The Kimball Group has been exposed to hundreds of successful data warehouses. Careful study of these successes has revealed a set of extract, transformation, and load (ETL) best practices. We first described these best practices in an Intelligent Enterprise column three years ago. Since then we have continued to refine the practices based on client experiences, feedback from students and continued research. As a result, we have carefully restructured these best practices into 34 subsystems that represent the key ETL architecture components required in almost every dimensional data warehouse environment. No wonder the ETL system takes such a large percentage of data warehouse and BI project resources!
The good news is that if you study these 34 subsystems, you’ll recognize almost all of them and will be on the way to leveraging your experience as you build your ETL system. While we understand and accept the industry’s accepted acronym, the “ETL” process really has four major components: Extracting, Cleaning and Conforming, Delivering and Managing. Each of these components and all 34 subsystems contained therein are explained below.
EXTRACTING: GETTING DATA INTO THE DATA WAREHOUSE
To no surprise, the initial subsystems of the ETL architecture address the issues of understanding your source data, extracting the data and transferring it to the data warehouse environment where the ETL system can operate on it independent of the operational systems. While the remaining subsystems focus on the transforming, loading and system management within the ETL environment, the initial subsystems interface to the source systems to access the required data. The extract-related ETL subsystems include:
- Data Profiling (subsystem 1) – Explores a data source to determine its fit for inclusion as a source and the associated cleaning and conforming requirements.
- Change Data Capture (subsystem 2) – Isolates the changes that occurred in the source system to reduce the ETL processing burden.
- Extract System (subsystem 3) – Extracts and moves source data into the data warehouse environment for further processing.
CLEANING AND CONFORMING DATA
These critical steps are where the ETL system adds value to the data. The other activities, extracting and delivering data, are obviously important, but they simply move and load the data. The cleaning and conforming subsystems change data and enhance its value to the organization. In addition, these subsystems should be architected to create metadata used to diagnose source-system problems. Such diagnoses can eventually lead to business process re-engineering initiatives to address the root causes of dirty data and to improve data quality over time.
The ETL data cleaning process is often expected to fix dirty data, yet at the same time the data warehouse is expected to provide an accurate picture of the data as it was captured by the organization’s production systems. It’s essential to strike the proper balance between these conflicting goals. The key is to develop an ETL system capable of correcting, rejecting or loading data as is, and then highlighting, with easy-to-use structures, the modifications, standardizations, rules and assumptions of the underlying cleaning apparatus so the system is self-documenting.
The five major subsystems in the cleaning and conforming step include:
- Data Cleansing System (subsystem 4) – Implements data quality processes to catch quality violations.
- Error Event Tracking (subsystem 5) – Captures all error events that are vital inputs to data quality improvement.
- Audit Dimension Creation (subsystem 6) – Attaches metadata to each fact table as a dimension. This metadata is available to BI applications for visibility into data quality.
- Deduplication (subsystem 7) – Eliminates redundant members of core dimensions, such as customers or products. May require integration across multiple sources and application of survivorship rules to identify the most appropriate version of a duplicate row.
- Data Conformance (subsystem 8 ) – Enforces common dimension attributes across conformed master dimensions and common metrics across related fact tables.
DELIVERING: PREPARE FOR PRESENTATION
The primary mission of the ETL system is the hand-off of the dimension and fact tables in the delivery step. There is considerable variation in source data structures and cleaning and conforming logic, but the delivery processing techniques are more defined and disciplined. Careful and consistent use of these techniques is critical to building a successful dimensional data warehouse that is reliable, scalable and maintainable.
Many of these subsystems focus on dimension table processing. Dimension tables are the heart of the data warehouse. They provide the context for the fact tables and hence for all the measurements. For many dimensions, the basic load plan is relatively simple: perform basic transformations to the data to build dimension rows to be loaded into the target presentation table.
Preparing fact tables is certainly important as they hold the key measurements of the business that users want to see. Fact tables can be very large and time consuming to load. However, preparing fact tables for presentation is typically more straightforward.
The delivery systems in the ETL architecture consist of:
- Slowly Changing Dimension (SCD) Manager (subsystem 9) – Implements logic for slowly changing dimension attributes.
- Surrogate Key Generator (subsystem 10) – Produces surrogate keys independently for every dimension.
- Hierarchy Manager (subsystem 11) – Delivers multiple, simultaneous, embedded hierarchical structures in a dimension.
- Special Dimensions Manager (subsystem 12) – Creates placeholders in the ETL architecture for repeatable processes supporting an organization’s specific dimensional design characteristics, including standard dimensional design constructs such as junk dimensions, mini-dimensions and behavior tags.
- Fact Table Builders (subsystem 13) – Construct the three primary types of fact tables: transaction grain, periodic snapshot and accumulating snapshot.
- Surrogate Key Pipeline (subsystem 14) – Replaces operational natural keys in the incoming fact table record with the appropriate dimension surrogate keys.
- Multi-Valued Bridge Table Builder (subsystem 15) – Builds and maintains bridge tables to support multi-valued relationships.
- Late Arriving Data Handler (subsystem 16) – Applies special modifications to the standard processing procedures to deal with late-arriving fact and dimension data.
- Dimension Manager (subsystem 17) – Centralized authority who prepares and publishes conformed dimensions to the data warehouse community.
- Fact Table Provider (subsystem 18) – Owns the administration of one or more fact tables and is responsible for their creation, maintenance and use.
- Aggregate Builder (subsystem 19) – Builds and maintains aggregates to be used seamlessly with aggregate navigation technologies for enhanced query performance.
- OLAP Cube Builder (subsystem 20) – Feeds data from the relational dimensional schema to populate OLAP cubes.
- Data Propagation Manager (subsystem 21) – Prepares conformed, integrated data from the data warehouse presentation server for delivery to other environments for special purposes.
MANAGING THE ETL ENVIRONMENT
A data warehouse will not be a success until it can be relied upon as a dependable source for business decision making. To achieve this goal, the ETL system must constantly work toward fulfilling three criteria:
- Reliability. The ETL processes must run consistently to provide data on a timely basis that is trustworthy at any level of detail.
- Availability. The data warehouse must meet its service level agreements. The warehouse should be up and available as promised.
- Manageability. A successful data warehouse is never done; it constantly grows and changes along with the business. Thus, ETL processes need to evolve gracefully as well.
The ETL management subsystems are the key architectural components that help achieve the goals of reliability, availability and manageability. Operating and maintaining a data warehouse in a professional manner is not much different than other systems operations: follow standard best practices, plan for disaster and practice. Many of you will be very familiar with the following requisite management subsystems:
- Job Scheduler (subsystem 22) – Reliably manages the ETL execution strategy, including the relationships and dependencies between ETL jobs.
- Backup System (subsystem 23) – Backs up the ETL environment for recovery, restart and archival purposes.
- Recovery and Restart (subsystem 24) – Processes for recovering the ETL environment or restarting a process in the event of failure.
- Version Control (subsystem 25) – Takes snapshots for archiving and recovering all the logic and metadata of the ETL pipeline.
- Version Migration (subsystem 26) – Migrates a complete version of the ETL pipeline from development into test and finally into production.
- Workflow Monitor (subsystem 27) – Ensures that the ETL processes are operating efficiently and that the warehouse is being loaded on a consistently timely basis.
- Sorting (subsystem 28) – Serves the fundamental, high-performance ETL processing role.
- Lineage and Dependency (subsystem 29) – Identifies the source of a data element and all intermediate locations and transformations for that data element or, conversely, start with a specific data element in a source table and reveals all activities performed on that data element.
- Problem Escalation (subsystem 30) – Support structure that elevates ETL problems to appropriate levels for resolution.
- Paralleling and Pipelining (subsystem 31) – Enables the ETL system to automatically leverage multiple processors or grid computing resources to deliver within time constraints.
- Security (subsystem 32) – Ensures authorized access to (and a historical record of access to) all ETL data and metadata by individual and role.
- Compliance Manager (subsystem 33) – Supports the organization’s compliance requirements typically through maintaining the data’s chain of custody and tracking who had authorized access to the data.
- Metadata Repository (subsystem 34) – Captures ETL metadata including the process metadata, technical metadata and business metadata, which make up much of the metadata of the total DW/BI environment.
SUMMING IT UP
As you may now better appreciate, building an ETL system is unusually challenging. The required ETL architecture requires a host of subsystems necessary to meet the demanding requirements placed on the data warehouse. To succeed, carefully consider each of these 34 subsystems. You must understand the breadth of requirements and then place an appropriate and effective architecture in place. ETL is more than just extract, transform and load; it’s a host of complex and important tasks.