This paper describes how software language engineering is applied to the process of data warehouse creation. The creation of a data warehouse is a complex process and therefore costly. My approach decomposes the data warehouse creation process into di erent aspects. These aspects are described with di erent languages which are integrated by a metamodel. Based on this metamodel, large parts of the data warehouse creation process can be generated. With my approach data warehouses are created more comfortable in less time.
Health Reporting describes the preparation and presentation of health relevant issues relating to population. It is used to give information to stakeholders in the health care system, politicians and interested non-professionals. Furthermore, risks are identi ed and appropriate warnings issued. In the Federal State of North Rhine-Westphalia this task is carried out by the government agency for public health called LIGA (Landesinstitut fur Gesundheit und Arbeit). LIGA provides a variety of di erent reports and is able to answer ad hoc questions. The reports are based on data from di erent sources and di erent systems with di erent formats. These sources are e.g. data from di erent public health departments or insurances. To ful ll these requirements, software support is needed. This means data has to be integrated into one dataset so that di erent sources can be related. Support is also needed for transforming data on regular basis into the integrated dataset. Also frontend and report generation has to be developed.
One software system providing this support is the MUSTANG platform
developed at OFFIS. MUSTANG is used at LIGA. The base of each MUSTANG
instance is an integrated dataset, this is also called data warehouse (DWH). In
an integrated dataset all relevant organizational knowledge is stored for
complex analysis. For fast data navigation and analysis, Online Analytical
Processing (OLAP) [
Analysis of organizational data: To nd data that can be used in the resulting DWH, existing data sources have to be analyzed. This analysis includes the content, format and the accessibility of the data. This kind of data is called fact data. In a DWH this fact data is extracted and integrated into the so called integrated dataset. De ne information demand: De ne what information should be provided by the DWH. This can be simple gures or complex computations. Data source transformation: Fact data has to be transformed in the data format of the integrated dataset. Therefore, for every data source a transformation has to be designed that translates data into the format of the integrated data set and stores it there. De ne multidimensional data model: De nes how fact data can be described multidimensionally and grouped together in hierarchies. Data quality: Based on DWH, analysis decisions are made so it is important to de ne data quality standards and how to identify invalid data.
To perform these steps no standardized process exists. Documentation of these steps is usually done with a large number of documents. A problem with this kind of documentation is missing, distributed or inconsistent information. Another aspect is that during realization a lot schematic work has to be done. For example, a multidimensional schema has to be designed and realized in the OLAP system, the integrated dataset, and at the frontend software. 2
Data Warehouse analysis and design as described in [
Another eld of related work is automated creation of DWHs. The feasibility
to connect MDA with the DWH process has been shown in [
More related to SLE is the work of Rizzi. His group deals with modeling di
erent aspects of DWHs. For example in [
All these approaches only deal with a certain aspect of DWH creation, not with the whole process with language support. In my thesis, I will develop an approach with languages that cover the whole process of DWH creation. These languages are integrated through a common metamodel and can deal with multidimensional structures. Based on the metamodel I will create transformations that allow generating large parts of the resulting DWH. With these transformations, schematic work in the step of realization is reduced. Furthermore, I will create a process model that orders the steps described in combination with the developed languages to improve documentation. With the process model, the di erent aspects are connected and re ned.
Data warehouses, as describes here, are very complex systems with di erent views, aspects, and levels of detail. To create a single language, these systems are di cult and not easy to use and maintain. Therefore, it is necessary to decompose a DWH creation into di erent aspects and create languages for each. The di erent languages will be used by di erent roles and provide a di erent level of detail. A rst result is that the process can be decomposed in six aspects:
Data Sources Schemas: Describes all relevant or available data stored in its operational system of an organization. This aspect contains the subject, the representation, and technical accessibility. The advantage is that all relevant sources are described together with their formats and accessibility at one place. For this aspect the development of an own language may not be necessary but a meta model will be su cient.
Data Source Transformation: Describes how data from the sources have to be transformed to match the analysis schema. Such a description makes it possible to abstract from the concrete target system and some automatic matching process can be applied.
Analysis Schema: This aspect describes the multidimensional schema of
the resulting DWH. The multidimensional schema consists of cubes and
dimensions. In a cube, fact data is stored. Each cell represents fact data that is
numerical and can be aggregated. They are described by dimensional metadata. In a
cube with a time dimension, fact data is stored for dates so monthly and yearly
values are computed by aggregation. The language is based on ADAPT. With
such a language it is possible to reduce schematic work in the realization phase,
as shown in [
Measures: Describes what kind of information is intended to be stored in the DWH. This can be simple fact data like infections. It also includes the designated granularity of information. For example, to predict an epidemic infect, data should be available for every date. Measures link fact data with mathematical operations. In Health Reporting, these are mostly crude rates, interest, and average. Measures are usually de ned by domain experts, in case of LIGA by epidemiologists. Measures are re ned in the analysis schema. With a language for measures the de nition can be used in the realization process and does not need to be reimplemented.
Hierarchy: The hierarchy aspect is a central one in my thesis because all other aspects use this directly or indirectly. This aspect describes how data is aggregated. In most cases hierarchies have many members and a complex structure and they are used in the multidimensional model. For example, imagine a geographical dimension that contains countries and cities. Using an own language, these structures can be modeled appropriate. It can help to build a repository that can be reused in a di erent DWH. The concrete syntax of the hierarchy language is based on ADAPT. However, it only allows to model hierarchies conceptually. To model hierarchies logically a tabular extension was created. To create parent-child relationships a query language was also integrated.
Data Quality: In a DWH it is important to ensure that certain quality
issues are met. Naturally, data quality is an aspect of the data sources but when
integrating di erent systems it would be very costly to deal with quality at the
sources because many di erent systems have to be considered and changing these
systems is rarely possible. Data quality is a large research eld. In my approach,
I want to integrate existing approaches to show how data quality issues can be
integrated. InDaQu [
The languages for the described aspects are developed independently. Each
language is developed with SLE techniques [
The integrated metamodel covers all aspects of a DWH at one place. A
common metamodel is a standardized documentation of the whole DWH system.
Other possiblities of an integrated metamodel have been shown in [
The advantage of using SLE for the creation of DWHs is that experts can design and analyze all aspects of the DWH independently in adequate domain speci c languages. Using the integrated metamodel the generation of a DWH is easier because all information and transformations are in that single model. 4
My hypothesis will be validated via implementation. I will implement the described languages, metamodels, and transformations on basis of the MUSTANG platform. My prototype will be able to generate a con guration for a DWH. I will regenerate parts of the LIGA DWH that was developed by OFFIS. I will compare the steps taken. When using my approach these steps will be reduced. 5
Currently I have developed three languages for hierarchies, analysis schema, and data quality. I integrated them in a common metamodel. Based on this, transformations for generating multidimensional schemas in databases and integration interface with consistency were built. The next action is to develop languages for data sources and data integration as well as the extension of the common metamodel. The current state of my thesis shows that modeling and generation of data warehouses can be possible and reasonable with SLE.