A COMPARISON OF DATA WARE HOUSE DESIGN MODELS

A MASTERS THESIS in Computer Engineer ing Atilim Univer sity

by BERIL PINAR BAARAN J ANUARY 2005

A COMPARISON OF DATA WARE HOUSE DESIGN MODELS

A THESIS SUBMITTED TO THE GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES OF ATILIM UNIVERSITY BY BERIL PINAR BAARAN

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE IN THE DEPARTMENT OF COMPUTER ENGINEERING

J ANUARY 2005

Approval of the Graduate School of Natural and Applied Sciences

_____________________ Prof. Dr. Ibrahim Akman Director I certify that this thesis satisfies all the requirements as a thesis for the degree of Master of Science. _____________________ Prof. Dr. Ibrahim Akman Head of Department This is to certify that we have read this thesis and that in our opinion it is fully adequate, in scope and quality, as a thesis for the degree of Master of Science. _____________________ Prof. Dr. Ali Yazici Co-Supervisor _____________________ Dr. Deepti Mishra Supervisor

Examining Committee Members Prof. Dr. Ali Yazici Dr. Deepti Mishra Asst. Prof. Dr. Nergiz E. altay Dr. Ali Arifolu Asst. Prof. Dr. idem Turhan ii _____________________ _____________________ _____________________ _____________________ _____________________

ABSTRACT

A COMPARISON OF DATA WARE HOUSE DESIGN MODELS Baaran, Beril Pnar M.S., Computer Engineering Department Supervisor: Dr. Deepti Mishra Co-Supervisor: Prof. Dr. Ali Yazici January 2005, 90 pages

There are a number of approaches in designing a data warehouse both in conceptual and logical design phases. The generally accepted conceptual design approaches are dimensional fact model, multidimensional E/R model, starER model and object-oriented multidimensional model. And in the logical design phase, flat schema, terraced schema, star schema, fact constellation schema, galaxy schema, snowflake schema, star cluster schema and starflake schemas are widely used approaches. This thesis proposes a comparison of both the conceptual and the logical design models and a sample data warehouse design and implementation is provided. It is observed that in the conceptual design phase, object-oriented model provides the best solution and for the logical design phase, star schema is generally the best in terms of performance and snowflake is generally the best in terms of redundancy. Keywords: Data Warehouse, Design Methodologies, DF, starER, ME/R, OOMD, flat schema, terraced schema, star schema, fact constellation schema, galaxy schema, snowflake schema, star cluster schema, starflake schema, DTS, Data Analyzer

iii

VER AMBARI TASARIM MODELLER KARILATIRMASI Baaran, Beril Pnar Yksek Lisans, Bilgisayar Mhendislii Blm Tez Yneticisi: Dr. Deepti Mishra Ortak Tez Yneticisi: Prof. Dr. Ali Yazici Ocak 2005, 90 sayfa

Veri ambar tasarmnn kavramsal ve mantksal tasarm aamalar iin birden fazla yaklam vardr. Kavramsal tasarm safhas iin genel olarak kabul grm yaklamlar dimensional fact, multidimensional E/R, starER ve object-oriented multidimensional modelleridir. Mantksal tasarm safhas iin genel olarak kabul grm yaklamlar flat, terraced, star, fact constellation, galaxy , snowflake, star cluster ve starflake emalardr. Bu tez, kavramsal ve mantksal tasarm modellerini karlatrr, rnek bir veri ambar tasarmn ve uygulamasn ierir. Bu tezde, kavramsal tasarm aamasnda object-oriented multidimensional modelinin; mantksal tasarm aamasnda performans kriteri asndan star emann, veri tekrar kriteri a sndan snowflake emann en iyi zmler olduu gzlendi.

Anahtar Kelimeler: Veri Ambar, Tasarm Yntemleri, DF, starER, ME/R, OOMD, flat ema, terraced ema, star ema, fact constellation ema, galaxy ema, snowflake ema, star cluster ema, starflake ema, DTS, Data Analyzer

iv

To my dear husband Thanks for his endless support

ACK NO WLEDGEMENTS

First, I would like to thank my thesis advisor Dr. Deepti MISHRA and cosupervisor Prof. Dr. Ali YAZICI for their guidance, insight and encouragement throughout the study.

I should also express my appreciation to examination committee members Asst. Prof. Dr. Nergiz E. AILTAY, Dr. Ali ARIFOLU, Asst. Prof. Dr. idem TURHAN for their valuable suggestions and comments.

I would like to express my thanks to my husband for his assistance, encouragement and all members of my family for their patience, sympaty and support during the study.

vi

TABLE OF CONTENTS
ABSTRACT ..........................................................................................................................iii Z........................................................................................................................................... iv ACKNOWLEDGEMENTS.................................................................................................. vi TABLE OF CONTENTS..................................................................................................... vii LIST OF TABLES ................................................................................................................. x LIST OF FIGURES............................................................................................................... xi LIST OF ABBREVIATIONS ............................................................................................xiii CHAPTER 1 INTRODUCTION.............................................................................................................. 1 1.1. Scope and outline of the thesis................................................................................... 2 2 DATA WAREHOUSE CONCEPTS ................................................................................. 3 2.1. Definition of Data Warehouse ................................................................................... 3 2.2. Why OLAP systems must run with OLTP ................................................................ 5 2.3. Requirements for Data Warehouse Database Management Systems...................... 8 3 FUNDAMENTALS OF DATA WAREHOUSE ............................................................ 10 3.1. Data acquisition......................................................................................................... 12 3.1.1. Extraction, Cleansing and Transformation Tools............................................ 13 3.2. Data Storage and Access .......................................................................................... 13 3.3. Data Marts ................................................................................................................. 14 4 DESIGNING A DATA WAREHOUSE.......................................................................... 16 4.1. Beginning with Operational Data ............................................................................ 16 4.2. Data/Process Models ................................................................................................ 18 4.3. The DW Data Model ................................................................................................ 19 4.3.1. High-Level Modeling ........................................................................................ 19 4.3.2. Mid-Level Modeling ......................................................................................... 21 vii

4.3.3. Low-Level Modeling......................................................................................... 23 4.4. Database Design Methodology for DW .................................................................. 24 4.5. Conceptual Design Models ...................................................................................... 27 4.5.1. The Dimensional Fact Model............................................................................ 27 4.5.2. Multidimensional E/R Model ........................................................................... 30 4.5.3. starER ................................................................................................................. 33 4.5.4. Object-Oriented Multidimensional Model (OOMD) ...................................... 35 4.6. Logical Design Models............................................................................................. 36 4.6.1. Dimensional Model Design .............................................................................. 37 4.6.2. Flat Schema........................................................................................................ 39 4.6.3. Terraced Schema ............................................................................................... 40 4.6.4. Star Schema........................................................................................................ 41 4.6.5. Fact Constellation Schema................................................................................ 43 4.6.6. Galaxy Schema .................................................................................................. 43 4.6.7. Snowflake Schema ............................................................................................ 44 4.6.8. Star Cluster Schema .......................................................................................... 45 4.6.9. Starflake Schema ............................................................................................... 47 4.6.10. Cube.................................................................................................................. 48 4.7. Meta Data .................................................................................................................. 53 4.8. Materialized views .................................................................................................... 53 4.9. OLAP Server Architectures ..................................................................................... 54 5 COMPARISON OF MULTIDIMENSIONAL DESIGN MODELS............................. 56 5.1. Comparison of Dimensional Models and ER Models ............................................ 56 5.2. Comparison of Dimensional Models and Object-Oriented Models ...................... 57 5.3. Comparison of Conceptual Multidimensional Models........................................... 58 5.4. Comparison of Logical Design Models................................................................... 60 5.5. Discussion on Data Warehousing Design Tools..................................................... 61 6 IMPLEMENTING A DATA WAREHOUSE................................................................. 64 6.1. A Case Study............................................................................................................. 64 6.2. OOMD Approach...................................................................................................... 65 6.3. starER Approach ....................................................................................................... 68

viii

6.4. ME/R Approach ........................................................................................................ 70 6.5. DF Approach ............................................................................................................. 72 6.6. Implementation Details............................................................................................. 74 7 CONCLUSIONS AND FUTURE WORK...................................................................... 83 7.1. Contributions of the Thesis ...................................................................................... 85 7.2. Future Work .............................................................................................................. 86 REFERENCES ..................................................................................................................... 87

ix

LIST OF TABLES

TABLE 2.1 Comparison of OLTP and OLAP.................................................................................... 7 4.1 2-dimensional pivot view of an OLAP Table ............................................................. 49 4.2 3-dimensional pivot view of an OLAP Table ............................................................. 49 5.1 Comparison of ER, DM and OO methodologies ......................................................... 58 5.2 Comparison of conceptual design models .................................................................... 60 5.3 Comparison of logical design models........................................................................... 61

LIST OF FIGURES

FIGURE 2.1 Consolidation of OLTP information............................................................................... 4 2.2 Same attribute with different formats in different sources............................................ 4 2.3 Simple comparison of OLTP and DW systems ............................................................. 5 3.1 Architecture of DW........................................................................................................ 10 4.1 Data Extraction............................................................................................................... 16 4.2 Data Integration.............................................................................................................. 17 4.3 Same data, different usage............................................................................................. 17 4.4 A Simple ERD for a manufacturing environment........................................................ 20 4.5 Corporate ERD created by departmental ERDs........................................................... 20 4.6 Relationship between ERD and DIS............................................................................. 21 4.7 Midlevel model members .............................................................................................. 21 4.8 A Midlevel model sample.............................................................................................. 22 4.9 Corporate DIS formed by departmental DISs. ............................................................. 23 4.10 An example of a departmental DIS............................................................................. 23 4.11 Considerations in low-level modeling ........................................................................ 24 4.12 A dimensional fact schema sample............................................................................. 28 4.13 The graphical notation of ME/R elements.................................................................. 31 4.14 Multiple cubes sharing dimensions on different levels ............................................. 32 4.15 Combining ME/R notations with E/R......................................................................... 33 4.16 Notation used in starER.............................................................................................. 33 4.17 A sample DW model using starER ............................................................................. 35 4.18 Flat Schema ................................................................................................................. 40 4.19 Terraced Schema......................................................................................................... 41 4.20 Star Schema................................................................................................................. 42

xi

4.21 Fact Constellation Schema ......................................................................................... 43 4.22 Galaxy Schema............................................................................................................ 44 4.23 Snowflake Schema...................................................................................................... 45 4.24 Star Schema with fork .............................................................................................. 46 4.25 Star Cluster Schema.................................................................................................... 47 4.26 Starflake Schema......................................................................................................... 47 4.27 Comparison of schemas............................................................................................... 48 4.28 3-D Realization of a Cube ........................................................................................... 50 4.29 Operations on a Cube................................................................................................... 52 6.1 ER model of sales and shipping systems...................................................................... 65 6.2 Use case diagram of sales and shipping system........................................................... 66 6.3 Statechart diagram of sales and shipping system......................................................... 67 6.4 Static structure diagram of sales and shipping system ................................................ 67 6.5 Sales subsystem starER model..................................................................................... 69 6.6 Shipping subsystem starER model................................................................................ 70 6.7 Sales subsystem ME/R model....................................................................................... 71 6.8 Shipping subsystem ME/R model................................................................................. 72 6.9 Sales subsystem DF model............................................................................................ 73 6.10 Shipping subsystem DF model.................................................................................... 73 6.11 Snowflake schema for the sales subsystem............................................................... 74 6.12 Snowflake schema for the shipping subsystem......................................................... 75 6.13 General architecture of the case study........................................................................ 75 6.14 Sales DTS Package ...................................................................................................... 77 6.15 Shipping DTS Package ................................................................................................ 77 6.16 Transformation details for delimited text file ........................................................... 78 6.17 Transact-SQL query as the transformation source.................................................... 79 6.18 Pivot Chart using Excel as client ............................................................................... 80 6.19 Pivot Table using Excel as client ............................................................................... 80 6.20 Data Analyzer as client............................................................................................... 81

xii

LIST OF ABBREVIATIONS

3GL 4GL DAG DB DBMS DDM DF DIS DSS DTS DW ER ERD ETL HOLAP I/O IT ME/R MOLAP ODBC OID OLAP OLTP OO

Third Generation Language Fourth Generation Language Directed Acyclic Graph Database Database Management Systems Data Dimensional Modeling Dimensional Fact Data Item Set Decision Support System Data Transformation Services Data Warehouse Entity Relationship Entity Relationship Diagram Extract, Transform, Load Hybrid OLAP Input/Output Information Technology Multidimensional E/R Multidimensional OLAP Open Database Connectivity Object Identifier Online Analytical Processing Online Transaction Processing Object Oriented

xiii

OOMD RDBMS ROLAP SQL UML XML

Object Oriented Multidimensional Relational Database Management Systems Relational OLAP Structured Query Language Unified Modeling Language Extensible Markup Language

xiv

CHAPTER 1

INTRODUCTION

Information is an asset that provides benefit and competitive advantage to any organization. Today, every corporation have a relational database management system that is used for organizations daily operations. The companies desire to increase the value of their organizational data by turning it into actionable information. As the amount of the organizational data increases, it becomes harder to access and get the most information out of it, because it is in different formats, exists on different platforms and resides on different structures. Organizations have to write and maintain several programs to consolidate data for analysis and reporting. Also, the corporate decisionmakers require access to all the organizations data at any level, which may mean modifications on existing or development of new consolidation programs. This process would be costly, inefficient and time consuming for an organization. Data warehousing provides an excellent approach in transforming operational data into useful and reliable information to support the decision making process and also provides the basis for data analysis techniques like data mining and multidimensional analysis. Data warehousing process contains extraction of data from heterogenous data sources, cleaning, filtering and transforming data into a common structure and storing data in a structure that is easily accessed and used for reporting and analysis purposes.

As the need for building an organizational data warehouse is clear, now the question is how. There are generally accepted design methodologies in designing and implementing a data warehouse. The focus of this thesis is discussing the data warehouse conceptual and logical design models and comparing these approaches. 1.1. Scope and outline of the thesis The thesis organized as follows: Chapter 2 presents an overview of data warehouse concepts and makes a comparison between operational and analytical processing systems. Chapter 3 provides information on data warehousing fundamentals and process. Chapter 4 gives information on data warehouse design approaches used in conceptual and logical design phases. In chapter 5, the design approaches described in chapter 4 are discussed and compared. Finally in chapter 6, a sample conceptual model is logically implemented using the logical design models and the physical implementation of a data warehouse is described.

CHAPTER 2

DATA WAREHOUSE CONCEPTS

2.1. Definition of Data War ehouse A data warehouse (DW) refers to a database that is different from the organizations Online Transaction Processing (OLTP) database and that is used for the analysis of consolidated historical data. According to Barry Devlin, IBM Consultant, a DW is simply a single, complete and consistent store of data obtained from a variety of sources and made available to end users in a way they can understand and use it in a business context [1, 3]. According to W.H. Inmon, a DW is a subject-or iented , integr ated , time-var iant , and nonvolatile collection of data in support of managements decision making process [1, 2, 3, 6, 10, 11]. The description of the four key features of the DW is given below. Subject-or iented: In general, an enterprise contains information that is very detailed to meet all requirements needed for related subsets of the organization (sales dept, human resources dept, marketing dept etc.) and optimized for transaction processing. Usually, this type of data is not suitable for decision-makers to use. Decision-makers need subject-oriented data. DW should include only key business information. The data in the warehouse should be organized based on subject and only subject-oriented data should be moved into a warehouse. 3

If the decision-maker needs to find all information about a spesific product, he/she would need to use all systems like rental sales system, order sales system and catalog sales system, which is not the preferable and the practical way. Instead, all the key information must be consolidated in a warehouse and organized into subject areas as illustrated in Figure 2.1.

Figur e 2.1 Consolidation of OLTP infor mation Integr ated: DW is an architecture constructed by integrating data from multiple heterogeneous sources (like relational database (DB), flat files, excel sheets, XML data, data from the legacy systems) to support structured and/or ad hoc queries, analytical reporting and decision making. DW also provides mechanisms for cleaning and standardizing data. Figure 2.2 emphasizes various uses and formats of Product Code attribute.

Figur e 2.2 Same attr ibute with differ ent for mats in differ ent sour ces Time-var ian t: DW provides information from a historical prospective. Every key structure in the DW contains, either implicitly or explicitly, an element of time. A DW generally stores data that is 5-10 years old, to be used for comparisons, trends and forecasting. Nonvolatile: Data in the warehouse are not updated or changed (see Figure 2.3), so it does not require transaction processing, recovery and concurrency control mechanisms. 4

The operations needed in the DW are initial loading of data and access of data and refresh.

Figur e 2.3 Simple compar ison of OL TP and DW systems Some of the DW characteristics are given below; It is a database that is maintained separately from organizations operational databases. It allows for integration of various application systems. It supports information processing by consolidating historical data. User interface aimed at decision-makers. It contains large amount of data. It is updated infrequently but periodically updates are required to keep the warehouse meaningful and dynamic. It is subject-oriented. It is non-volatile. Data is longer-lived. Transaction systems may retain data only until processing is complete, whereas data warehouses may retain data for years. Data is stored in a format that is structured for querying and analysis. Data is summarized. DWs usually do not keep as much detail as transactionoriented systems. 2.2. Why OLAP systems must r un with OLTP In this section, I aim to make a comparison of OLTP and Online Analytical Processing (OLAP) systems and explain the reasons why an OLAP system is needed.

The nature of OLTP and OLAP systems are completely different both in technical and in business needs. The following table compares OLTP systems OLAP systems in main technical topics

OLTP User and System Or ientation Thousands users, customeroriented, used for transactions and querying clerks, clients and Information Technology (IT) professionals Data Contents Manages current data, very detail-oriented

OLAP Hundreds users, marketoriented, used for data analysis by knowledge workers

Manages large amounts of historical data, provides facilities for summarization and aggregation, stores information at different levels of granularity to support decision making process

Data is continuously updated Data is volatile and normalized (EntityRelationship (ER) Model) Database Design

Data is refreshed

Data is non-volatile and denormalized (Dimensional Model)

Adopts an ER model and an Adopts star, snowflake, or application-oriented database design, index/hash on primary key. fact constellation model and a subject-oriented database design, lots of scans.

View

Focuses on the current data within an enterprise or department, detailed, flat relational.

Spans multiple versions of a database schema due to the evolutionary process of an organization; integrates information from many organizational locations and data stores, summarized, multidimensional.

Access Patter ns

Short, atomic, simple transactions; requires concurrency control and recovery mechanisms, day to day activities, mostly updates.

Mostly read-only operations, complex query, although many could be complex queries, long term informational requirements, decision support, mostly reads.

Table 2.1 Comp ar ison of OLTP and OLAP It may seem questionable to implement a DW system for companies running their business on OLTP systems. The following list compares the main reasons for using a DW; To gain high performance of both systems by proper data organization (DB design). OLTP deals with transactions, concurrency, locking and logging. OLTP deals with many records at ones. Transaction performance of OLTP and selection performance of OLAP would be in conflict. Different structures, contents, and uses of the data. OLTP requires the current data. OLAP requires historical data. Data Cleanness. Data in OLTP might be dirty because it is collected by clerks that may make mistakes and for other reasons. Data that goes into OLAP should be cleaned and standardized.

OLTP and OLAP systems need to run different types of queries. They may provide different functionality and use different types on queries.

The main roles in a company that will use a DW solution are [4]; Top executives and decision makers Middle/operational managers Knowledge workers Non-technical business related individuals

The main advantages of using a DW solution are summarized in the list below [2, 3, 6]; High query performance Does not interfere with local processing at sources Information copied at warehouse (can modify, summarize, restructure, etc.) Potential high Return on Investment Competitive advantage Increase productivity of corporate decision makers

As discussed above, a DW solution has many advantages and benefits to an organization. Also implementing a DW solution solves some business problems, it may bring some new self-owned problems mentioned below [2, 6];

Underestimation of resources for data loading Hidden problems with source systems Required data not captured Increased end-user demands High maintenance Long duration projects Complexity of integration Data homogenization High demand for resources Data ownership

2.3. Requir ements for Data War ehouse Database Management Systems In the implementation of a DW solution, many technical points must be considered. While an OLTP database management systems (DBMS) must only consider transaction 8

processing performance (which is basically; a transaction must be completed in the minimum time; without deadlocks; and with support of thousands of transactions per second) The relational DBMS (RDBMS) suitable for data warehousing has the following requirements [6]; Load per for m ance : Data warehouses need incremental loading of data periodically so the load process performance should be like gigabytes of data per hour. Load pr ocessing: Data conversion, filtering, indexing and reformatting may be necessary during loading data into the data warehouse. This process should be executed as a single unit of work. Data quality m anagement: The warehouse must ensure consistency and referential integrity despite various data sources and big data size. The measure of success for a data warehouse is the ability to satisfy business needs. Quer y Per for mance: Complex queries must complete in acceptable periods. Ter abyte scalability: The data warehouse RDBMS should not have any

database size limitations and should provide recovery mechanisms. Mass user scalability: The data warehouse RDBMS should be able to support hundreds of concurrent users. War ehouse administr ation: Easy-to-use and flexible administrative tools should exists for data warehouse administration. Advanced quer y functionality: The data warehouse RDBMS should supply advanced analytical operations to enable end-users perform advanced calculations and analysis.

CHAPTER 3

FUNDAMENTALS OF DATA WAREHOUSE

The main reason for building a DW is to improve the quality of information in the organization. Data coming from both internal and external sources in various formats and structures is consolidated and integrated into a single repository. DW system comprises the data warehouse and all components used for building, accessing and maintaining the data warehouse.

Figur e 3.1 Ar chitectur e of DW

10

A general architecture of a DW is given in Figure 3.1 and the main components are described below [5, 32]. The data import and preparation component is responsible for data acquisition. It includes all programs (like Data Transformation Services (DTS)) that are responsible for extracting data from operational sources, preparing and loading it into the warehouse. The access component includes all applications (like OLAP) that use the information stored in the warehouse. Additionally, a metadata management component is responsible for the management, definition and access of all different types of metadata. Metadata is defined as data describing the meaning of data. In data warehousing, there are various types of metadata, e.g., information about the operational sources, the structure and semantics of the data warehouse data, the tasks performed during the construction, the maintenance and access of a data warehouse, etc. Implementing a DW is a complex task containing two major phases. In the configuration phase, a conceptual view of the warehouse is first specified according to user requirements (DW design). Then, the related data sources and the Extraction-LoadTransform (ETL) process (data acquisition) are determined. Finally, decisions about persistent storage of the warehouse using database technology and the various ways data will be accessed during analysis are made. After the initial load (the first load of the DW according to the configuration), during the operation phase, warehouse data must be regularly refreshed, i.e., modifications of operational data since the last DW refreshment must be propagated into the warehouse such that data stored in the data warehouse reflect the state of the underlying operational systems. A more natural way to consider multidimensionality of warehouse data is provided by the multidimensional data model. In this model, the data cube is the basic modeling construct. Operations like pivoting (rotate the cube), slicing-dicing (select a subset of the cube), roll-up and drill-down (increasing and decreasing the level of aggregation) can be applied to a data cube. For the implementation of multidimensional databases, there are two main approaches. In the first approach, extended RDBMSs, called relational OLAP 11

(ROLAP) servers, use a relational database to implement the multidimensional model and operations. ROLAP servers provide SQL extensions and translate data cube operations to relational queries. In the second approach, multidimensional OLAP (MOLAP) servers store multidimensional data in non-relational specialized storage structures. These systems usually precompute the results of complex operations (during storage structure building) in order to increase performance. 3.1. Data acquisition Data extraction is one of the most time-consuming tasks of DW development. Data consolidated from heterogenous systems may have problems, and may need to be first transformed and cleaned before loaded into the DW. Data gathered from operational systems may be incorrect, inconsistent, unreadable or incomplete. Data cleaning is an essential task in data warehousing process in order to get correct and qualitative data into the DW. This process contains basically the following tasks: [5] converting data from heterogenous data sources with various external representations into a common structure suitable for the DW identifying and eliminating redundant or irrelevant data transforming data to correct values (e.g., by looking up parameter usage and consolidating these values into a common format) reconciling differences between multiple sources, due to the use of homonyms (same name for different things), synonyms (different names for same things) or different units of measurement As the cleaning process is completed, the data that will be stored in the warehouse must be merged and set into a common detail level containing time related information to enable usage of historical data. Before loading data into the DW, tasks like filtering, sorting, partitioning and indexing may need to be performed. After these processes, the consolidated data may be imported into the DW using one of bulk data loaders, a custom application or an import/export wizard provided by the DBMS administration applications.

12

3.1.1. Extr action, Cleansing and Tr ansfor mation Tools The tasks of capturing data from a source system, cleansing, and transforming the data and loading the consolidated data into a target system can be done either by separate products or by a single integrated solution. Integrated solutions fall into one of the following categories [6]: Code generators Database data replication tools Dynamic transformation engines

There are solutions that fulfill all of the requirements mentioned above. One of these products is Microsoft Data Transformation Services is described in chapter 6. Code gener ator s Code generators create customized 3GL, 4GL transformation programs based on source and target data definitions. The main issue with this approach is the management of the large number of programs required to support a complex corporate DW. Database data r eplication tools Database data replication tools employ database triggers or a recovery log to capture changes to a single data source on one system and apply the changes to a copy of the source data located on a different system. Most replication products dont support the capture of changes to non-relational files and databases and often not provide facilities for significant data transformation and enhancement. These tools can be used to rebuild a database following failure or to create a database for a data mart, provided that the number of data sources is small and the level of data transformation is relatively simple. Dynamic tr ansfor mation engines Rule-driven dynamic transformation engines capture data from a source system at userdefined intervals, transform the data and then send and load the results into a target environment. Most products support only relational data sources, but products are now emerging that handle non-relational source files and databases. 3.2. Data Stor age and Access Because of the special nature of warehouse data and access, accustomed mechanisms for data storage, query processing and transaction management must be

13

adapted. DW solutions need complex querying requirements and operations involving large volumes of data access. These operations need special access methods, storage structures and query processing techniques. The storage approaches of a DW is described in detail in section 4.9. One of these physical storage methods may be chosen concerning the trade-off between query performance and amount of data. Once the DW is available for end-users, there are a variety of techniques to enable end-users access the DW data for analysis and reporting. There are several tools and products that are commercially available. In common all client tools use generally OLEDB, ODBC or native client providers to access the DW data. The most commercially used client application is Microsoft Excel with pivot tables. A company that makes business in several countries througout the world may need to analyse regional trends and my need to compete in regions. A centric DW may not be feasible for these companies. These organizations may need to establish data marts which are selected parts of the DW that support specific decision support application requirements of a companys department or geographical region. Data marts usually contain simple replicas of warehouse partitions or data that has been further summarized or derived from base warehouse data. Data marts allow the efficient execution of predicted queries over a significantly smaller database. 3.3. Data Mar ts A data mart is a subset of the data in a DW and is summary data relating to a department or a specific function [6]. Data marts focus on the requirements of users in a particular department or business function of an organization. Since data marts are specialized for departmental operations, they contain less data and the end-users are much capable of exploiting data marts than DWs. The main reasons for implementing a data mart instead of a DW may be summarized as follows: Data marts enable end-users to analyze the data they need most often in their daily operations.

14

Since data marts contain less data, the end-user response time in queries is much quicker. Data marts are more specialized and contain less data, therefore data transformation and integration tasks are much faster in data marts than DWs and setting up a data mart is a simpler and a cheaper task compared to establishing an organizational DW in terms of time and resources.

In terms of software engineering, building a data mart may be a more feasible project than building a DW, because the requirements of building a data mart are much more explicit than a corporate wide DW project.

Although data marts seem to have advantages over DWs, there are some issues that must be addressed about data marts. Size: Although data marts are considered to be smaller than data warehouses, size and complexity of some data marts may match a small corporate DW. As the size of a data mart increases, it is likely to have a performance decrease. Load per for mance: Both end-user response time and data loading performance are critical tasks of data marts. For increasing the response time, data marts usually contain lots of summary tables and aggregations which have a negative effect on load performance. User access to data in multiple data mar ts: A solution to this problem is building virtual data marts which are views of several physical data marts. Administr ation: With the increase in number of data marts, the management need arises to coordinate data mart activities such as versioning, consistency, integrity, security and performance tuning.

15

CHAPTER 4

DESIGNING A DATA WAREHOUSE

Designing a warehouse means to complete all the requirements mentioned in section 2.3 and obviously is a complicated process. There are two major components to build a DW; the design of the interface from operational systems and the design of the DW [11]. DW design is different from a classical requirements-driven systems design. 4.1. Beginning with Oper ational Data Creating the DW does not only involve extracting operational data and entering it into the warehouse (Figure 4.1) .

Figur e 4.1 Data Extr action

16

Pulling the data into the DW without integrating it is a big mistake ( Figure 4.2 ).

Figur e 4.2 Data Integr ation Existing applications were designed with their own requirements and integration with other applications was not concerned much. These results in data redundancy, i.e. same data may exist in other applications with same meaning, with different name or with different measure ( Figure 4.3 ).

Figur e 4.3 Same data, differ ent usage Another problem is the performance of accessing existing systems data. The existing systems environment holds gigabytes and perhaps terabytes of data, and attempting to scan all of it every time a DW load needs to be done is resource and time consuming and unrealistic. Three types of data are loaded into the DW from the operational system: Archival data Data currently contained in the operational environment

17

Changes to the DW environment from the updates that have occurred in the operational system since the last refresh

Five common techniques are used to limit the amount of operational data scanned to refresh the DW. Scan data that has been timestamped in the operational environment. Scan a 'delta' file. A delta file contains only the changes made to an application as a result of the transactions that have run through the operational environment. Scan a log file or an audit file created by the transaction processing system. A log file contains the same data as a delta file. Modify application code. Rubbing a 'before' and an 'after' image of the operational file together.

Another difficulty is that operational data must undergo a time-basis shift as it passes into the DW. The operational datas accuracy is valid at the instant it is accessed, after that it may be updated. However when the data is loaded into the warehouse, it cannot be updated anymore, so a time element must be attached to it. Another problem when passing data is the need to manage the volume of data that resides in and passes into the warehouse. Volume of data in the DW will grow fast. 4.2. Data/Pr ocess Models The process model applies only to the operational environment. The data model applies to both the operational environment and the DW environment. A process model consists: Functional decomposition Context-level zero diagram Data flow diagram Structure chart State transition diagram Hierarchical input process output(HIPO) chart Pseudo code

18

A process model is invaluable, for instance, when building the data mart. The process model is requirements-based; it is not suitable for the DW. The data model is applicable to both the existing systems environment and the DW environment. An overall corporate data model has been constructed with no regard for a distinction between existing operational systems and the DW. The corporate data model focuses on only primitive data. Performance factors are added into the corporate data model as the model is transported to the existing systems environment. Although few changes are made to the corporate data model for operational environment, more changes are made to the corporate data model to use in DW environment. First, data that is used purely in the operational environment is removed. Next, the key structures of the corporate data model are enhanced with an element of time. Derived data is added to the corporate data model where the derived data is publicly used and calculated once, not repeatedly. Finally, data relationships in the operational environment are turned into artifacts in the DW. A final design activity in transforming the corporate data model to the data warehouse data model is to perform stability analysis. Stability analysis involves grouping attributes of data together based on their tendency for change. 4.3. The DW Data Model There are three levels in data modeling process: high-level modeling (called the ERD, entity relationship level), midlevel modeling (called the data item set, or DIS), and low-level modeling (called the physical model). 4.3.1. High-Level Modeling The high level of modeling features entities and relationships. The name of the entity is surrounded by an oval. Relationships among entities are depicted with arrows. The direction and number of the arrowheads indicate the cardinality of the relationship, and only direct relationships are indicated.

19

Figur e 4.4 A Simple ERD for a manufactur ing envir onment The entities that are shown in the ERD level (see Figure 4.4) are at the highest level of abstraction. The corporate ERD as shown in Figure 4.5 is formed of many individual ERDs that reflect the different views of people across the corporation. Separate high-level data models have been created for different communities within the corporation. Collectively, they make up the corporate ERD.

Figur e 4.5 Cor p or ate ERD cr eated b y depar tmental ERDs

20

4.3.2. Mid-Level Modeling After the high-level data model is created, the next level is establishedthe midlevel model or the DIS. For each major subject area, or entity, identified in the highlevel data model, a midlevel model is created. Each area is subsequently developed into its own midlevel model (see Figure 4.6)

Figur e 4.6 Relationship between ERD and DIS Four basic constructs are found at the midlevel model (also shown in Figure 4.7): A primary grouping of data A secondary grouping of data A connector, suggesting the relationships of data between major subject areas Type of data

Figur e 4.7 Midlevel model members

21

The primary grouping exists once, and only once, for each major subject area. It holds attributes that exist only once for each major subject area. As with all groupings of data, the primary grouping contains attributes and keys for each major subject area. The secondary grouping holds data attributes that can exist multiple times for each major subject area. This grouping is indicated by a line drawn downward from the primary grouping of data. There may be as many secondary groupings as there are distinct groups of data that can occur multiple times. The third construct is the connector. The connector relates data from one grouping to another. A relationship identified at the ERD level results in an acknowledgement at the DIS level. The convention used to indicate a connector is an underlining of a foreign key. The fourth construct in the data model is type of data. Type of data is indicated by a line leading to the right of a grouping of data. The grouping of data to the left is the supertype. The grouping of data to the right is the subtype of data. These four data modeling constructs are used to identify the attributes of data in a data model and the relationship among those attributes. When a relationship is identified at the ERD level, it is manifested by a pair of connector relationships at the DIS level. A sample model is drawn in Figure 4.8 below.

Figur e 4.8 A Midlevel model sample Like the corporate ERD that is created from different ERDs reflecting the community of users, the corporate DIS is created from multiple DISs. Figure 4.9 shows a sample corporate DIS formed by many departments DISs.

22

Figur e 4.9 Cor p or ate DIS for med by depar tmental DISs. Figure 4.10 shows an individual departments DIS.

Figur e 4.10 An example of a depar tmental DIS 4.3.3. Low-Level Modeling The physical data model is created from the midlevel data model just by extending the midlevel data model to include keys and physical characteristics of the model. At this point, the physical data model looks like a series of tables, sometimes called

23

relational tables. With the DW, the first step in doing so is deciding on the granularity and partitioning of the data. After granularity and partitioning are factored in, a variety of other physical design activities are embedded into the design. At the heart of the physical design considerations is the usage of physical input/output (I/O). Physical I/O is the activity that brings data into the computer from storage or sends data to storage from the computer. The job of the DW designer is to organize data physically for the return of the maximum number of records from the execution of a physical I/O. Figure 4.11 illustrate the major considerations in low-level modeling.

Figur e 4.11 Consider ations in low-level modeling There is another mitigating factor regarding physical placement of data in the data warehouse: Data in the warehouse normally is not updated. This frees the designer to use physical design techniques that otherwise would not be acceptable if it were regularly updated. 4.4. Database Design Methodology for DW In the next few sections of this thesis I will be discussing both conceptual and logical design methods of data warehousing. Adopting the terminology of [23, 36, 37, 38] three different design phases are distinguished; conceptual design manages concepts that are close to the way users perceive data; logical design deals with concepts related to a certain kind of DBMS; physical design depends on the specific DBMS and describes how data is actually stored [35, 40].

24

Prior to beginning the discussion, the basic concepts of dimensional modeling should be mentioned which are: facts, dimensions and measures [7, 24]. A fact is a collection of related data items, consisting of measures and context data. It typically represents business items or business transactions. A dimension is a collection of data that describe one business dimension. Dimensions determine the contextual background for the facts; they are the parameters over which we want to perform OLAP. A measure is a numeric attribute of a fact, representing the performance or behavior of the business relative to the dimensions. Before this discussion, I also prefer to summarize the methodology proposed by Kimball [21], who is accepted as a guru on data warehousing and whose studies have encouraged many academicians on the study of data warehousing. The nine step methodology by Kimball is as follows[6, 42, 43]: 1. Choosing the process: The process (function) refers to the subject matter of a particular data mart. The first data mart to be built should be the one that is most likely to be delivered on time with in budget and to answer the most important business question. 2. Choosing the grain: This means deciding exactly what a fact table record represents. Only when the grain for the fact table is chosen can we identify the dimensions of the fact table. The grain decision for the fact table also determines the grain of each of the dimension tables. 3. Identifying and conforming the dimensions: Dimensions set the context for asking questions about the facts in the fact table. A well-built set of dimensions makes the data mart understandable and easy to use. A poorly presented or incomplete set of dimensions will reduce the usefulness of a data mart to an enterprise. When a dimension is used in more than one data mart, the dimension is referred to as being conformed. 4. Choosing the facts : The grain of the fact table determines which facts can be used in the data mart. All the facts must be expressed at the level implied by the grain. The facts should be numeric and additive. Additional facts can be added to a fact table at any time provided they are consistent with the grain of the table. 25

5. Storing pre-calculations in the fact table : Once the facts have been selected each should be re-examined to determine whether there are opportunities to use precalculations. 6. Rounding out the dimension tables : We return to the dimension tables and add as much text description to the dimensions. The text descriptions should be as intuitive and understandable to the users. The usefulness of a data mart is determined by the scope and nature of the attributes of the dimension tables. 7. Choosing the duration of the database: The duration measures how far back in time the fact table goes. There is requirement to look at the same time period a year or two earlier. Very large fact tables raise at least two very significant DW design issues. First, it is often increasingly difficult to source increasingly old data. The older data, the more likely there will be more problems in reading and interpreting the old files or the old tapes. Second, it is mandatory that the old versions of the important dimensions be used, not the most current versions. This is known as the slowly changing dimension problem. 8. Tracking slowly changing dimensions: There are three basic types of slowly changing dimensions: o Type1: where a changed dimension attribute is overwritten, o Type2: where a changed dimension attribute causes a new dimension record to be created, o Type3: a changed dimension attribute causes an alternate attribute to be created so that both the old and new values of the attribute are simultaneously accessible in the same dimension record. 9. Deciding the query priorities and the query modes: We consider physical design issues. The most critical physical design issues affecting the end-users perception of the data mart are physical sort order of the fact table on disk and the presence of pre-stored summaries or aggregations. There are additional physical design issues affecting administration, backup, indexing performance, and security.We have a design for data mart that supports the requirements of a particular business process and also allows the easy integration with other related data marts to ultimately form the enterprise-wide DW.

26

4.5. Conceptual Design Models The main goal of conceptual design modeling is developing a formal, complete, abstract design based on the user requirements [34]. At this phase of a DW there is the need to: Represent facts and their properties: Facts properties are usually numerical and can be summarized (aggregated). Connect the dimension to facts: Time is always associated to a fact. Represent objects and capture their properties with the associations among them: Object properties (summary properties) can be numeric. Additionally there are three special types of associations; specialization/generalization (showing objects as subclasses of other objects), aggregation (showing objects as parts of a layer object), membership (showing that an object is a member of another higher object class with the same characteristics and behavior). Strict membership (or not) (all members belong to only one higher object class), Complete membership (or not) (all members belong to one higher object class and that object class is consisted by those members only). Record the associations between objects and facts: Facts are connected to objects. Distinguish dimensions and categorize them into hierarchies: dimensions governed by associations of type membership forming hierarchies that specify different granularities. 4.5.1. The Dimensional Fact Model This model is built from ER schemas [9, 15, 16, 17, 33]. The Dimensional Fact (DF) Model is a collection of tree structured fact schemas whose elements are facts, attributes, dimensions and hierarchies. Fact attributes additivity, optional dimension attributes and non-dimension attributes existence may also be represented on fact schemas. Compatible fact schemas may be overlapped in order to relate and compare data. A fact schema is structured as a tree whose root is a fact. The fact is represented by a box which reports the fact name. 27

Figur e 4.12 A dimensional fact schema sample Sub-trees rooted in dimensions are hierarchies. The circles represent the attributes and the arcs represent relationship between attribute pairs. The non-dimension attributes (address attribute as shown in Figure 4.12) are represented by lines instead of circles. A non-dimension attribute contains additional information about an attribute of the hierarchy, is connected to it by a -to-one relationship and cannot be used for aggregation. The arcs represented by dashes express optional relationships between pairs of attributes. A fact expresses a many-to-many relationship among the dimensions. Each combination of values of the dimensions defines a fact instance, one value for each fact attribute. Most attributes are additive along all dimensions. This means that the sum operator can be used to aggregate attribute values along all hierarchies. A fact attribute is called semi-additive if it is not additive along one or more dimensions, non-additive if it is additive along no dimension. DF model consists of 5 steps; Defining facts (a fact may be represented on the E/R schema either by an entity F or by an n-ary relationships between entities E1 to En). For each fact; o Building the attribute tree. (Each vertex corresponds to an attribute of the schema; the root corresponds to the identifier of F; for each vertex v, the corresponding attribute functionally determines all the attributes corresponding to the descendants of v. If F is identified by the combination of two or more attributes, identifier (F) denotes their 28

concatenation. It is worth adding some further notes: It is useful to emphasize on the fact schema the existence of optional relationships between attributes in a hierarchy. Optional relationships or optional attributes of the E/R schema should be marked by a dash; A one-to-one relationship can be thought of as a particular kind of many-to-one relationship, hence, it can be inserted into the attribute tree; Generalization hierarchies in the E/R schema are equivalent to one-to-one relationships between the super-entity and each sub-entity; x-to-many relationships cannot be inserted into the attribute tree. In fact, representing these relationships at the logical level, for instance by a star schema, would be impossible without violating the first normal form; an n-ary relationship is equivalent to n binary relationships. Most n-ary relationships have maximum multiplicity greater than 1 on all their branches; they determine n one-to-many binary relationships which cannot be inserted into the attribute tree.) o Pruning and grafting the attribute tree (Not all of the attributes represented in the attribute tree are interesting for the DW. Thus, the attribute tree may be pruned and grafted in order to eliminate the unnecessary levels of detail. Pruning is carried out by dropping any subtree from the tree. The attributes dropped will not be included in the fact schema, hence, it will be impossible to use them to aggregate data. Grafting is used when its descendants must be preserved.). o Defining dimensions (The dimensions must be chosen in the attribute tree among the children vertices of the root. E/R schemas can be classified as snapshot and temporal. A snapshot schema describes the current state of the application domain; old versions of data varying over time are continuously replaced by new versions. A temporal schema describes the evolution of the application domain over a range of time; old versions of data are explicitly represented and stored. When designing a DW from a temporal schema, time is explicitly represented as an E/R attribute and thus it is an obvious candidate to define a dimension. Time is not

29

explicitly represented however, should be added as a dimension to the fact schema). o Defining fact attributes (Fact attributes are typically either counts of the number of instances of F, or the sum/average/maximum/minimum of expressions involving numerical attributes of the attribute tree. A fact may have no attributes, if the only information to be recorded is the occurrence of the fact.). o Defining hierarchies (Along each hierarchy, attributes must be arranged into a tree such that an x-to-one relationship holds between each node and its descendants. It is still possible to prune and graft the tree in order to eliminate irrelevant details. It is also possible to add new levels of aggregation by defining ranges for numerical attributes. During this phase, the attributes which should not be used for aggregation but only for informative purposes may be identified as non-dimension attributes.). 4.5.2. Multidimensional E/R Model It is argued that ER approach is not suited for multidimensional conceptual modeling because the semantics of the main characteristics of the model cannot be effectively represented. Multidimensional E/R (ME/R) model includes some key considerations [14]: Specialization of the ER Model Minimal extension of the ER Model; this model should be easy to learn and use for an experienced ER Modeler. There are few additional elements. Representation of the multidimensional aspects; despite the minimality, the specialization should be powerful enough to express the basic multidimensional aspects, namely the qualifying and quantifying data and the hierarchical structure of the qualifying data. This model allows the generalization concepts. There are some specializations: A special entity set: dimension level Two special relationship sets connecting dimension levels: o a special n-ary relationship set: the fact relationship set

30

o a special binary relationship set: the roll-up to relationship set The roll-up to relationship set; it relates a dimension level A to a dimension level B representing concepts of a higher level of abstraction (city roll-up to country). The fact relationship set is a specialization of a general n-ary relationship set. It connects n different dimension level entities. The fact relationship set models the natural separation of qualifying and quantifying data. The attributes of the fact relationship set model the measures of the fact while dimension levels model the quantifying data. This model uses a special graphical notation which a sample notation is shown in Figure 4.13 .

Figur e 4.13 The gr aphical notation of ME/R elements Individual characteristics of ME/R model may be summarized as follows; A central element in the multidimensional model is the concept of dimensions that span the multidimensional space. The ME/R model does not contain an explicit counterpart for this idea. This is not necessary because a dimension consists of a set of dimension levels. The information which dimension-levels belong to a given dimension is included implicitly within the structure of the rolls-up graph. The hierarchical classification structure of the dimensions is expressed by dimension level entity sets and the roll-up relationships. The rolls-up relationship sets define a directed acyclic graph on the dimension levels. This enables the easy modeling of multiple hierarchies, alternative paths and shared hierarchy levels for different dimensions. Thus no redundant modeling of the shared levels is necessary. Dimension level attributes are modeled as attributes of dimension level entity sets. This allows a different attribute structure for each dimension level.

31

By modeling the multidimensional cube as a relationship set it is possible to include an arbitrary number of facts in the schema thus representing a multicube model. Remarkably the schema also contains information about the granularity level on which the dimensions are shared.

Concerning measures and their structure, the ME/R model allows record structured measures as multiple attributes for one fact relationship set. The semantic information that some of the measures are derived cannot be included in the model. Like the E/R model the ME/R model captures the static structure of the application domain. The calculation of measures is functional information and should not be included in the static model. An orthogonal functional model should capture these dependencies.

Schema contains rolls-up relationship between entities. Therefore levels of different dimensions may roll up to a common parent level. This information can be used to avoid redundancies.

This model is used is a relationship. ME/R and ER models notations can be used together.

Figure 4.14 shows multiple cubes that share dimensions on different levels.

Figur e 4.14 Multiple cubes shar ing dimensions on differ ent levels As mentioned above, the ME/R and ER model notations can be used together as illustrated in Figure 4.15.

32

Figur e 4.15 Combining ME/R notations with E/R 4.5.3. star ER This model combines star structure with constructs of ER model [13]. The starER contains facts, entities, relationships and attributes. This model has the following constructs: Fact set: represents a set of real world facts sharing the same characteristics or properties. It is always associated with time. It is represented as a circle. Entity set: represents a set of real world objects with similar properties. It is represented as a rectangle. Relationship set: represents a set of associations among entity sets or among entity sets and fact sets. Its cardinality can be many-to-many, many-to-one, oneto-many. It is represented as a diamond. Relationship sets among entity sets can be type of specialization/generalization, aggregation and membership. Figure 4.16 shows the notation for relationship set types.

Figur e 4.16 Notation used in star ER

33

Attribute: static properties of entity sets, relationship sets, fact sets. It is represented as an oval. Fact properties can be of type stock (S) (the state of something at a specific point in time), flow (F) (the commutative effect over a period of time for some parameter in the DW environment and which is always summarized) or valueper-unit (V) (measured for a fixed-time and the resulted measures are not summarized).

The following criteria are satisfied by the starER schema; Explicit hierarchies in dimensions Symmetric treatment of dimensions and summary attributes (properties) Multiple hierarchies in each dimension Support for correct summary or aggregation Support of non-strict hierarchies Support of many-to-many relationships between facts and dimensions Handling different levels of granularity at summary properties Handling uncertainty Handling change and time Relationships between dimensions and facts in starER arent only many-to-one, but also many-to-many, which allows for better understanding of the involved information. Object participating in the data warehouse, but not in the form of a dimension are allowed in the starER. Specialized relationships on dimensions are permitted

There following list shows the main differences between DF Schema and starER model;

(specialization/generalization, aggregation, membership) and represent more information. DF requires only a rather straight forward transformation to fact and dimension tables. This is an advantage of DF Schema. But this is not a drawback for the starER model, since well-known rules of how to transform an ER Schema (Which is the basic structural difference between the two approaches) to relations do exist. 34

starER model combines the powerful constructs of the ER model with the star schema.

A sample DW model using starER is illustrated in Figure 4.17.

Figur e 4.17 A sample DW model using star ER 4.5.4. Ob ject-Or iented Multidimensional Model (OOMD) Unified Modeling Language (UML) has been widely accepted as a standard objectoriented modeling language for software design. OOMD modeling approach is based on UML. In OOMD model, dimensions and facts are represented by dimension classes and fact classes [18, 19]. Fact classes are considered as composite classes in a sharedaggregation relationship of n-dimension classes. In this way, many-to-many relationships between facts and particular dimensions are represented by indicating 1..* cardinality on the dimension class. Fact classes cardinality is defined as * to indicate that a dimension object can be part of one, zero or more fact object instances. The minimum cardinality of dimension classes is defined as 1 to indicate that a fact object is always related to object instances from all dimensions. Derived measures are placed in fact class by notation after /. Derivation rules appear between braces. Identifying attribute can be defined in fact classes by notation after {OID} (Object Identifier). All measures are additive (Sum operator can be applied to aggregate measure values along all dimensions). For dimensions, every classification hierarchy level is specified by a class

35

(base class). An association of classes specifies the relationships between two levels of a classification hierarchy. These classes must define DAG (Directed Acyclic Graph) rooted in the dimension class. The DAG structure can represent both alternative path and multiple classification hierarchies. Descriptor attribute ({D}) define in every class that represents a classification hierarchy level. Strictness means that an object at a hierarchys lower level belongs to only one higher level object. Completeness means that all members belong to one higher-class object and that object consists of those members only. OOMD approach uses a generalization-specialization relationship to categorize entities that contain subtypes. Cube classes represent initial user requirements as the starting point for subsequent dataanalysis phase. Cube classes contain; Head area; contains the cube classs name. Measures area; contains the measures to be analyzed. Slice area; contains the constraints to be satisfied. Dice area; contains the dimensions and their grouping conditions to address the analysis. Cube operations; cover the OLAP operations for a further data analysis phase.

4.6. Logical Design Models DW logical design involves the definition of structures that enable an efficient access to information. The designer builds multidimensional structures considering the conceptual schema representing the information requirements, the source databases, and non functional (mainly performance) requirements. This phase also includes specifications for data extraction tools, data loading processes, and warehouse access methods. At the end of logical design phase, a working prototype should be created for the end-user. Dimensional models represent data with a cube structure, making more compatible logical data representation with OLAP data management. The objectives of dimensional modeling are [10]:

36

To produce database structures that are easy for end-users to understand and write queries against, To maximize the efficiency of queries.

It achieves these objectives by minimizing the number of tables and relationships between them. Normalized databases have some characteristics that are appropriate for OLTP systems, but not for DWs [7]: Its structure is not easy for end-users to understand and use. In OLTP systems this is not a problem because, usually end-users interact with the database through a layer of software. Data redundancy is minimized. This maximizes efficiency of updates, but tends to penalize retrievals. Data redundancy is not a problem in DWs because data is not updated on-line. Dimensionality modeling uses the ER Modeling with some important restrictions. Dimensional model composed of one table with a composite primary key, called fact table, and a set of smaller tables called dimension tables . Each dimension table has a simple (non-composite) primary key that corresponds exactly to one of the components of the composite key in the fact table. This characteristic structure is called star schema or star join . Another important feature, all natural keys are replaced with surrogate keys. This means that every join between fact and dimension tables is based on surrogate keys, not natural keys. Each surrogate key should have a generalized structure based on simple integers. The use of surrogate keys allows the data in the DW to have some independence from the data used and produced by the OLTP systems. 4.6.1. Dimensional Model Design This section describes a method for developing a dimensional model from an Entity Relationship model [12]. This data model is used by OLTP systems. It contains no redundancy, but high efficiency of updates, shows all data and relationships between them. Simple queries require multiple table joins and complex subqueries. It is suitable for technical specialist.

37

Classify Entities: For producing a dimensional model from ER model, first classify the entities into three categories. o Transaction Entities: These entities are the most important entities in a DW. They have highest precedence. They construct fact tables in star schema. These entities record details about particular events (orders, payments, etc.) that decision makers want to understand and analyze. There are some characteristics; It describes an event that occurs at a point in time. It contains measurements or quantities that may be summarized (sales amount, volumes) o Component Entities: These entities are directly related with a transaction entity with a one-to-many relationship. They have lowest precedence. They define the details or components of each transaction. They answer the who, what, when, where, how and why of event (customer, product, period, etc.). Time is an important component of any transaction. They construct dimension tables in star schema. o Classification Entities : These entities are related with component entities by a chain of one-to-many relationship. They are functionally dependent on a component entity. These entities represent hierarchies embedded in the data model, which may be collapsed in to component entity to form dimension tables in star schema.

Identify Hierarchies: Most dimension tables in star schema include embedded hierarchies. A hierarchy is called maximal if it cannot be extended upwards or downwards by including another entity. An entity is called minimal if it has no one-to-many relationship. An entity is called maximal if it has no many-to-one relationship.

Produce Dimensional Models: There are two operators to produce dimensional models from ER. o Collapse Hierarchy: Higher level entities can be collapsed into lower level entities within hierarchies. Collapsing a hierarchy is a form of denormalization. This increases redundancy in the form of a transitive

38

dependency, which is a violation to 3NF. We can continue doing this until we reach the bottom of the hierarchy and end up with a single table. o Aggregation: This operator can be applied to a transaction entity to create a new entity containing summarized data. There are 8 models used in dimensional modeling [6, 12]: Flat Schema Terraced Schema Star Schema Fact Constellation Schema Galaxy Schema Snowflake Schema Star Cluster Schema Starflake Schema

4.6.2. Flat Schema This schema is the simplest schema. This is formed by collapsing all entities in the data model down into the minimal entities. This minimizes the number of tables in the database and joins in the queries. We end up with one table for each minimal entity in the original data model [12]. This structure does not lose information from the original data model. It contains redundancy, in the form of transitive and partial dependencies, but does not involve any aggregation. It contains some problems; first it may lead to aggregation errors when there are hierarchical relationships between transaction entities. When we collapse numerical amounts from higher level transaction entities in to other they will be repeated. Second this schema contains large number of attributes. Therefore while the number of tables (system complexity) is minimized, the complexity of each table (element complexity) is increased. Figure 4.18 shows a sample flat schema.

39

Figur e 4.18 Flat Schema 4.6.3. Ter r aced Schema This schema is formed by collapsing entities down maximal hierarchies, end with when they reach a transaction entity. This results in a single table for each transaction entity in the data model. It causes some problems for inexperienced user, because the separation between levels of transaction entities is explicitly shown [12]. The Figure 4.19 illustrates a sample terraced schema.

40

Figur e 4.19 Ter r aced Schem a 4.6.4. Star Schema It is the basic structure for a dimensional model. It has one fact table and a set of smaller dimension tables arranged around the fact table. The fact data will not change over time. The most useful fact tables are numeric and additive because data warehouse applications almost never access a single record. They access hundreds, thousands, millions of records at a time and aggregate them. The fact table is linked to all the dimension tables by one to many relationships. It contains measurements which may be aggregated in various ways [10, 12, 39]. Dimension tables contain descriptive textual information. Dimension attributes are used as the constraints in the data warehouse queries. Dimension tables provide the basis for aggregating the measurements in the fact table. They generally consist of embedded hierarchies. Each star schema is formed in the following way;

41

A fact table is formed for each transaction entity. The key of the table is the combination of the keys of its associated component entities. A dimension table is formed for each component entity, by collapsing hierarchically related classification entities into it. Where hierarchical relationships exist between transaction entities, the child entity inherits all dimensions (and key attributes) from the parent entity. This provides the ability to drill down between transaction levels.

Numerical attributes within transaction entities should be aggregated by key attributes (dimensions). The aggregation attributes and functions used depend on the application.

Star schemas can be used to speed up query performance by denormalizing reference information into a single dimension table. Denormalization is appropriate when there are a number of entities related to the dimension table that are often accessed, avoiding the overhead of having to join additional tables to access those attributes. Denormalization is not appropriate where the additional data is not accessed very often, because the overhead of scanning the expanded dimension table may not be offset by gain in the query performance. The advantage of using this schema; it reduces the number of tables in the database and the number of relationships between them and also the number of joins required in user queries. The Figure 4.20 shows a sample star schema.

Figur e 4.20 Star Schema 42

4.6.5. Fact Constellation Schem a A fact constellation schema consists of a set of star schemas with hierarchically linked fact tables. The links between the various fact tables provide the ability to drill down between levels of detail [10, 12]. The following figure, Figure 4.21, illustrates a sample of a fact constellation schema.

Figur e 4.21 Fact Constellation Schema 4.6.6. Galaxy Schema Galaxy schema is a schema where multiple fact tables share dimension tables. Unlike a fact constellation schema, the fact tables in a galaxy do not need to be directly related [12]. The following figure, Figure 4.22, illustrates a sample of a galaxy schema.

43

Figur e 4.22 Galaxy Schema 4.6.7. Snowflake Schem a In a star schema, hierarchies in the original data model are collapsed or denormalized to form dimension tables. Each dimension table may contain multiple independent hierarchies. A snowflake schema is a variant of star schema with all hierarchies explicitly shown and dimension tables do not contain denormalized data [10, 12]. The many-to-one relationships among sets of attributes of a dimension can separate new dimension tables, forming a hierarchy. The decomposed snowflake structure visualizes the hierarchical structure of dimensions very well. A snowflake schema can be produced by the following procedure: A fact table is formed for each transaction entity. The key of the table is the combination of the keys of the associated component entities.

44

Each component entity becomes a dimension table. Where hierarchical relationships exist between transaction entities, the child entity inherits all relationships to component entities (and key attributes) from the parent entity.

Numerical attributes within transaction entities should be aggregated by the key attributes. The attributes and functions used depend on the application.

The following figure, Figure 4.23, illustrates a sample of a snowflake schema.

Figur e 4.23 Snowflake Schema 4.6.8. Star Cluster Schema While snowflake contains fully expanded hierarchies, which adds complexity to the schema and requires extra joins, star schema contains fully collapsed hierarchies, which leads to redundancy. So, the best solution may be a balance between these two schemas [12]. Overlapping dimensions can be identified as forks in hierarchies. A fork occurs when an entity acts as a parent in two different dimensional hierarchies. Fork entities can be identified as classification entities with multiple one-to-many relationships. In Figure 4.24, Region is parent of both Location and Customer entities and the fork occurs at the Region entity.

45

Figur e 4.24 Star Schema with fork A star cluster schema is a star schema which is selectively snowflaked to separate out hierarchical segments or sub dimensions which are shared between different dimensions. A star cluster schema has the minimal number of tables while avoiding overlap between dimensions. A star cluster schema can be produced by the following procedure: A fact table is formed for each transaction entity. The key of the table is the combination of the keys of the associated component entities. Classification entities should be collapsed down their hierarchies until they reach either a fork entity or a component entity. If a fork is reached, a sub dimension table should be formed. The sub dimension table will consist of the fork entity plus all its ancestors. Collapsing should begin again after the fork entity. When a component entity is reached, a dimension table should be formed. Where hierarchical relationships exist between transaction entities, the child entity should inherit all dimensions (and key attributes) from the parent entity. Numerical attributes within transaction entities should be aggregated by the key attributes (dimensions). The attributes and functions used depend on the application. The Figure 4.25 illustrates a sample diagram of star cluster schema.

46

Figur e 4.25 Star Cluster Schema 4.6.9. Star flake Schema Starflake schema is a hybrid structure that contains a mixture of star and snowflake schemas. The most appropriate database schemas use a mixture of denormalized star and normalized snowflake schemas [6, 41]. The Figure 4.26 illustrates a sample diagram of starflake schema.

Figur e 4.26 Star flake Schem a 47

Whether the schema star, snowflake or starflake, the predictable and standard form of the underlying dimensional model offers important advantages within a DW environment including; Efficiency; the consistency of the database structure allows more efficient access to the data by various tools including report writers and query tools. Ability to handle changing requirements; The star schema can adapt to changes in the user requirements, as all dimensions are equivalent in terms of providing access to the fact table. Extensibility; the dimensional model is extensible. It must support adding new dimensions, adding new dimensional attributes, breaking existing dimension records down to lower level of granularity from a certain point in time forward. Ability to model common business situations Predictable query processing

The following figure, Figure 4.27, shows a comparison of the logical design methods in complexity versus redundancy trade-off.

Figur e 4.27 Compar ison of schemas 4.6.10. Cube Cubes are the logical storage structures for OLAP databases. A cube defines a set of related dimensions; each cell of the cube hold one value, the value of each cell is an

48

intersection of the dimensions. A 2-dimensional view of an OLAP table is given in Table 4.1.

Table 4.1 2-dimensional pivot view of an OLAP Table A 3-dimensional view of an OLAP table is given in Table 4.2.

Table 4.2 3-dimensional pivot view of an OLAP Table A 3-D realization of the cube shown in Table 4.2 is illustrated in Figure 4.28.

49

Figur e 4.28 3-D Realization of a Cube The cube has three dimensions which are time, state and product. Each dimension enables you to perform specific OLAP operations on the cube. The basic OLAP operations are as follows [10, 23, 24]: Roll up: An operation for moving up the hierarchy level and grouping into larger units along a dimension. Using roll up capability, users can zoom out to see a summarized level of data. Roll up operation is also called the drill up operation. Drill down: An operation for moving down the hierarchy level and stepping down the hierarchy. Using drill down capability, users can navigate to higher levels of detail. Drill down operation is the reverse of roll up operation. Slice: Slicing performs a selection on one dimension of a cube and results in a sub cube. Slicing cuts through the cube so that users can focus on more specific perspectives. Dice: Slicing defines a sub cube by performing a selection on two or more dimensions of a cube. 50

Pivot: Pivot operation is also called rotate operation. Pivoting is a visualization operation which rotates the data axes in view in order to provide an alternative presentation of the data.

A detailed figure describing the operations above is illustrated below in Figure 4.29.

51

Figur e 4.29 Oper ations on a Cube

52

4.7. Meta Data An important component of the DW environment is meta data. Meta data, or data about data, provides the most effective use of the DW. Meta data allows the end user/DSS analyst to navigate through the possibilities. In other words, when a user approaches a data warehouse where there is no meta data, the user does not know where to begin the analysis. Meta data acts like an index to the data warehouse contents. It sits above the warehouse and keeps track of what is where in the warehouse. Typically, items the meta data store tracks are as follows [6]: Structure of data as known to the programmer and to the DSS analyst Source data Transformation of data Data model DW History of extracts

Metadata has several functions within the DW that relates to the processes associated with data transformation and loading, DW management and query generation. The metadata associated with data transformation and loading must describe the source data and any changes that were made to the data. The metadata associated with data management describes the data as it is stored in the DW. Every object in the database needs to be described including the data in each table, index and view and any associated constraints. The metadata is also required by the query manager to generate appropriate queries. 4.8. Mater ialized views They address the problem of selecting a set of views to materialize in a DW taking into account [7]: the space allocated for materialization the ability of answering a set of queries (defined against the source relations) using exclusively these views 53

the combined query evaluation and view maintenance cost

In this proposal they define a graph based on states and state transitions. They define a state as a set of views plus a set of queries, containing an associated cost. Transitions are generated when views or queries are changed. They demonstrate that there is always a path from an initial state to the minimal cost state. 4.9. OLAP Ser ver Ar chitectur es Logically, OLAP engines present business users with multidimensional data from data warehouses or data marts, without concerns regarding how or where the data are stored. However, the physical architecture and implementation of OLAP engines must consider data storage issues. Implementations of a warehouse server engine for OLAP processing include [10]: Relation al OLAP (ROLAP) servers: These are the intermediate servers that stand in between a relational back-end server and client front-end tools. They use a relational or extended-RDBMS to store and manage warehouse data, and OLAP middleware to support missing pieces. ROLAP servers include optimization for each DBMS back-end, implementation of aggregation navigation logic, and additional tools and services. ROLAP technology tends to have greater scalability than MOLAP technology. Multidimensional OLAP (MOLAP) servers: These servers support multidimensional views of data through array-based multidimensional storage engines. They map multidimensional views directly to data cube array structures. The advantage of using a data cube is that it allows fast indexing to precomputed summarized data. Notice that with multidimensional data stores, the storage utilization may be low if the data set is sparse. In such cases, sparse matrix compression techniques should be explored. Many OLAP servers adopt a two-level storage representation to handle sparse and dense data sets: the dense subcubes are identified and stored as array structures, while the sparse subcubes employ compression technology for efficient storage utilization. Hybr id OLAP (HOLAP) ser vers: The hybrid OLAP approach combines ROLAP and MOLAP technology, benefiting from the greater scalability of ROLAP and the faster computation of MOLAP. For example, a HOLAP server may allow large volumes of

54

detail data to be stored in a relational database, while aggregations are kept in a separate MOLAP store.

55

CHAPTER 5

COMPARISON OF MULTIDIMENSIONAL DESIGN MODELS

5.1. Compar ison of Dimensional Models and ER Models The main objective of ER modeling is to remove redundancy from data. ER modeling aims to optimize performance for transaction processing. To remove redundancy, designers must use hundreds of entities and relations between entities, which makes ER model complex. There is no easy way to enable endusers navigate through the data in ER models. It is also hard to query ER models because of the complexity; many tables should be joined to obtain a result set. Therefore ER models are not suitable for high performance retrieval of data. The dimensional model is a standard framework. End user tools can make strong assumptions about the dimensional model to make user interfaces more user friendly and to make processing more efficient [20]. Dimensional model is more adaptable to unexpected changes in user behavior and requirements. The logical design can be made independent of expected query patterns. All dimensions can be thought as symmetrically equal entry points into the fact table. Dimensional model is extensible to new design decisions and data elements. All existing fact and dimension tables can be changed in place without having to reload data. End user query and reporting tools are not affected by the change.

56

ER modeling does not involve business rules, it involves data rules. Dimensional model involves business rules. 5.2. Compar ison of Dimensional Models and Ob ject-Or iented Models Dimensional data modeling (DDM) is a dimensional model design methodology that for each business process, it enumerates relevant measures and dimensions. A DDM provides a multidimensional conceptual view of data. Although DDM is the favorite approach in data warehousing, it focuses mainly on data and its proper structuring to maximize query performance. DDM approach lacks modeling business goals and processes. Although the final logical and physical model will be a dimensional data model, object-oriented (OO) model is much stronger in the logical and conceptual design phases. We should consider DDM approach and OO approach as complementary to each other, a logical design modeled by an OO model can be mapped easily to a DDM model. Various approaches have been developed for the conceptual design of multidimensional systems in the last years to represent multidimensional structural and dynamic properties. However, none of them has been accepted as a standard for multidimensional modeling. On the other hand, UML has been accepted as the standard OO modeling language for describing and designing various aspects of software systems. Using UML, OO model allows modeling of the business process, sub

processes, use cases, system actors, classes, objects, collaboration of objects, relations between object and finally components, which are basically reusable software packages. Objects have types, properties (attributes), behaviors, methods and relations with other objects like aggregation, inheritance and association. A DDM approach is basically an approach in which tables are associated with SQL methods to support set-oriented processing of data and return result set to the caller. On the other hand, OO approach provides an object layer to a DW application unifying behavior and data within the object components. An OO approach provides a tighter conceptual association between strategic business goals and objectives and the DDM model. It starts with the specification of business goals, then modeling of the processes and use cases. Based on the use cases, 57

objects are modeled. The resulting DDM model will be valid with the corresponding business requirements. In addition to the discussion above, I summarized the comparison of ER, DM and OO methodologies according to factors, I consider as the most important, in Table 5.1. ER Standard notation Business rules focus Data rules focus Ability to model all business requirements in detail Specialization / Generalization Commercial case tool availability High association with business objectives Adaptability to changing requirements Table 5.1 Comp ar ison of ER, DM and OO meth odologies 5.3. Compar ison of Conceptual Multidimensional Models In sections 5.1 and 5.2, the main conceptual modeling approaches are mentioned. This section gives a comparison of conceptual multidimensional models according multidimensional modeling properties [13, 14, 15, 16, 17, 18, 19]. Additivity of measures: DF, starER and OOMD support this property. Using ME/R model, only static data structure can be captured. No functional aspect can be implemented with ME/R model; therefore ME/R does not support this property. Many-to-many relationships with dimensions: StarER and OOMD support this property. DF and ME/R models do not support many-to-many relationships. Derived measures: None of the conceptual models include derived measures as part of their conceptual schema except OOMD model. 58 DM OO (UML)

Nonstrict and complete classification hierarchies: Although DF, starER and ME/R can define certain attributes for classification hierarchies, only starER model can define exact cardinality for nonstrict and complete classification hierarchies. OOMD can represent nonstrict and complete classification hierarchies.

Categorization of dimensions (specialization/generalization): DF does not support this property. Since starER and ME/R models derive from ER model, these two models use is-a relationship to categorize dimensions. Since OOMD model is object-oriented, it has support for dimension categorization. Note that specialization/generalization is a basic aspect of object-orientation.

Graphic notation and specifying user requirements: All modeling techniques provide a graphical notation to help designers in conceptual modeling phase. Only ME/R model provides state diagrams to model systems behavior and provides a basic set of OLAP operations to be applied from these user requirements. OOMD provide complete set of UML diagrams to specify user requirements and help define OLAP functions.

Case tool support: All conceptual design models except starER have case tool support. Conceptual design using DF approach can be implemented using the WAND case tool [31]. Conceptual design using ME/R approach can be implemented using GramMi case tool [29]. With the OOMD approach, the conceptual design may be implemented using Microsoft Visio, Rational Rose or GOLD case tools [18, 30].

The Table 5.2 summarizes the comparison given above. starER Additivity of measures Many-to-many with dimensions Derived measures Nonstrict and complete relationships DF ME/R OOMD

classification hierarchies Categorization of dimensions

59

(specialization/generalization) Graphic notation Specifying user requirements Case tool support MS Visio Table 5.2 Compar ison of conceptu al design models 5.4. Compar ison of L ogical Design Models Among the logical design models, star schema, snowflake schema and the fact constellation schema are the mostly used models commercially. In this section, I want to compare these three models in terms of efficiency, usability, reusability and flexibility quality factors. I think efficiency is the most important factor in DW modeling. A DW is usually a very large database. Because many queries will access large amounts of data and involve multiple join operations, efficiency becomes a major consideration [22]. A star schema is generally the most efficient design for two reasons. First, a design with denormalized tables need fewer joins. Second, most optimizers recognize star schemas and can generate efficient star join operations. A fact constellation schema is a set of star schemas with hierarchically. A fact constellation schema may need more join operations on fact tables. Similarly, a snowflake schema will require more joins on dimension tables. In some cases where the denormalized dimension tables in star schema becomes very large, a snowflake schema may be the most efficient design approach. In terms of usability, a number of advantages may be considered for star schema design approach. The star schema is the simplest structure among the three schemas. Because a star schema has the fewest number of tables, users need to execute fewer join operations which makes it easier to formulate analytic queries. It is easier to learn star schema compared to other two schemas. Fact constellation schema and snowflake schema are more complex than star schema which is a disadvantage in terms of usability. Considering reusability, the snowflake schema is more reusable than star and fact constellation schemas. Dimension tables in a snowflake schema do not contain 60

WAND GRAMMI GOLD, Rational Rose,

denormalized data. This makes it easier to share dimension tables between snowflake schemas in a DW. In star schema and fact constellation schema design approaches, dimension tables are denormalized and this makes it less convenient to share dimension tables between schemas. In terms of flexibility, a star schema is more flexible in adapting to changes in user requirements. The star schema can adapt to changes in the user requirements easier, as all dimensions are equivalent in terms of providing access to the fact table. Table 5.3 summarizes the comparison of the three logical design models in terms of quality factors. Star Schema Snowflake Schem a Fact Constellation Schema Efficiency Usability Reusability Flexibility High High Low High Low Low High Low Moderate Moderate Moderate Moderate

Table 5.3 Compar ison of logical design models 5.5. Discussion on Data War ehousing Design Tools There are CASE tools that enable a user to design DW architecture. Some of the CASE tools are mentioned in section 5.3. Current commercial CASE tools have great design options to model that enable modelers to model databases and software solutions even in the enterprise level. These tools can generate code that may be used in the development phase in forms of VB.NET, C#, C++ and may reverse engineer a given source code project into software models. Using these CASE tools, a designer can also generate databases via the tool and reverse engineer databases into model using the design model diagrams. Unfortunately, the development of the CASE tools on the data warehousing area is not as mature as the development in ER and software modeling areas. Very few commercially available tools may help in designing data warehousing solutions and may still not cover the requirements you need.

61

A CASE tool without basing the modeling notation on UML may never cover the needs of data warehousing design. Note that, the main purpose of UML is to represent business rules [26]. Unfortunately, a case tool for a complete data warehousing design is not available. But there are still solutions using the existing CASE tools in data warehousing arena. As Kimball mentions [27] I am cautious about recommending methodologies and CASE tools from traditional environments for data warehouse software development. Methodologies scare me the most. So often I see a methodology becoming the goal of a project, rather than just being a tool. Recently someone asked me: What is the difference between a methodologist and a terrorist? I answered, At least you can negotiate with a terrorist. . Finally, a designer/developer should use the existing CASE tools for the development of a DW. There are some factors in which data warehouse software developers must consider to make more use of existing software development tools. These topics are summarized as follows: Definition and declaration of business rules: Communication and documentation are very important for data warehouse business rules design. Unfortunately, we should not expect significant automatic conversion of the most complex business rules into code using the present commercial CASE tools. Complex business rules still remain in a documentation-only form, and probably will only be enforced by a human designer. The best CASE tool meeting this requirement may be Microsoft Visio 2003 Enterprise Edition. Database Design: Most of the commercial CASE tools enable designing ER databases. It should be addressed that although these CASE tools may support generating ER databases from models and reverse engineering from an ER generated database into an ER model, these tools do not provide a complete solution for an OLAP solution. For designing a database Microsoft Visio, ERWin or WarehouseArchitect may be used. ETL Process: The commercially available tools for ETL process meets the requirements needed by this process. The widely used ETL product is

62

Microsoft Data Transformation Services (DTS) which is a built-in feature of Microsoft SQL Server 2000. DTS allows defining ETL processes on both its built-in designer or using the COM API of DTS or using scripting language. This tool can accept any data source that has an ODBC or OLEDB provider. As an example, a data source may be Microsoft SQL Server 7.0, Microsoft SQL Server 2000, Oracle, Sybase, DB2, a text file, an Microsoft Access database, an Microsoft Excel spreadsheet (all versions), etc. In this thesis, the ETL process of the sample DW solution is implemented using DTS and described in section 6.5. Metadata Repository Design: Vendors like Microsoft, IBM, Oracle have all either defined a global metadata repository format or promised the market that they will do so in the future. But, the world needs a single standard instead of multiple repository definitions. Most analysts believe that Microsofts repository effort is the one most likely to succeed. In 1997, Microsoft turned over the content of its proprietary repository design to the Metadata Coalition, and worked with a group of vendors to define the Open Information Model (OIM), intended to support software development and data warehousing [27]. Microsoft and many other vendors are actively programming their tools to read and write from the Microsoft Repository. Communication: While the nature of data warehousing requires the need of consolidating data from heterogeneous data sources, data sources may become homogenous. As XML become a standard of communication, the XML web services appeared. An XML Web service is a programmable entity that provides a particular element of functionality, such as application logic, and is accessible to any number of potentially disparate systems using ubiquitous Internet standards, such as XML and HTTP [28] . Web services change the nature of DWs.

63

CHAPTER 6

IMPLEMENTING A DATA WAREHOUSE

6.1. A Case Study This section describes a case study of implementing a DW. A database is designed for simulating an OLTP application. The aim of this case study is to build a DW to enable analysis of the data in the OLTP database. The conceptual design is modeled using OOMD, starER, ME/R and DF models to give example designs for all approaches. The DW is designed using snowflake schema. And finally this case study illustrates a implementation of a data warehousing solution covering all phases of design. The ER model of the OLTP database that simulates the basis for our DW is shown in Figure 6.1.

64

Figur e 6.1 ER model of sales and shipping systems 6.2. OOMD Appr oach Data warehousing has largely developed with little or no reference to objectoriented software engineering. This is consistent with (a) its development out of two-tier client/server relational database methodology, and (b) its character as a kind of highlevel systems integration, rather than software development, activity. Data Warehousing assembles components, rather than creating them. The initial top-down centralized data warehousing model with its single database server, limited middleware, and GUI frontend, could get by with a pragmatic "systems integration" orientation at the global level of component interaction, while restricting explicit object-orientation to the various GUI, middleware, analysis, web-enabling, data transformation, and reporting tools that comprise the data warehousing arsenal. The days of two-tier client/server-based data warehouses are gone now. The dynamic of the data warehousing business and its own development of new technology,

65

has caused centralized data warehouses to be increasingly supplemented with data marts, with data stores of diverse type and content, with specialized application servers, with design and metadata repositories, and with internet, intranet, and extranet front-ends. The two-tier client/server paradigm has given way to a multi-tier reality characterized by an increasing diversity of objects/components, and by an increasing complexity of object relations. Moreover this reality is one of physically distributed objects across an increasingly complex network architecture. Data warehouse development now requires the creation of a specific distributed object/component solution, and the evolutionary development of this solution over time, rather than the creation of a classical two-tier client/server application [25]. I strongly believe it is more convenient to model complex software systems with OO design approach using UML. So, before going to implementation details, I would like to introduce some sample diagrams for the sales and shipping system using the OO conceptual modeling approach. Respectively, Figure 6.2 illustrates the use case

diagram, Figure 6.3 shows the statechart diagram and Figure 6.4 illustrates the static structure diagram of sales and shipping system using UML notation.

Figur e 6.2 Use case diagr am of sales and shipping system

66

Use case diagrams are used to describe real-world activities. A use case is a set of events that occurs when an actor uses a system to complete a process. Normally, a use case is a relatively large process.

Figur e 6.3 Statechar t diagr am of sales and shipping system Statechart diagrams are used to show the sequence of states an object goes through during its life.

Figur e 6.4 Static str uctur e diagr am of sales and shipping system 67

Static structure diagrams are used to create conceptual diagrams that represent concepts from the real world and the relationships between them, or class diagrams that decompose a software system into its parts. UML modeling has also the following diagrams that give modeler great flexibility and helps understandability during the conceptual and logical design phases. Package diagrams are used to group related elements in a system. One package can contain subordinate packages, diagrams, or single elements. Activity diagrams are used to describe the internal behavior of a method and represent a flow driven by internally generated actions. Sequence diagrams are used to show the actors or objects participating in an interaction and the events they generate arranged in a time sequence. Collaboration diagrams are used to show relationships among object roles such as the set of messages exchanged among the objects to achieve an operation or result. Component diagrams are used to partition a system into cohesive components and show the structure of the code itself. Deployment diagrams are used to show the structure of the run-time system and communicate how the hardware and software elements that make up an application will be configured and deployed. Static structure diagram in Figure 6.4 forms the basis of the MD model. This diagram can easily be mapped to MD model. Sales and Shipping classes form the fact tables, Employee, Shipper, Region, State, Country, Product, Product_SubCategory and Product_Category classes form the dimension tables. 6.3. star ER Appr oach In this section, the conceptual model is designed using starER approach. Figure 6.5 illustrates the sales subsystem starER model. The following list describes the items in the starER model. The items represented as circles are fact sets. The items represented as rectangles are entity sets. The items represented as diamonds are relationship sets. The items respresented as ovals are attributes.

68

The model contains one fact set (sales) and four dimensions (employee, product, time and state). The time, product and state dimensions contain hierarchies that enable summarization of the fact set by different granularities. Non-complete membership is shown in these dimensions.

Figur e 6.5 Sales subsystem star ER model Figure 6.6 illustrates the shipping subsystem starER model.

69

Figur e 6.6 Shipping subsystem star ER model The model contains one fact set (shipping) and five dimensions (shipper, product, time and state_from, state_to). The time, product and states dimensions contain hierarchies that enable summarization of the fact set by different granularities. Noncomplete membership is shown in these dimensions. It is relatively easy and straightforward to design the conceptual model using starER modeling technique having the ER of the OLTP system. 6.4. ME/R Appr oach In this section, the conceptual model is designed using ME/R approach. Figure 6.7 illustrates the sales subsystem with ME/R approach. The first design step is determining dimensions and facts. Fact relationship connects the sales fact with the dimensions state, product, day and employee. The rolls-up relationships (arrow shapes) are shown in the model below. The fact relationship in the middle of the diagram connects the atomic dimension levels. Each dimension is

70

represented by a subgroup that starts at the corresponding atomic level. The actual facts (sales units, sales dollars) are modelled as attributes of the fact relationship. The dimension hierarchies are shown by the rolls-up relationships (e.g. Product rolls-up to product_subcategory and product_category). Additional attributes of a dimension level (e.g. employee_id and employee_name of an employee) are depicted as dimension attributes of the corresponding level.

Figur e 6.7 Sales subsystem ME/R model Figure 6.8 illustrates the shipping subsystem with ME/R approach and is similar to the sales subsytem.

71

Figur e 6.8 Shipping subsystem ME/R model 6.5. DF Appr oach In this section, the conceptual model is designed using DF approach. Figure 6.9 illustrates the sales subsystem with DF approach. The sales fact scheme is structured as a tree whose root is the sales fact. The fact is illustrated as a rectangle and contains the fact name (sales) and measures (sales_dollars, sales_units). Each vertex directly attached to the fact is a dimension. The dimensions in the sales model are product, time, state and employee. Subtrees rooted in dimensions are hierarchies. Their vertices, represented by circles, are attributes (product_ID); their arcs represent -to-one relationships between pairs of attributes. The vertices in the fact schema represented by lines instead of circles are non-dimension attributes (product_name).

72

Figur e 6.9 Sales subsystem DF model Figure 6.10 illustrates the shipping subsystem with DF approach.

Figur e 6.10 Shipping subsystem DF model

73

6.6. Implementation Details In this section, using the OOMD approach and taking the static structure diagram in Figure 6.4 as the basis, the physical implementation of the sample DW is given. In the model, there are two fact tables; sales and shipping. I have chosen snowflake schema for the implementation of the DW. The main reason for choosing the snowflake schema is that, the sample OLTP database I have prepared as the data source of my sample DW is completely formed of normalized tables and therefore using snowflake schemas in the design of the DW is more applicable and easier to implement. The sample DW is implemented using Microsoft OLAP Server [8]. The snowflake schema for the sales subsystem is illustrated in Figure 6.11.

Figur e 6.11 Snowflake schema for the sales subsystem.

74

The snowflake schema for the shipping subsystem is illustrated in Figure 6.12.

Figur e 6.12 Snowflake schema for the shipping subsystem. The diagram in Figure 6.13 illustrates the general architecture of the case study.

Figur e 6.13 Gener al ar chitectur e of the case study The ETL process is implemented using Microsoft Data Transformation Services (DTS). DTS is a set of graphical tools and programmable objects that let the designer extract, transform, and consolidate data from different sources into single or multiple destinations. A DTS package is an organized collection of connections, DTS tasks, DTS

75

transformations, and workflow constraints assembled either with a DTS tool or programmatically and saved to Microsoft SQL Server, Microsoft SQL Server 2000 Meta Data Services, a structured storage file, or a Microsoft Visual Basic file. Each package contains one or more steps that are executed sequentially or in parallel when the package is run. When executed, the package connects to the correct data sources, copies data and database objects, transforms data, and notifies other users or processes of events. Packages can be edited, password protected, scheduled for execution, and retrieved by version. A DTS transformation is one or more functions or operations applied against a piece of data before the data arrives at the destination. The source data is not changed. For example, the user can extract a substring from a column of source data and copy it to a destination table. The particular substring function is the transformation mapped onto the source column. The user also can search for rows with certain characteristics (for example, specific data values in columns) and apply functions only against the data in those rows. Transformations make it easy to implement complex data validation, data scrubbing, and conversions during the import and export process. DTS is based on an OLE DB architecture that allows copying and transforming data from a variety of data sources. Some of these are listed below: SQL Server and Oracle directly, using native OLE DB providers. ODBC sources, using the Microsoft OLE DB Provider for ODBC. Microsoft Access 2000, Microsoft Excel 2000, Microsoft Visual FoxPro, dBase, Paradox, HTML, and additional file data sources. Text files, using the built-in DTS flat file OLE DB provider. Microsoft Exchange Server, Microsoft Active Directory and other nonrelational data sources. Other data sources provided by third-party vendors.

In the sample implementation, the sales fact table and the shipping fact table are populated from a delimited text file, an Access database, an Excel spreadsheet and a SQL Server database. The DTS packages implemented are shown in Figure 6.14 and Figure 6.15 respectively.

76

Figur e 6.14 Sales DTS Package

Figur e 6.15 Shipping DTS Package

77

The arrow from delimited sales text file to DW Data Source indicates the transformation from the text file to the SalesFact table. The transformation is a copy column transformation with the details shown in Figure 6.16.

Figur e 6.16 Tr ansfor mation details for delimited text file Transformations from Microsoft Excel and Microsoft Access are similar to transformation from the delimited text file. Transformation from the SQL Server database is a little different. The table with the data is not loaded as it is, but with a custom query. The detail of this transformation is shown Figure 6.17. As seen in the figure, the transformation source is a custom Transact-SQL query with grouping and aggregation of data.

78

Figur e 6.17 Tr ansact-SQL quer y as the tr ansfor mation sour ce Another feature of DTS is lookup queries. Lookup queries allow running queries and stored procedures against other connections besides the source and destination. For example, by using a lookup query, you can make a separate connection during a query and include data from that connection in the destination table. Lookup queries are especially useful in validating input data before loading it. After the populating data from different sources and transforming and loading the data into the DW using DTS, we need some client application for end users to enable them query the DW. In the sample implementation I used Microsoft Excel and Microsoft Data Analyzer as the client applications and mainly the pivot table and pivot chart technologies. The Figure 6.18 shows a snapshot of pivot chart view of the sales cube.

79

Figur e 6.18 Pivot Ch ar t using Excel as client The Figure 6.19 shows pivot table view of the sales cube.

Figur e 6.19 Pivot Table using Excel as client

80

The Figure 6.20 shows the sales cube using Microsoft Data Analyzer.

Figur e 6.20 Data Analyzer as client Microsoft Data Analyzer is business-intelligence software that enables you to apply the intelligence of an organization to the challenges of displaying and analyzing data in a quick and meaningful manner. Data Analyzer accomplishes this by giving the user a complete overview in one screen, which helps the user to quickly find hidden problems, opportunities, and trends. By using Data Analyzer, non-technical business users can get answers to their questions immediately and independently, which puts knowledge directly into the hands of the people who need it most decision-makers at all levels in the organization. Data Analyzer provides a number of advantages for displaying and analyzing the data: A complete overview on a single screen replaces masses of grids, graphs, and reports.

81

Multidimensional views show relationships throughout all aspects of your business. Customizable displays guide the user to hidden problems, opportunities, and trends. The dynamic use of color highlights anomalies. Saving and reporting functions allow the user to save views for future use and export them to Microsoft PowerPoint or Microsoft Excel. Power users can do more advanced analysis in less time.

82

CHAPTER 7

CONCLUSIONS AND FUTURE WORK

Successful data management is an important factor in developing support systems for the decision-making process. Traditional database systems, called operational or transactional, do not satisfy the requirements for data analysis of the decision-making users. An operational database supports daily business operations and the primary concern of such database is to ensure concurrent access and recovery techniques that guarantee data consistency. Operational databases contain detailed data, do not include historical data, and since they are usually highly normalized, they perform poorly for complex queries that need to join many relational tables or to aggregate large volumes of data. A DW represents a large repository of integrated and historical data needed to support the decision-making process. The structure of a DW is based on a multidimensional model. This model includes measures that are important for analysis, dimensions allowing the decision-making users to see these measures from different perspectives, and hierarchies supporting the presentation of detailed or summarized measures. The characteristics of a multidimensional model specified for the DW can be applied for a smaller structure, a data mart, which is different from the DW in the scope of its analysis. A data mart refers to a part of an organization and contains limited amount of data.

83

DWs have become the main technology for DSS. DSSs require not only the data repository represented by DW, but also the tools that allow analysing data. These tools include different kinds of applications; for example, software that include statistics and data mining techniques offers complex analysis for a large volume of data to identify profiles, behaviour, and tendencies. On the other hand, OLAP tools can manage high volumes of historical data allowing for dynamic data manipulations and flexible interactions with the end-users through the drill-down, roll-up, pivoting, and slicingdicing operations. Furthermore, OLAP tools are based on multidimensional concepts similar to DW multidimensional model using for it measures, dimensions, and hierarchies. If the DW data structure has a well-defined multidimensional model, it is easier to fully exploit OLAP tools capabilities. In this thesis, widely accepted conceptual and logical design approaches in DW design are discussed. In the conceptual design phase DF, starER, ME/R and OOMD design models are compared. OO design model is significantly better than the other design approaches. OOMD supports conceptual design phase with a rich set of diagrams that enables the designer model all the business information and requirements using a case tool with UML. OOMD design model meets the following factors while the others lack one or more: Additivity of measures Many-to-many relationships with dimensions Derived measures Nonstrict and complete classification hierarchies Categorization of dimensions (specialization/generalization) Graphic notation Specifying user requirements Case tool support

In the logical design phase flat, terraced, star, fact constellation, galaxy, snowflake, star cluster and starflake schemas are discussed. Among these logical design models, star schema, snowflake schema and the fact constellation schema are the mostly used models commercially. These three models are compared in terms of efficiency, usability,

84

reusability and flexibility quality factors among which efficiency is the most important one considering DW modeling. Considering these factors and the requirements of the business and considering the trade-off between redundancy and the query performance, either snowflake or star schema may be the best choice in the design. There are CASE tools that enable a user to design DW architecture. Current commercial CASE tools have great design options to model that enable modelers to model databases and software solutions even in the enterprise level. Unfortunately, the development of the CASE tools on the data warehousing area is not as mature as the development in ER and software modeling areas. Very few commercially available tools may help in designing data warehousing solutions and may still not cover the requirements you need. A CASE tool without basing the modeling notation on UML may never cover the needs of data warehousing design. Note that, the main purpose of UML is to represent business rules. Unfortunately, a case tool for a complete data warehousing design is not available. But there are still solutions using the existing CASE tools in data warehousing arena. Two of the commercial CASE tools that support OOMD design are Microsoft Visio and Rational Rose. The sample OOMD model in the thesis is implemented using Microsoft Visio. Likewise, there are a number of OLAP Servers like Microsoft SQL Server Analysis Services, Hyperion Essbase, PowerPlay, DB2 OLAP Server. As mentioned in the thesis, data warehousing is a complete process starting with data acquisition, designing and storage of data and finally enabling end users access this data. It is important for a platform to be able to offer a complete solution in data warehousing. In this thesis, the data warehousing application is implemented using Microsoft technologies. For the data acquisition phase Microsoft DTS, for design and storage phase Microsoft SQL Server Analysis Services and finally for end user access Microsoft Excel and Microsoft Data Analyzer are used. 7.1. Contr ibutions of the Thesis This thesis contributes to both theory and practice of data warehousing. The data warehouse design models and approaches in the literature are researched and grouped

85

according to the phase in the project development cycle. These models are refined according to the acceptance in the literature. The first contribution of the thesis is the comparison of methodologies namely E/R, DM and OO. The second contribution of the thesis is the comparison of the conceptual design models. The third contribution of the thesis is the comparison of the logical design models in terms of the quality factors. The fourth contribution of the thesis is a case study covering the phases of a data warehousing solution. In addition, issues and problems identified in this thesis that might impact the data warehouse implementation will enable project managers to develop effective strategies for successfully implementing data warehouses. According to my research, there is no complete study in literature on DW models providing a mapping of models to development phases and giving a comparison of the models according to these phases. Also, very few articles point to an implementation covering all phases of data warehousing. So, the last contribution of the thesis is providing a complete study on these missing points. 7.2. Futur e Wor k One possible future work may be comparing the pyhsical design models for a data warehousing solutions and extend the case study to cover these physical design approaches. Another future work may be implementing a more complex case study using real world application data, perform performance tests using the three logical models compared to support the comparison on logical design models presented in the thesis. Another future work may be improving the comparison of logical design models by both covering more model and more quality factor in the comparison. A case tool for a complete data warehousing design is not available. One future work may be implementing of a case tool meeting the requirements of a data warehousing solution.

86

REFERENCES

[1] Romm M., Introduction to Data Warehousing, San Diego SQL User Group [2] Goyal N., Introduction to Data Warehousing, BITS, Pilani Lecture Notes [3] Franconi E., Introduction to Data Warehousing, Lecture http://www.inf.unibz.it/~franconi/teaching/2002/cs636/2 ,2002 Notes,

[4] Pang L., Data Warehousing and Data Mining, Leslie Pang Web Site and Lecturer Notes [5] Gatziu S. and Vavouras A., Data Warehousing: Concepts and Mechanisms, 1999 [6] Thomas Connolly & Carolyn Begg., Database Systems, 3th Edition, AddisonWesley, 2002 [7] Gatierrez A. and Marotta A., An Overview of Data Warehouse Design Approaches and Techniques, Uruguay, 2000 [8] Reed Jacobson., Microsoft SQL Server 2000 Analysis Services, ISBN 0-73560904-7, 2000 [9] Rizzi S., Open Problems in Data Warehousing., http://sunsite.informatik.rwthaachen.de/Publications/CEUR-WS/Vol-77/ DMDW 2003, Berlin, Germany [10] J. Han and M. Kamber, Data Mining: Concepts and Techniques, Chapter2: Data Warehouse and OLAP Technology for Data Mining, Barnes & Nobles, 2000 [11] W. H. Inmon, Building the Data Warehouse, 3th Edition, John Wiley, 2002

87

[12] Moody D. L. and Kortink M. A. R., From Enterprise Models to Dimensional Models: Methodology for Data Warehouse and Data Mart Design, http://sunsite.informatik.rwth-aachen.de/Publications/CEUR-WS//Vol-28/ http://sunsite.informatik.rwth DMDW 2000 , Stockholm, Sweden [13] Tryfona N., Busborg F., Christiansen J. G., starER: A Conceptual Model for Data Warehouse Design, Proceeding of the ACM 2nd International Workshop Data Warehousing and OLAP (DOLAP99), 1999 [14] Sapia C., Blaschka M., Hfling G., Dinter B., Extending the E/R Model for the Multidimensional Paradigm, Proceeding 1st International Workshop on Data Warehousing and Data Mining (DWDM98), 1998 [15] Golfarelli M., Maio D., Rizzi S., Conceptual Design of Data Warehouses from E/R Schemas, Proceeding of the 31st Hawaii International Conference on System Sciences (HICSS-31), Vol. VII,1998 [16] Golfarelli M., Maio D., Rizzi S., The Dimensional Fact Model: A Conceptual Model For Data Warehouses, International Journal of Cooperative Information Systems (IJCIS), Vol. 7, 1998 [17] Golfarelli M, Rizzi S., A Methodological Framework for Data Warehouse Design, Proceeding of the ACM DOLAP98 Workshop, 1998 [18] Lujan-Mora S., Trujillo J., Song I., Multidimensional Modeling with UML Package Diagrams, 21st International Conference on Conceptual Modeling (ER2002), 2002 [19] Trujillo J., Palomar M., An Object Oriented Approach to Multidimensional Database Conceptual Modeling (OOMD) , Proceeding 1st International Workshop on Data Warehousing and OLAP (DOLAP98), 1998 [20] Kimball R., http://www.dbmsmag.com/9708d15.html A Dimensional Modeling Manifesto, DBMS Magazine, Aug 1997 [21] Kimball R., The Data Warehouse Toolkit , John Wiley, 1996 [22] Martyn T., Reconsidering Multi-Dimensional Schemas, SIGMOD Record, Vol. 33, No. 1, 2004 [23] Elmasri R., Navathe S., Fundamentals of Database Systems, 3rd Edition, Addison-Wesley, 2000 [24] Ballard C., Herreman D., Schau D., Bell R., Kim E., and Valencic A., Data Modeling Techniques for Data Warehousing, IBM Redbook, IBM International Technical Support Organization, 1998

88

[25] Firestone J., Object-Oriented Data Warehousing, 1997 [26] Kimball R., Enforcing the Rules, 2000 , http://www.intelligententerprise.com/000818/webhouse.jhtml?_requestid=380244 [27] Kimball R., The Software Developer in Us, 2000, http://www.intelligententerprise.com/000908/webhouse.jhtml http://www.intelligententerprise.com/000908/webho [28] Microsoft Developer Network (MSDN) Library, XML Web Services Overview, October 2004 [29] Hahn K., Sapia C., and Blaschka M., Automatically Generating OLAP Schemata from Conceptual Graphical Models, Proceedings ACM 3rd International Workshop Data Warehousing and OLAP (DOLAP 2000), 2000 [30] Mora-Lujan S., Multidimensional Modeling Using UML and XML, Proceedings 16th European Conference on Object-Oriented Programming (ECOOP 2002), 2002 [31] Golfarelli M., Rizzi S., WAND: A Case Tool for Data Warehouse Design, Demo Proceedings of The 17th International Conference on Data Engineering (ICDE 2001), 2001 [32] Chaudhuri S., Dayal U., An Overview of Data Warehousing and OLAP Technology, ACM Sigmod Record, vol.26, 1997 [33] Golfarelli M., Rizzi S., Designing the Data Warehouse: Key Steps and Crucial Issues, Journal of Computer Science and Information Management, 1999 [34] Phipps C., Davis K., Automating Data Warehouse Conceptual Schema Design and Evaluation, DMDW02, 2002 [35] Peralta V., Marotta A., Ruggia R., Towards the Automation of Data Warehouse Design, 2003 [36] Batini C., Ceri S., Navathe S., Conceptual Database Design-An Entity Relationship Approach, Addison-Wesley, 1992 [37] Abello A., Samos J., Saltor F., A Data Warehouse Multidimensional Data Models Classification, Technical Report, 2000 [38] Abello A., Samos J., Saltor F., A Framework for the Classification and Description of Multidimensional Data Models, Database and Expert Systems Applications, 12th International Conference, 2001 [39] Teklitz F., The Simplification of Data Warehouse Design, Sybase, 2000

89

[40] Prosser A., Ossimitz M., Data Warehouse Management , University of Economics and Business Admin., Vienna, 2000 [41] Ahmad I., Azhar S., Data Warehousing in Construction: From Conception to Application, First International Conference on Construction in the 21st Century (CITC2002) Challenges and Opportunities in Management and Technology , 2002 [42] Kimball R., Letting the Users Sleep, Part 1, DBMS, 1996, http://www.dbmsmag.com/9612d05.html http://www.dbmsmag.com/9612d05.ht [43] Kimball R., Letting the Users Sleep, Part 2, DBMS, 1997, http://www.dbmsmag.com/9701d05.html

90