Data warehousing is a process for collecting, storing, and delivering decision-support data for some or all of an enterprise. Data warehousing is a broad subject that is described point-by-point in this Refcard. A data warehouse is one of the artifacts created in the data warehousing process.

William (Bill) H. Inmon has provided an alternate and useful definition of a data warehouse: "A subject-oriented, integrated, time-variant, and nonvolatile collection of data in support of management's decision-making process."

As a total architecture, data warehousing involves people, processes, and technologies to achieve the goal of providing decision-support data that is consistent, integrated, standardized, and easy to understand.

See the book The Analytical Puzzle: Profitable Data Warehousing, Business Intelligence and Analytics (ISBN 978-1935504207) for details.

What a Data Warehouse is and is Not

A data warehouse is a database whose data includes a copy of operational data. This data is often obtained from multiple data sources and is useful for strategic decision-making. It does not, however, contain original data.

"Data warehouse," by the way, is not another name for "database." Some people incorrectly use the term "data warehouse" as if it's a generic name for a database. A data warehouse does not only consist of historic data — it can be made up of analytics and reporting data, too. Transactional data that is managed in application data stores will not reside in a data warehouse.

Data Warehouse Architecture Components

The data warehouse's technical architecture includes data sources, data integration, BI/analytics data stores, and data access.

Data Warehouse Tech Stack

Infrastructure Architecture

The data warehouse tech stack is built on a fundamental framework of hardware and software known as the infrastructure.

Using a data warehouse appliance or a dedicated database infrastructure helps support the data warehouse. This technique tends to yield the highest performance. The data warehouse appliance is optimized to provide database services using massively parallel processing (MPP) architecture. It includes multiple tightly coupled computers with specialized functions, plus at least one array of storage devices that are accessed in parallel. Specialized functions include system controller, database access, data load, and data backup.

Data warehouse appliances provide high performance. They can be up to 100x faster than the typical database server. Consider the data warehouse appliance when more than 2TB of data must be stored.

Data Architecture

Data architecture is a blueprint for the management of data in an enterprise. The data architect builds a picture of how multiple sub-domains work. Some of these subdomains are data governance, data quality, ILM (information lifecycle management), data framework, metadata and semantics, master data, and, finally, business intelligence.

Data Architecture Sub-Domains

Data Flow

The diagram below displays how data flows through the data warehouse system. Data first originates from the data sources, such as inventory systems (systems stored in data warehouses and operational data stores). The data stores are formatted to expose data in the data marts that are then accessed using BI and analytics tools.

Data is the raw material through which we can gain understanding. It is a critical element in data modeling, statistics, and data mining. It is the foundation of the pyramid that leads to wisdom and to informed action.

Data Attribute Characteristics

Three levels of data modeling are developed in sequence:

1. Conceptual Data Model - a high level model that describes a problem using entities, attributes, and relationships.
2. Logical Data Model - a detailed data model that describes a solution in business terms, and that also uses entites, attributes, and relationships.
3. Physical Data Model - a detailed data model that defines database objects, such as tables and columns. This model is needed to implement the models in a database and produce a working solution.

Entities

An entity is a core part of any conceptual and logical data model. An entity is an object of interest to an enterprise — it can be a person, organization, place, thing, activity, event, abstraction, or idea. Entities are represented as rectangles in the data model. Think of entities as singular nouns.

Attributes

An attribute is a characteristic of an entity. Attributes are categorized as primary keys, foreign keys, alternate keys, and non-keys, as depicted in the diagram below.

Relationships

A relationship is an association between entities. Such a relationship is diagrammed by drawing a line between the related entities. The following diagram depicts two entities — Customer and Order — that have a relationship specified by the verb phrase "places" in this way: Customer Places Order.

Cardinality

Cardinality specifies the number of entities that may participate in a given relationship, expressed as one-to-one, one-to-many, or many-to-many, as depicted in the following example.

Cardinality is expressed as minimum and maximum numbers. In the first example below, an instance of entity A may have one instance of entity B, and entity B must have one and only one instance of entity A. Cardinality is specified by putting symbols on the relationship line near each of the two entities that are part of the relationship.

In the second case, entity A may have one or more instances of entity B, and entity B must have one and only one instance of entity A.

Minimum cardinality is expressed by the symbol farther away from the entity. A circle indicates that an entity is optional, while a bar indicates that an entity is mandatory. At least one is required.

Maximum cardinality is expressed by the symbol closest to the entity. A bar means that a maximum of one entity can participate, while a crow's foot (a three-prong connector) means that many entities may participate. This means a large unspecified number.

Normalization is a data modeling technique that organizes data by breaking it down to its lowest level, i.e. its "atomic" components, to avoid duplication. This method is used to design the atomic data warehouse part of the data warehousing system.

The atomic data warehouse (ADW) is an area where data is broken down into low-level components in preparation for exporting to data marts. The ADW is designed using normalization and methods that make for speedy history loading and recording.

Header and Detail Entities

The ADW is organized into non-changing data with logical keys and changeable data that supports tracking of changes and rapid load/insert. Use an integer as the primary surrogate key. Then, add the effective date to track changes.

Associative Entities

Track the history of relationships between entities using an associative entity with effective dates and expiration dates.

Atomic DW Specialized Attributes

Use specialized attributes to improve ADW efficiency and effectiveness. Identify these attributes using a prefix of ADW_.

Supporting data is required to enable the data warehouse to operate smoothly. Here is some supporting data:

• Code management and translation
• Data source tracking
• Error logging

Code Translation

Data warehousing requires that codes, such as gender code and units of measure, be translated to standard values aided by code-translation tables like these:

• Code Set – Group of codes, such as "Gender Code"
• Code – An individual code value
• Code Translation – Mapping between code values

Data-Source Tracking and Logging

Data-source tracking provides a means of tracing where data originated within a data warehouse:

• Data Source – identifies the system or database
• Data Process – traces the data-integration procedure
• Data Process Log – traces each data warehouse load

Message Logging

Message logging provides a record of events that occur while loading the data warehouse:

• Data Process Log – traces each data warehouse load
• Message Type – specifies the kind of message
• Message Log – contains an individual message

A dimensional database is a database that is optimized for query and analysis and is not normalized like the atomic data warehouse. It consists of fact and dimension tables, where each fact is connected to one or more dimensions.

Sales Order Fact

The sales order fact includes the measurer's order quantity and currency amount. Dimensions of Calendar Date, Product, Customer, Geo Location, and Sales Organization put the sales order fact into context. This star schema supports looking at orders in a cubical way, enabling slicing and dicing by customer, time, and product.

A fact is a set of measurements. It tends to contain quantitative data that gets presented to users. It often contains amounts of money and quantities of things. Facts are surrounded by dimensions that categorize the fact.

Anatomy of a Fact

Facts are SQL tables that include:

• Table Name – a descriptive name usually containing the word 'Fact'
• Primary Keys – attributes that uniquely identify each fact occurrence and relate it to dimensions
• Measures – quantitative metrics

Event Fact Example

Event facts record single occurrences, such as financial transactions, sales, complains, or shipments.

Snapshot Fact

The snapshot fact captures the status of an item at a point in time, such as a general ledger balance or inventory level.

Cumulative Snapshot Fact

The cumulative snapshot fact adds accumulated data, such as year-todate amounts, to the snapshot fact.

Aggregated Fact

Aggregated facts provide summary information, such as general ledger totals during a period of time, or complaints per product per store per month.

Fact-less Fact

The fact-less fact tracks an association between dimensions rather than quantitative metrics. Examples include miles, event attendance, and sales promotions.

A dimension is a database table that contains properties that identify and categorize. The attributes serve as labels for reports and as data points for summarization. In the dimensional model, dimensions surround and qualify facts.

Data and Time Dimensions

Date dimensions support trend analysis. Date dimensions include the date and its associated week, month, quarter, and year. Time-of-day dimensions are used to analyze daily business volume.

Multiple-Dimension Roles

One dimension can play multiple roles. The date dimension could play roles of a snapshot date, a project start date, and a project end date.

Degenerate Dimension

A degenerate dimension has a dimension key without a dimension table. Examples include transaction numbers, shipment numbers, and order numbers.

Slowly-Changing Dimensions

Changes to dimensional data can be categorized into levels:

Data integration is a technique for moving data or otherwise making data available across data stores. The data integration process can include extraction, movement, validation, cleansing, transformation, standardization, and loading.

Extract Transform Load (ETL)

In the ETL pattern of data integration, data is extracted from the data source and then transformed while in flight to a staging database. Data is then loaded into the data warehouse. ETL is strong for batch processing of bulk data.

Extract Load Transform (ELT)

In the ELT pattern of data integration, data is extracted from the data source and loaded to staging without transformation. After that, data is transformed within staging and then loaded to the data warehouse.

Change Data Capture (CDC)

The CDC pattern of data integration is strong in event processing. Database logs that contain a record of database changes are replicated near real time at staging. This information is then transformed and loaded to the data warehouse.

CDC is a great technique for supporting real-time data warehouses.

Innovation in Data Warehouse Technology

Thanks to the agility offered by today’s cloud-based data warehouse solutions, there are cutting-edge innovations that can automate some of the key aspects of data warehousing. For example, the ETL process described earlier has changed considerably thanks to machine learning and natural language processing, ultimately resulting in its complete automation. What’s more, data warehouse storage and compute have also benefited from automated optimization, saving data analysts time on tasks associated with querying, storage, and scalability, which dramatically cuts down costs, coding time, and resources.

As a modern solution, cloud data warehouse automation saves endless hours of coding and modeling for data ingestion, integration, and transformation. The tasks listed below can be now easily automated and seamlessly connected to third-party solutions, such as business intelligence visualization tools, via the cloud:

• Automate data source connections.
• Seamlessly connect to third-party SaaS APIs.
• Easily connect to the most common storage services.

New technology now exists that automates data schema modeling, where adaptive schema changes at real time along with the data, and changes are seamless. You would only need to just upload the data sources, everything else is automated including the following tasks:

• Data types are automatically discovered, and a schema is generated based on the initial data structure.
• Likely relationships between tables are automatically detected and used to model a relational schema.
• Aggregations are automatically generated.
• Table history feature allows you store data uploaded from API data sources, so you can compare and analyze data from different time periods.

Automated Query Optimization

Automated cloud data warehousing technology exists that can automatically re-index the schema and perform a series of optimizations on the queries and data structure to improve runtime based on algorithms that assess usage so that:

• Re-indexing happens automatically whenever the algorithm detects changes in query patterns.
• Redistributing the data across nodes to improve data locality and join performance is done automatically.

Solving Concurrency Issues
To remedy concurrency issues, new cloud data warehousing technologies today can separate storage from compute and increase the compute nodes based on the amount of connections. Consequently, the number of available clusters scales with the number of users and the intensity of the workload, supporting hundreds of parallel queries that are load-balanced between clusters.

Storage Optimization
Data warehouse automation has also vastly improved how data is stored and used. New, “smart” data warehouse technologies constantly run periodic processes to mark data and optimize the storage based on usage. Smart data warehouse technology scales up and down based on the data volume. Scaling happens automatically behind the scenes, keeping clusters available for both reads and write, and thus ingestion can continue uninterrupted. When the scaling is complete, the old and new clusters are swapped instantly. In addition, the data warehouse maintenance itself has been greatly improved, as well, by automating the cleaning and compressing of tables to boost database performance.