Published in TDAN.com January 2000
Many of the Business Intelligence tools look way cool. They provide graphs, moving targets, drill-downs, and drill-through. But much of the work in an operational data warehouse involves getting
the data from operational systems into the data warehouse so that business intelligence tools can display those pretty pictures. This paper addresses the extraction, transformation, and load
components of data warehousing. We’ll look at issues in extraction, transformation, and loading and common approaches to loading data. We assume that source data structures are generally not
similar to target data structures (e.g., flat files and normalized tables).
At the risk of being a bit simplistic, extraction, transformation, and load requires three main steps:
- Read the source data
- Apply business, transformation, and technical rules
- Load the data
Figure 1 shows this data flow. After the process reads the data, it must transform the data by applying technology, transformation, and business rules to it. An example of a transformation rule is:
“Convert values in field x to integer.” An example of a business rule is: “Customers must purchase products in the list of ‘Washer’, ‘Dryer’, “Refrigerator’.”
Applying business, transformation, and technology rules to data means generating keys, transforming codes, converting datatypes, parsing, merging, and many other operations. Once the data is in an
appropriate technical and business format, the ETL process can load it into target tables. Note that these steps can potentially be performed many places in the system chain. For example, extract
programs could transform code values as they read from DB2 tables or VSAM datasets. A Perl program could parse a file and generate keys for use by bcp or DTS. Finally, stored procedures could split
staging table rows in a set-oriented fashion.
The extraction, transformation, and load environment consists of three architectures (data, application, and technology), and a group of people (management, developers, and support). Figure 2 shows
an abstract ETL environment.
The data architecture includes the data itself and its quality as well as the various models that represent the data, data structures, business and transformation rules, and business meaning
embodied in the data and data structures. Data architecture models include conceptual data models, logical data models, physical data models, and physical representations such as COBOL
copybooks, C structures, and SQL DDL statements.
Technology architecture includes computers, operating systems, data management systems (e.g., Oracle, Sybase, Btrieve), networks, network elements, and the models that represent these. Technology
models include network diagrams, computer specifications, and technology standards such as TCP/IP, SQL, and ODBC.
The application architecture includes operational systems, including any support programs such as the ETL programs. Application models include context diagrams, functional decomposition diagrams,
state transition diagrams, and data flow diagrams.
process. If these models don’t exist, you’ll have to create many of them. Good data, application, and technology models are essential to creating a data warehouse that meets the
needs of the business.
Common Data Sources and Systems
ETL processes populate SQL Server tables primarily from operational systems, such as Accounts Receivable, Customer Care, and Billing systems. Other data sources may include files from external data
providers, the Internet, or data from user desktops (such as a list of products). File formats may include spreadsheets, text files, dBASE, or image files in JPG or GIF format. IBM host data files
(in EBCDIC) usually get translated to ASCII either by FTP or a gateway. These days, almost all relational systems support ODBC or OLE-DB connections.
The data source also depends on what subject area(s) is implemented in the data warehouse. Common subject areas include customer, product, marketing, and sales. Financial data usually comes from
the accounting package. Customer (people) data may come from customer care systems, marketing databases, or third-party databases. Products, features, and prices often come from the marketing
While an ETL solution belongs in the application architecture, the data, technology, and “person” architectures also influence the ETL approach. Data issues include:
- Quality of data
- How similar are the source and target data structures?
- What kinds of dependencies exist in the data?
- How is meta data used?
- How “complex” are the data relationships?
People issues include:
- What technology does management feel comfortable with?
- What in-house expertise does your shop have?
- Who will support the ETL processes?
Technology issues include:
- What is the volume and frequency of load?
- How much disk space will be needed?
- To what extent are source and target platforms interoperable?
- How will the ETL processes be scheduled?
These factors influence the approach to loading the data warehouse, which also affects the cost of the solution (in labor and/or products), and its ease of development, understandability, and
maintenance. We’ll address each area in the following sections.
Data Architecture Issues
Similarity of Source and Target Data Structures. Target data structures should have been created from a physical data model, which in turn should have been created from a logical
data model that was the result of modeling sessions with the business people. The more different the source and target data structures, the more complex the ETL algorithms, processing, and
maintenance effort. In terms of physical implementation, flat files are often the most common data source, though these are usually derived from other file formats. If the source and target data
structures are not similar, the load processes will typically have to parse the records, transform values, validate values, substitute code values, and generate keys, etc.
Quality of Data. Poor quality data should have been identified in a data assessment phase and, ideally, cleaned in the originating operational systems. Common data quality issues
include missing values, code values not in the correct list of values, dates, and referential integrity issues. It makes no sense to load the data warehouse with poor quality data. As an example,
if the data warehouse will be used for database marketing, the addresses should be validated to avoid returned mail.
Complexity of the Source Data. Depending on the sourcing team’s background, some data sources are more complex than others. Examples of complex source data may include
multiple record types, bit fields, COBOL OCCURS clauses, and packed decimal fields. This kind of data will translate into requirements for the ETL tool or custom-written solution since they are
unlikely to exist in the target data structures. Individuals on the sourcing team that are unfamiliar with these types may need to do some research in these areas.
Dependencies in the Data. Dependencies in the data will determine the order in which you load tables. Dependencies also tend to reduce parallel loading operations, especially if
data is merged from different systems, which are on a different business cycle. Complex dependencies will also tend to make the load processes more complex, encourage bottlenecks, and make support
Meta data. Technical meta data describes not only the structure and format of the source and target data sources, but also the mapping and transformation rules between them. Meta
data should be visible (and usable) to both programs and people. Microsoft’s DTS uses meta data in the repository to transform and load data sources.
Logging. ETL processes should log information about the data sources they read, transform, and write. Key information includes date processed, number of rows read/rows written,
errors encountered, and rules applied. This information is critical for quality assurance, and serves as an audit trail. The logging process should be rigorous enough so that you can trace data in
the data warehouse back to the source. In addition, this information should be available as the processes are running to assist in estimated completion times.
Notification. The ETL requirements should specify what makes an acceptable load. The ETL process should notify the appropriate support people when a load fails or has errors.
Ideally, the notification process should plug into your existing error tracking system.
Cold Start, Warm Start. Unfortunately, systems crash. You need to be able to take the appropriate action if the system crashes with your ETL process running. Partial loads
can be a pain. Depending on the size of your data warehouse and volume of data, you may want to start over (cold start) or start from the last known successfully-loaded records (warm start). The
logging process should provide information about the state of the ETL process.
Management’s Comfort Level with Technology. How conversant is management with data warehousing architecture? Will you have a data warehouse manager? Does management have
development background? They may suggest doing all the ETL processes with Visual Basic. Comfort level is a valid concern, and these concerns will constrain your options. Seminars, magazine
articles, and industry “expert opinion” will help you get your points across.
In-House Expertise. What is your shop’s tradition? VB/SQL Server? MVS/CICS/COBOL? VAX/VMS/DCL? UNIX/Sybase? Do your UNIX machines have the Perl distributions? ETL solutions
will be drawn from current conceptions, skills, and toolsets. Acquiring, transforming, and loading the data warehouse is an ongoing process and will need to be maintained and extended as more
subject areas are added to the data warehouse.
Support. Once the ETL processes have been created, support for them, ideally, should plug into existing support structures, including people with the appropriate skill-sets,
notification mechanisms, and error-tracking systems. If you use a tool for ETL, the support staff may need to be trained. The ETL process should be documented, especially in the area of auditing
Interoperability between Platforms. There must be a way for systems on one platform to talk to systems on another. FTP (File Transfer Protocol) is a common way to transfer data
from one system to another. FTP requires a physical network path from one system to another as well as the Internet Protocol on both systems. External data sources usually come on a floppy, tape,
or an Internet server. More flexible (and more expensive) options for data access include database gateways such as OmniConnect.
Volume and Frequency of Loads. Since the data warehouse is loaded via batch programs, a high volume of data will tend to reduce the batch window. The volume of data also affects
the back out and recovery work. Fast load programs (e.g., from Platinum) reduce the time it takes to load data into the data warehouse.
Disk Space. Not only does the data warehouse potentially have requirements for a lot of disk space, but there is also a lot of “hidden” disk space needed for staging areas and
intermediate files. For example, you may want to extract data from source systems into flat files and then transform the data to other flat files for load by native DBMS utilities.
Scheduling. Loading the data warehouse could involve hundreds of source files, which originate on different systems, use different technology, and are produced at different times.
A monthly load may be common for some portions of the warehouse and a quarterly load for others. Some loads may be on-demand, such as lists of products or external data. Some extract programs may
run on a different type of system than your scheduler.
Listed below are some common operations you may need to perform on the data you load.
|Generate a key||Identity column; primary key generator|
|Translate a code||if-then logic; lookup|
|Split data from one source into two targets||Multiple write statements|
|Merge two data sources into one||Join from multiple source tables|
|Log errors and progress information||Log/Schedule table that all load processes write to|
|Load code tables||Set of scripts|
There are four common approaches to transforming and loading data into the data warehouse.
- Build a custom solution
- Buy a code generator
- Buy a data transformation engine
- Use a mixture of the above
All the common data management systems have utilities for extraction and load, and a key challenge in choosing the ETL approach is deciding what tools to use for what parts of the process. All the
data, technology, application, and people issues we talked about earlier will influence your approach to the ETL processes. Let’s look at the four main approaches.
Build a Custom Solution. There are really three approaches here. The first makes use of SQL Server staging tables and stored procedures. The second uses a programming language such
as C, C++, Perl, or Visual Basic. The third approach involves building your own code generator using technical meta data.
The staging table approach loads data into a staging table and uses stored procedures to transform the data. In general, procedural SQL is much easier and quicker to write, debug,
and maintain than 3GL code, but this solution is likely to be “slower” than a 3GL or engine solution. Transact-SQL has a number of useful functions for data validation and transformations,
including isdate(), dateadd(), substring(), convert(), and “case.”
A custom-written solution written in a language like C or C++ is very powerful and flexible, but it may also be more expensive to write and maintain than the results of some of the
other approaches. The basic algorithm is (1) extract source data, (2) apply business, technology, and transformation rules, and (3) load data. Depending on the language you chose, there are likely
to be a number of libraries and components that may assist in the process. Abstracting the operations via classes and table-driven rules can help reuse.
There are a number of ways to approach building your own code generator, and they all are interesting. A code generator reads meta data or a specification language and generates
useful code based on the meta data. Examples of meta data include SQL Server system catalogs, the Erwin metamodel, and the Microsoft repository. Examine the system procedures such as “sp_help” to
see how they use the system catalog tables. A C-based approach might use lex and yacc.
Buy Code Generator. These tools read meta data and generate transformation source code, which can then be compiled and run. Typical target languages are COBOL and C++. Vendors such
as Prism, ETI, and Carleton market these kinds of tools.
Buy Data Transformation Engine. Transformation engines usually operate at a higher level of abstraction than even visual development tools. Generally you can accomplish much more
with an engine in far fewer statements than you can with 3GLs. DTS is an example of a data transformation engine. It’s extensible, reads a number of data sources, and fits right into the SQL
Server environment. Other tools include PowerCenter (Informatica), Reliant Data Bridge (Reliant Data Systems), Data Stage (Ardent), Sagent Solution (Sagent), and Data Junction (Data Junction).
Engines typically have a framework that defines how the extraction and transformation works. While they are not as flexible as custom-written solutions, some are extensible. Engines will also
generally import COBOL copybooks and SQL DDL for use as meta data. Some use a repository housed in a standard relational database system.
Mixed Solution. In many cases, you won’t be able to create an end-to-end solution with just one tool. You may use native unload and load utilities with a code generator. Or
you may use a transformation engine and post-process the data once loaded into the data warehouse. The challenge is to determine where each tool plays best in your specific environment.
This article examined issues and approaches in transforming and loading data into the data warehouse. The ETL process is a development process like any other. To create or purchase a viable ETL
solution, it takes an understanding of the requirements, source and target data and data structures, and the technical and political environment.