives a correction to a fact row or a fact row arrives late these dates can be helpful. When the boss got the report of sales for the first quarter at April 2. We can mark the report with an ExecuteDatetime 2010-04-02 4pm to be used when the boss at August 16 wants to recreate an old report as it actually was looked at the time of creation, we can search the fact data:
InsertTime <= ExecuteDatetime.
And we can make a new report as it should have looked given corrections made to the data after its creation: InsertTime <= Now.
See more in chapter 2 temporal snapshots and in section 6.4.
Early arriving dimension
When a dimension has members that is not yet been referred to from a fact row. There will never be a method of payment called »Blood«, but I still keep it as a member of the dimension.
Other inferred members
Variety of constant inferred members and fixed key values for a dimension could be:
»Missing« (-1) when a source legacy system provide a business key value that is null, meaning it is not present in some data, not registered, not reported or »hasn’t happened yet« or »be available later« because the business key to be determined is expected to be available later and fact table will be updated thereby. Therefore the fact table column for the dimension gets the value -1 for missing.
»Not available« (-2) when a source legacy system provide a business key value that is not known and not found in the data warehouse e.g. CustomerNumber 424-15-90. Therefore the fact table column for the dimension gets the value -2 for unknown in a general term for a unknown member because it is not all values I like to make as an inferred member for a dimension. A CustomerNumber in a source system could be type in wrongly and therefore is it not available for a dimension in the data warehouse, and I find it better to mark the dimension in the fact with -2 instead of inserting a wrong value into the Customer dimension. Some wrong CustomerNumbers will in the fact table gets the value -2 and therefore become homeless, but an Audit trail can fetch them and reporting them and the source system can be corrected. In a real time load it could happen that a sales transaction is referring to a CustomerNumber that is not yet in the Customer dimension and inferred member is the only way out if it is not possible to use a snapshop on the source legacy system to create a read-only and static view as a transactionally consistent with the database as of the moment of the snapshot's creation. (In danish data værdien er ikke tilgængelig og ikke til rådighed i dimensionen).
»Not applicable« (-3) when the dimension is not relevant for the fact row (N/A, N.A.). It is best to avoid a N/A dimension by making separate fact tables. When a type 2 dimension is expanded with new columns the old historical rows do not have a value for the columns, we can use »Not applicable« as the value or a new member »Before column« (-5). The same apply to a fact table expanded with a new dimension column, and back in time we have no value for that dimension. (In danish ikke anvendelig, ikke relevant, vil ikke svare på spørgsmålet).
»Bad«, »Corrupt« or »Dirty« (-4) when a source legacy system provide bad business key value or not enough data to determine the appropriate dimension key identity. This may be due to corrupted data in the source legacy system or incomplete knowledge of the business rules for the source data for the dimension.
The term »Not applicable« I will normally replace with a relevant term for a dimension e.g. a FactClassInstructor has two dimensions for a primary instructor and a secondary instructor and a role-playing DimInstructor. Sometimes a class has no secondary instructor therefore I use dimension key value -3 in the fact row with the term »No instructor« in DimInstructor.
The key values -1 to -4 is not carved in stone, so please make the values that fit the dimension, for example a FactQuestionnaire has a DimRespondent with following key values:
A dimension table can have a metadata column like IsInferred as a boolean (bit) with two values (1 = true and 0 = false):
An example of a Customer dimension
Will an inferred dimension member value data row be updated when the source legacy system provide a Region and a Name?
Kimball page 479 and 516 says it is okay to do a type 1 overwrite in a type 2 dimension and overwrite IsInferred with 0. I say it is up to you to think about the overwrite, because will it be okay after a week or a month that a report back in time will have a different output in present time?
The fact table is untouched or unchanged because the foreign key Customer_key values 134 - 136 are at the right place already, it is only Unknown values of the dimension that is changed. Kimball says there is no need for revisiting the fact table for making an inferred member to a normal member of the dimension.
When other levels in a hierarchy has text value »Unknown« like Region and users of the dimension in a cube has already made reports with fact data connected to »Unknown«, it is important to keep the history in the dimension so a user some months later still can make the same report with same summary value »Unknown«. If a user instead want to use the current values of the previous inferred members, the cube should provide such a dimension too.
The design of a dimensional model has a four-step process described by Kimball:
1. Select the business process to model by gathering and understanding business needs and available data in source legacy systems.
2. Declare the grain of fact as Transaction, Periodic or Accumulating/Aggregated.
3. Identify the dimensions with columns, hierarchy and group.
4. Identify the facts with measures as additive or non-additive.
Dimensional modeling ends up in relational database tables for dimensions and facts with one-to-many relationships between primary keys and foreign keys, meaning it becomes a Entity Relationship data model where entities are labeled with Kimball naming.
An entity or table can have a normal name e.g. Product and Sales and they can be placed into multiple database schemas based on Kimball naming to labeling each table in the dimensional modeling e.g. of some database schema names:
Extra schemas for Usage supporting tables in the ETL process:
Sometimes a data modeler is adding a label Dim and Fact to the name of the table e.g. DimProduct and FactSales or Product_dim and Sales_fact.
The most important of all is a good table name e.g. in Kimball’s Healthcare example he has tables like:
An invoice has multiple line items. Why make invoice as a dimension, when the parent invoice header fact table has 4 dimensions and 5 measures, and the child invoice line item fact table has 6 dimensions and 3 measures? Kimball Design Tip #25: We can't roll up our business by product! If we constrain by a specific product, we don't know what to do with invoice level discounts, freight charges and tax. Have to take the invoice level data and allocate down to the line item level. Therefore we get one Invoice fact table with dimensions and measures from the invoice header e.g. invoice no, order no, invoice date, customer, ship-to and sales person, and from the invoice line item e.g. product, quantity, unit price and amount. Invoice number becomes a degenerate dimension. We get a »capture a relationship in the fact«.
Utilizing dimensional modeling, end users easily understand and navigate data structure and fully exploit the data.
5. More readings
https://www.kimballgroup.com/2002/09/two-powerful-ideas/ (data staging area)
Steps to Modeling the EDW (video)
Ralph Kimball and Margy Ross: The Data Warehouse Toolkit, The Definitive Guide to Dimensional Modeling, Third Edition, 2013, first edition published in 1996 to introducing the star schema as the basic solution for modeling multidimensional data on a DW.
Ralph Kimball and Joe Caserta: The Data Warehouse ETL Toolkit, Practical Techniques for Extracting, Cleaning, Conforming, and Delivering Data, 2004.
Page 215 shows a pipeline for a ETL process to lookup key values in dimensions. Read how to implement a pipeline with date range lookup for SSIS in section 11.4
William H. Inmon: Building the Data Warehouse, Fourth Edition, 2005, first edition published in 1990 to be considered as the birth of DW as »a subject-oriented, integrated, time-variant, and non-volatile collection of data in support of managements decision making process.«
Dan Linstedt and Michael Olschimke: Building a Scalable Data Warehouse with Data Vault 2.0, 2015. Introduction and good debate
6. Examples with sample values
6.1. Customer snowflake dimension
Customer table in a source legacy system with CustomerId as a primary key (selected as business key for the data warehouse because the value is immutable) and CustomerNumber build of the nationality (country) and the gender of the customer (the value is not immutable because customer can change nationality and gender):
Customer dimension table as a snowflake dimension:
Customer_key is a unique sequence number as primary key for the dimension for having independence on the business key CustomerId from the source legacy system. From CustomerNumber I derived two extra columns for Country and Gender and they become snowflake dimensions and can be used together with other dimension tables and fact tables. Payment is taking from another table in the CRM system and becomes a dimension itself and is merged into the Customer dimension to do an enrichment of the customers. Country dimension has been enriching inside the data warehouse by the business users with ISOcode and Currency values to fulfill reporting requirements. Each country is placed in a continent hierarchy (Africa, Antarctica, Asia, Australia, Europe, North America and South America) in case a report wants a continent overview of the customers and sales.
Country dimension table becomes an Outrigger dimension:
Gender dimension table:
Payment dimension table with an extra data warehouse column to specify the sort order of payments for reporting purpose and a hierarchy of types:
For a olap multidimensional cube loading data to a Customer dimension will be implemented with a view that joins the relevant tables and with several columns and hierarchies to make an easy search, zoom and drill down to specific data in access data tools like Excel, Power BI, QlikView, Tableau or Targit:
· Nationality (Country renamed to play a role)
· Payment method
In terms of granularity I want to see daily sales by customer and product for each individual employee which gives the grain declaration of date, customer, product, employee. The time-of-day from the operational system is left out so quantity has been summarized to date level in case a customer purchase the same product at the same employee in the one day e.g. 10.15am and 3.45pm.
An example of a sales statistic would be: »What is the amount of sale to female customers from France that have been paid with credit card in April 2014?«
We need a Sales fact table like this:
The last row shows a unwanted sale because it is a stolen item therefore the customer is missing together with the amount.
A report want to show the total amount per gender including Transgender even though there is no customer yet. Since snowflaking has created a Gender dimension with all genders independent of customer, a query will use an outer join to include all genders:
SELECT g.Gender, TotalAmount = ISNULL(SUM(s.Amount),0)
FROM Fact.Sales s
INNER JOIN Dim.Customer c ON c.Customer_Key = s.Customer_key
RIGHT OUTER JOIN Dim.Gender g ON g.Gender_Key = c.Gender_key
GROUP BY g.Gender_key
When snowflaking is not wanted for the Customer dimension, it will contain all the relevant descriptive columns:
The Customer dimension will contain redundant data e.g. a PaymentSortorder value is repeated for all the customers that is using same payment. One day business ask for a new sort order of payments because it will be more suitable for the reporting. I recommend to have a Usage supporting database in the data warehouse with tables where business users can update data through an application. With a Payment table with columns Payment, PaymentSortorder and PaymentType a business user can change the values in column PaymentSortorder. The Usage supporting database will be a data source to the data warehouse where the ETL process will fetch the Payment table to the Input data area and the ETL process will update the Customer dimension table in the Data mart. Since the business does not care of old sort order of payment, the column PaymentSortorder in Customer dimension will be treated as a type 1 column meaning a simple update of all the rows in the dimension table.
In the Usage supporting database a business user can add extra data to the countries and hereby enrich the reporting with data that don't exists in the source legacy systems.
Since there is no customer of Transgender, the value will not exists in the Customer dimension and therefore Transgender can’t been shown in a report except when it is added through a Union sql statement but data belongs in a row in a table.
When all dimension values is wanted as fact, the fact table needs to include extra dimensions for Gender, Country and Payment and drop them from Customer. We get a »capture a relationship in the fact«.
6.2. Type 1, type 2 and type 7 with most recent current value
A customer has Id 421 in a source legacy system and her maiden name is Marie Beaulieur, and she is inserted into a Customer dimension. At 2007-11-16 she got married and took her husband’s surname, therefore her name was changed to Marie Lesauvage.
When the dimension is a type 1, Marie’s maiden name will be overwritten and will be forgotten. Customer_key is a surrogate key identity column a unique sequence number.
The type 1 dimension will always display the customers current name (most recent name). A Customer_key column will be in a fact table and in a data access tool like Tableau and QlikView fact column Customer_key will be joined to dimension column Customer_key inside the tool’s data model.
When the dimension is a type 2, Marie’s maiden name will remain and will be remembered and her changed name will continue in an extra row. To keep track of when a value is changed, there is two datetime timeline columns called ValidFrom and ValidTo. Marie’s maiden name was valid from she was born or she becomes customer at 2000-05-08 and that date is known in the source legacy system. The first entering will be giving a ValidFrom e.g. 1900-01-01 as a beginning of time and a ValidTo e.g. 9999-12-31 as a end of time. Marie’s marriage at 2007-11-16 will create a new row in the dimension table with the ValidFrom 2007-11-16 because she took her husband’s surname and the previous row will get ValidTo 2007-11-16 meaning that until that date she kept her maiden name. Customer_key is a surrogate key identity column a unique sequence number.
Note: Kimball page 508: If ValidFrom and ValidTo are datetime stamps then the ValidTo must be set to exactly the ValidFrom of the next row so that no gap exists between rows and no overlap too with continuous time. Other kind of time is event time (in a point of time) and interval time with gaps or overlap.
ValidFrom represents »from now on« (in danish »fra og med«) and ValidTo represents »to and not included« (in danish »til og ikke medtaget«).
What happen to the date when Marie becomes a customer at 2000-05-08? We can add a column called CustomerBeginDate, or we can add two rows at the first time a customer is entering so ValidFrom in the second row will represent the date the person becomes a customer. It will look like this:
The above table is not used in the following example.
Marie Lesauvage remarried at 2014-07-15 and took her new husband’s surname therefore her name is changed to Marie Sainte and becomes a new row in the table.
The Business_key column is the »glue« that holds the multiple records together.
Columns ValidFrom and ValidTo is metadata information about the data row itself. A dimension can have extra columns for metadata for a data row in a table e.g.
IsCurrent, IsDelete, IsInferred, State column or Status column with values e.g. »Current, Deleted, Expired«. More about metadata columns in section 6.3.
Marie has made six purchases which is entered into a fact table and the date of purchase determines the value in column Customer_key after this rule.
Business_key = CustomerId AND
ValidFrom <= PurchaseDate AND ValidTo > PurchaseDate
It is called Date Range Lookup when we are going to find the right Customer_key in the Customer dimension for a specific date of purchase for example, 2014-05-08 for Marie becomes:
Business_key = 421 AND
ValidFrom <= 2014-05-08 AND ValidTo > 2014-05-08
The dimension row with ValidFrom 2007-11-16 and ValidTo 2014-07-15 meets the rule:
Business_key = 421 AND
2007-11-16 <= 2014-05-08 AND 2014-07-15 > 2014-05-08
The dimension data row has Customer_Key value 4 which we are using in the inserted fact row that represent her purchase at 2014-05-08 where her surname was Lesauvage at that time. The Date Range Lookup will be performed for each row of purchase from the source legacy system where each PurchageDate must fit the rule. An example of a fact table where Marie has made many purchages since year 2005 and each purchage has its own Customer_Key value as a navigation reference back to the Customer dimension to fetch full name when the purchage was happening, occurred and registered:
For each row in the fact table, the value in column Customer_key is pointing to a row in the dimension table and give the name of the customer at the time the purchase was happening, occurred and registered. Customer_key column in fact table represent as-was when it is joined to the dimension table.
When a report of purchases wants to show the customers current name (most recent name) independent of the date of purchase, it is solved with a database view that find the most recent CustomerName for each value of Customer_key with this result:
A type 2 current view dimension can be implemented like this:
CREATE VIEW DimCustomer_Current AS
SELECT d.Customer_key, c.CustomerName
FROM DimCustomer d
(SELECT Business_key, CustomerName -- the most recent CustomerName.
WHERE ValidTo = '9999-12-31' -- datetime2(7) '9999-12-31 23:59:59.9999999'
) c ON c.Business_key = d.Business_key
The view will do a query for always to select the current value for each key values in the dimension and show the result set or recordset as showned in the table above.
If your sql has windowing function the view can avoid the self join which will enhance performance and here is no use of ValidTo column:
CREATE VIEW DimCustomer_Current AS
SELECT Customer_key, CustomerName = LAST_VALUE(CustomerName)
OVER(PARTITION BY Business_key ORDER BY ValidFrom
ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)
Use the view in a join with the fact table to fetch the current value for each fact row:
SELECT c.CustomerName, f.Quantity, f.Amount
FROM FactSales f
INNER JOIN DimCustomer_Current c ON c.Customer_key = f.Customer_key
Customer_key column from the view represent as-is for most recent data when it is joined to the fact table.
The disadvance of the first view for current value of the dimension is the join on the business key because it can be a string representation or a composite business key is composed of multiple columns that needs to be part of the inner join columns, which can cost performance when the dimension has many rows. With type 7 dimension Kimball introducing an integer representation of the business key and he called it durable key, an artificial surrogate key value per business key, see later.
The disadvance of the second view is that the windowing function must be used for each column that is wanted to be displayed for the dimension.
Temporal tables in SQL Server 2016 is closest to type 4 because keeping history in the separate table while original dimension table keeps current dimension member, more reading.
In case you want to display the original value of the dimension as type 0:
CREATE VIEW DimCustomer_Original AS
SELECT Customer_key, CustomerName = FIRST_VALUE(CustomerName)
OVER(PARTITION BY Business_key ORDER BY ValidFrom
ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)
Use the dimension table in a join with the fact table to fetch registered value for each fact row:
SELECT c.CustomerName, f.PurchaseDate, f.Quantity, f.Amount
FROM FactSales f
INNER JOIN DimCustomer c ON c.Customer_key = f.Customer_key
Customer_key column from the table of dimension represent as-was when the fact row was entered into the fact table when the data occurred.
There is another type 2 approach that does not need the ValidTo column in the dimension table and therefore the ETL process shall not update the previous row, because a ValidTo column can be calculated when it is needed to determine the dimension key for the coming fact row:
CREATE VIEW DimCustomer_ValidTo AS
SELECT Customer_key, CustomerName, ValidFrom,
ValidTo = LEAD(ValidFrom, 1, '9999-12-31')
OVER(PARTITION BY Business_key ORDER BY ValidFrom)
In mysql 5.7:
CREATE VIEW DimCustomer_ValidTo AS
SELECT Customer_key, CustomerName, ValidFrom,
(SELECT IFNULL(MIN(b.ValidFrom), '9999-12-31 00:00:00') AS ValidFrom
FROM DimCustomer b
WHERE b.Business_key = a.Business_key AND b.Customer_key > a.Customer_key
) AS ValidTo
FROM DimCustomer a
In a data access tool like Tableau and QlikView the type 2 for showing most recent values needs two views that will inside the tool’s data model be joined together. A user can in a dropdown box select Marie Sainte (meaning Customer_key value 5) and all her six purchases facts will be shown or a sum of Amount even though she has changed her name over time. The dimension for current customers will show this unique list of most recent customers in a dropdown box based on a view:
CREATE VIEW dma.DimCustomer_Current AS
SELECT Customer_key AS Customer_Current_key, CustomerName
WHERE ValidTo = '9999-12-31' -- datetime2(7) '9999-12-31 23:59:59.9999999'
The fact will show this through a view that include a Customer_Current_key column to be joined to the view DimCustomer_Current in the tool’s data model:
Column Customer_Current_key represents a master value for the different slave values in column Customer_key, e.g. master value 5 has slave values 3, 4, 5 in a master slave relationship.
CREATE VIEW dma.FactSales AS
WITH dimCustomer_Current (Customer_key, Customer_Current_key) AS
MAX(Customer_key) OVER(PARTITION BY Business_key) AS Customer_Current_key
SELECT f.PurchaseDate, f.Customer_key, c.Customer_Current_key, f.Product_key,
FROM FactSales f
INNER JOIN dimCustomer_Current c ON c.Customer_key = f.Customer_key
When current customer Marie Sainte is selected in a dropdown box, the tool will do a query based on the tool’s data model to show all her six purchases facts even though she has changed her name over time:
SELECT c.CustomerName, f.PurchaseDate, f.Quantity, f.Amount
FROM dma.FactSales f
INNER JOIN dma.DimCustomer_Current c
ON c.Customer_Current_key = f.Customer_Current_key
WHERE c.CustomerName = 'Marie Sainte'
The fact view is still keeping the Customer_key column to be joined to the Customer dimension for showing the name of a customer when the purchase was happening.
The view dma.FactSales is using a With part (Common Table Expression) to make a temporary result set. It could be changed to a materialized view which store data in a table with a clustered primary key through the ETL process, and table FactSales will join to the materialized-view-table to get a better performance.
When a dimension is a type 7 it keeps history as type 2 and it does the same as type 1 current value with an extra column called Durable key that is an integer representation of the business key because a business key could be a mix of number and text and could be a composite key (durable in danish is holdbar or slidstærkt).
Customer_key is holding the type 2 complete history of the dimension columns changes, and Customer_Durable_key is holding the type 1 current version of the dimension columns.
Both the dimension primary key (here Customer_key) and the durable key (here Customer_Durable_key) will appear in the fact table as a dual foreign key for a given dimension.
The Durable key will also appear in the current view that will show each customer in a unique list of most recent customers.