by Craig Utley
Creating a Star Schema Database is one of the most important, and sometimes the final, step in creating a data warehouse. Given how important this process is to building a data warehouse, it is important to understand how to move from a standard, on-line transaction processing (OLTP) system to a final star schema. Please note that a general term is relational data warehouse and may cover both star and snowflake schemas.
This paper attempts to address some of the issues that many who are new to the data warehousing arena find confusing, such as:
- What is a Data Warehouse? What is a Data Mart?
- What is a Star Schema Database?
- Why do I want/need a Star Schema Database?
- The Star Schema looks very denormalized. Won’t I get in trouble for that?
- What do all these terms mean?
This paper will attempt to answer these questions, and show developers how to build a star schema database to support decision support within their organizations.
Usually, readers of technical articles are bored with terminology that comes either at the end of a chapter or is buried in an appendix at the back of a book. Here, however, I have the thrill of presenting some terms up front. The intent is not to bore readers earlier than usual, but to present a baseline off of which to operate. The problem in data warehousing is that the terms are often used loosely by different parties. The definitions presented here represent how the terms will be used throughout this paper.
OLTP stands for Online Transaction Processing. This is a standard, normalized database structure. OLTP is designed for transactions, which means that inserts, updates, and deletes must be fast. Imagine a call center that takes orders. Call takers are continually taking calls and entering orders that may contain numerous items. Each order and each item must be inserted into a database. Since the performance of the database is critical, database designers want to maximize the speed of inserts (and updates and deletes). To maximize performance, some businesses even limit the number of records in the database by frequently archiving data.
OLAP and Star Schema
OLAP stands for Online Analytical Processing. OLAP is a term that means many things to many people. Here, the term OLAP and Star Schema are basically interchangeable. The assumption is that a star schema database is an OLAP system. An OLAP system consists of a relational database designed for the speed of retrieval, not transactions, and holds read-only, historical, and possibly aggregated data.
While an OLAP/Star Schema may be the actual data warehouse, most companies build cube structures from the relational data warehouse in order to provide faster, more powerful analysis on the data.
Data Warehouse and Data Mart
Data Warehouses and Data Marts differ in scope only. This means that they are built using the exact same methods and procedures, so the process is the same while only their intended scope varies.
A data warehouse (or mart) is way of storing data for later retrieval. This retrieval is almost always used to support decision-making in the organization. That is why many data warehouses are considered to be DSS (Decision-Support Systems). While some data warehouses are merely archival copies of data, most are used to support some type of decision-making process. The primary benefit of taking the time to create a star schema, and then possibly cube structures, is to speed the retrieval of data and format that data in a way that it is easy to understand. This means that a star schema is built not for transactions but for queries.
Both a data warehouse and a data mart are storage mechanisms for read-only, consolidated, historical data. Read-only means that the person looking at the data won’t be changing it. If a user wants to look at the sales yesterday for a certain product, they should not have the ability to change that number. Of course if the number is wrong, it should be corrected, but more on that later.
“Consolidated” means that the data may have come from various sources. Many companies have purchased different vertical applications from various vendors to handle such tasks as human resources (HR), accounting/finance, inventory, and so forth. These systems may run on multiple operating systems and use different database engines. Each of these applications may store their own copy of an employee table, product table, and so on. A relational data warehouse must take data from all these systems and consolidate it so it is consistent, which means it is in a single format.
The “historical” part means the data may be only a few minutes old, but often it is at least a day old. A data warehouse usually holds data that goes back a certain period in time, such as five years. In contrast, standard OLTP systems usually only hold data as long as it is “current” or active. An order table, for example, may move order data to an archive table once the order has been completed, shipped, and received by the customer.
The data in data warehouses and data marts may also be aggregated. While there are many different levels of aggregation possible in a typical data warehouse, a star schema may have a “base’ level of aggregation, which is one in which all the data is aggregated to a certain point in time.
For example: assume a company sells only two products: dog food and cat food. Each day, the company records the sales of each product. At the end of a couple of days, the data looks like this:
4/24/99 1 5 2 2 3 0 3 2 6 4 2 2 5 3 3 4/25/99 1 3 7 2 2 1 3 4 0
Clearly, each day contains several transactions. This is the data as stored in a standard OLTP system. However, the data warehouse might not record this level of detail. Instead, it could summarize, or aggregate, the data to daily totals. The records in the data warehouse might look something like this:
Quantity Sold Date Dog Food Cat Food 4/24/99 15 13 4/25/99 9 8
This summarization of data reduces the number of records by aggregating the individual transaction records into daily records that show the number of each product purchased each day.
In this simple example, it is easy to derive Table 2 simply by running a query against Table 1. However, many complexities enter the picture that will be discussed later.
There is no magic to the term “aggregations.” It simply means a summarized, typically additive value. The level of aggregation in a star schema depends on the scenario. Many star schemas are aggregated to some base level, called the grain, although this is becoming somewhat less common as developers rely on cube building engines to summarize to a base level of granularity.
OLTP, or Online Transaction Processing, systems are standard, normalized databases. OLTP systems are optimized for inserts, updates, and deletes; in other words, transactions. Transactions in this context can be thought of as the entry, update, or deletion of a record or set of records.
OLTP systems achieve greater speed of transactions through a couple of means: they minimize repeated data, and they limit the number of indexes. The minimization of repeated data is one of the primary drivers behind normalization.
When examining an order, systems typically break orders down into an order header and then a series of detail records. The header contains information such as an order number, a bill-to address, a ship-to address, a PO number, and other fields. An order detail record is usually a product number, a product description, the quantity ordered, the unit price, the total price, and other fields. Here is what an order might look like:
The data stored for this order looks very different. If stored in a flat structure, the detail records look something like this:
|Order Number||Order Date||Customer ID||Customer Name||Customer Address||Customer City|
|12345||4/24/99||451||ACME Products||123 Main Street||Louisville|
|Customer State||Customer Zip||Contact Name||Contact Number||
|Product Description||Category||Sub Category||Product Price||Quantity Ordered||Etc…|
|¼” Brass Widget||Brass Goods||Widgets||$1.00||200||Etc…|
Notice, however, that for each detail, much of the data is being repeated: the entire customer address, the contact information, the product information, and so forth. All of this information is needed for each detail record, but the system should have to store all the customer and product information for each record. Relational technology allows each detail record to tie to a header record, without having to repeat the header information in each detail record. The new detail records might look like this:
Order Number Product Number Quantity Ordered 12473 A4R12J 200
A simplified logical view of the tables might look something like this:
Notice that the extended cost is not stored in the OrderDetail table. OLTP schemas store as little data as possible to speed inserts, updates, and deletes. Therefore, any number that can be calculated at query time is calculated and not stored.
Developers also minimize the number of indexes in an OLTP system. Indexes are important but they slow down inserts, updates, and deletes. Therefore, most schemas have just enough indexes to support lookups and other necessary queries. Over-indexing can significantly decrease performance.
Database normalization is the process of removing repeated information. As shown above, normalization reduces repeated information of the order header record in each order detail record. Normalization is a process unto itself, and what follows is merely a brief overview.
Normailzation first removes repeated records in a table. For example, the following order table contains much repeated information and is not recommended:
In this example, there will be some limit on the number of order detail records in the Order table. If there were twenty repeated sets of fields for detail records, the table would be unable to handle an order for twenty one or more products. In addition, if an order has just has one product ordered, all the other fields are useless.
So, the first step in the normalization process is to break the repeated fields into a separate table, and end up with this:
Now, an order can have any number of detail records.
As stated before, OLTP allows for the minimization of data entry. For each detail record, only the primary key value from the OrderHeader table is stored, along with the primary key of the Product table, and then the order quantity is added. This greatly reduces the amount of data entry necessary to add a product to an order.
Not only does this approach reduce the data entry required, it greatly reduces the size of an OrderDetail record. Compare the size of the record in Table 3 to that in Table 4. The OrderDetail records take up much less space with a normalized table structure. This means that the table is smaller, which helps speed inserts, updates, and deletes.
In addition to keeping the table smaller, most of the fields that link to other tables are numeric. Queries generally perform much better against numeric fields than they do against text fields. Therefore, replacing a series of text fields with a numeric field can help speed queries. Numeric fields also index faster and more efficiently.
With normalization, there are frequently fewer indexes per table. Each transaction requires the maintenance of affected indexes. With fewer indexes to maintain, inserts, updates, and deletes run faster.
There are some disadvantages to an OLTP structure, especially when retrieving the data for analysis. First, queries must utilize joins across multiple tables to get all the data. Joins tend to be slower than reading from a single table, so minimizing the number of tables in a query will boost performance. With a normalized structure, developers have no choice but to query from multiple tables to get the detail necessary for a report.
One of the advantages of OLTP is also a disadvantage: fewer indexes per table. Fewer indexes per table are great for speeding up inserts, updates, and deletes. In general terms, the fewer indexes in the database, the faster inserts, updates, and deletes will be. However, again in general terms, the fewer indexes in the database, the slower select queries will run. For the purposes of data retrieval, a higher number of correct indexes helps speed retrieval. Since one of the design goals to speed transactions is to minimize the number of indexes, OLTP databases trade faster transactions at the cost of slowing data retrieval. This is one reason for creating two separate database structures: an OLTP system for transactions, and an OLAP system for data retrieval.
Last but not least, the data in an OLTP system is not user friendly. Most IT professionals would rather not have to create custom reports all day long. Instead, they would prefer to give their customers some query tools so customers could create reports without involving the IT organization. Most customers, however, don’t know how to make sense of the normalized structure of the database. Joins are somewhat mysterious, and complex table structures (such as associative tables on a bill-of-material system) are difficult for the average customer to use. The structures seem obvious to IT professionals, who sometimes wonder why customers can’t get the hang of it. Remember, however, that customers know how to do a FIFO-to-LIFO revaluation and other such tasks that IT people may not know how to do; therefore, understanding relational concepts just isn’t something customers should have to worry about.
If customers want to spend the majority of their time performing analysis by looking at the data, the IT group should support their desire for fast, easy queries. On the other hand, maintaining the speed requirements of the transaction-processing activities is critical. If these two requirements seem to be in conflict, they are, at least partially. Many companies have solved this by having a second copy of the data in a structure reserved for analysis. This copy is more heavily indexed, and it allows customers to perform large queries against the data without impacting the inserts, updates, and deletes on the main data. This copy of the data is often not just more heavily indexed, but also denormalized to make it easier for customers to understand.
Reasons to Denormalize
When database administrators are asked why they would ever denormalize, the first (and often only) answer is: speed. Recall one of the key disadvantages to the OLTP structure: It is built for data inserts, updates, and deletes, but not data retrieval. Therefore, one method of squeezing some speed out of it is by denormalizing some of the tables and having queries pull data from fewer tables. These queries are faster because they perform fewer joins to retrieve the same recordset.
Joins are relatively slow, as has already been mentioned. Joins are also confusing to many end users. By denormalizing, users are presented with a view of the data that is far easier for them to understand. Which view of the data is easier for a typical end-user to understand:
The second view is much easier for the end user to understand. While a normalized schema requires joins to create this view, putting all the data in a single table allows the user to perform this query without using joins. Creating a view that looks like this, however, still uses joins in the background and therefore does not achieve the best performance on the query. Fortunately, there is a better way.
How Humans View Information
All of this leads to the real question: how do humans view the data stored in the database? This is not the question of how humans view it with queries, but how do they logically view it? For example, are these intelligent questions to ask:
- How many bottles of Aniseed Syrup were sold last week?
- Are overall sales of Condiments up or down this year compared to previous years?
- On a quarterly and then monthly basis, are Dairy Product sales cyclical?
- In what regions are sales down this year compared to the same period last year? What products in those regions account for the greatest percentage of the decrease?
All of these questions would be considered reasonable, perhaps even common. They all have a few things in common. First, there is a time element to each one. Second, they all are looking for aggregated data; they are asking for sums or counts, not individual transactions. Finally, they are looking at data in terms of “by” conditions.
“By” conditions refer to looking at data by certain conditions. For example, take the question: “On a quarterly and then monthly basis, are Dairy Product sales cyclical?” This can be rephrased with the following statement: “We want to see total sales by category (just Dairy Products in this case), by quarter or by month.”
Here the customer is looking at an aggregated value, the sum of sales, by specific criteria. Customers can add further “by” conditions by saying they wanted to see those sales by brand and then the individual products.
Figuring out the aggregated values to be shown, such as the sum of sales dollars or the count of users buying a product, and then figuring out the “by” conditions is what drives the design of the star schema.
Making the Database Match Expectations
If the goal is to view the data as aggregated numbers broken down along a series of “by” criteria, why isn’t data simply stored in this format?
That’s exactly what is done with the star schema. It is important to realize that OLTP is not meant to be the basis of a decision support system. The “T” in OLTP stands for transactions, and a transaction is all about taking orders and depleting inventory, and not about performing complex analysis to spot trends. Therefore, rather than tie up an OLTP system by performing huge, expensive queries, the answer is to build a database structure that maps to the way humans see the world.
Humans see the world in a multidimensional way. Most people think of cube structures when speaking of multiple dimensions, but cubes are typically built from relational data that has already been put into a dimensional model. The dimensional model is a database structure to support queries, and cubes can then be built on it later.
Facts and Dimensions
When examining how people look at data, they usually want to see some sort of aggregated data. These data are called measures. These measures are numeric values that are measurable and usually additive. For example, sales dollars are a perfect measure. Every order that comes in generates a certain sales volume measured in some currency. If a company sells twenty products in one day, each for five dollars, they generate 100 dollars in total sales. Therefore, sales dollars is one measure most companies track. Companies may also want to know how many customers they had that day. Did five customers buy an average of four products each, or did just one customer buy twenty products? Sales dollars and customer counts are two measures businesses may want to track.
Just tracking measures isn’t enough, however. People need to look at measures using those “by” conditions. The “by” conditions are called dimensions. In order to examine sales dollars, people almost always wan to see them by day, or by quarter, or by year. There is almost always a time dimension on anything people ask for. They may also want to know sales by category or by product. These “by” conditions will map into dimensions: there is almost always a time dimension, and product and geography dimensions are very common as well.
Therefore, in designing a star schema, the first order of business is usually to determine what people want to see (the measures) and how they want to see it (the dimensions).
Mapping Dimensions into Tables
Dimension tables answer the “why” portion of a question: how do people want to slice the data? For example, people almost always want to view data by time. Users often don’t care what the grand total for all data happens to be. If the data happens to start on June 14, 1989, do users really care how much total sales have been since that date, or do they really care how one year compares to other years? Comparing one year to a previous year is a form of trend analysis and one of the most common things done with data in a star schema.
Relational data warehouses may also have a location or geography dimension. This allows users to compare the sales in one region to those in another. They may see that sales are weaker in one region than any other region. This may indicate the presence of a new competitor in that area, or a lack of advertising, or some other factor that bears investigation.
When designing dimension tables, there are a few rules to keep in mind. First, all dimension tables should have a single-field primary key. This key is typically a surrogate key and is often just an identity column, consisting of an automatically incrementing number. The value of the primary key is meaningless, hence the surrogate key; the real information is stored in the other fields. These other fields, called attributes, contain the full descriptions of the dimension record. For example, if there is a Product dimension (which is common) there are fields in it that contain the description, the category name, the sub-category name, the weight, and so forth. These fields do not contain codes that link to other tables. Because the fields contain full descriptions, the dimension tables are often fat; they contain many large fields.
Dimension tables are often short, however. A company may have many products, but even so, the dimension table cannot compare in size to a normal fact table. For example, even if a company has 30,000 products in the product table, the company may track sales for these products each day for several years. Assuming the company actually only sells 3,000 products in any given day, if they track these sales each day for ten years, they end up with this equation: 3,000 products sold X 365 day/year * 10 years equals almost 11,000,000 records! Therefore, in relative terms, a dimension table with 30,000 records will be short compared to the fact table.
Given that a dimension table is fat, it may be tempting to normalize the dimension table. Normalizing the dimension tables is called a snowflake schema and will be discussed later in this paper.
Developers have been building hierarchical structures in OLTP systems for years. However, hierarchical structures in an OLAP system are different because the hierarchy for the dimension is actually stored in a single dimension table (unless snowflaked as discussed later.)
The product dimension, for example, contains individual products. Products are normally grouped into categories, and these categories may well contain sub-categories. For instance, a product with a product number of X12JC may actually be a refrigerator. Therefore, it falls into the category of major appliance, and the sub-category of refrigerator. There may have more levels of sub-categories, which would further classify this product. The key here is that all of this information is stored in fields in the dimension table.
The product dimension table might look something like this:
Notice that both Category and Subcategory are stored in the table and not linked in through joined tables that store the hierarchy information. This hierarchy allows users to perform “drill-down” functions on the data. They can execute a query that performs sums by category and then drill-down into that category by calculating sums for the subcategories within that category. Users can then calculate the sums for the individual products in a particular subcategory.
The actual sums being calculated are based on numbers stored in the fact table. These will be examined when discussing the fact table later.
Consolidated Dimensional Hierarchies (Star Schemas)
The above example (Figure 7) shows a hierarchy in a dimension table. This is how the dimension tables are built in a star schema; the hierarchies are contained in the individual dimension tables. No additional tables are needed to hold hierarchical information.
Storing the hierarchy in a dimension table allows for the easiest browsing of the dimensional data. In the above example, users could easily choose a category and then list all of that category’s subcategories. They would drill-down into the data by choosing an individual subcategory from within the same table. There is no need to join to an external table for any of the hierarchical information.
In this overly-simplified example, there are two dimension tables joined to the fact table. The fact table will examined later. For now, examples will use only one measure: SalesDollars.
In order to see the total sales for a particular month for a particular category, a SQL query would look something like this:
SELECT Sum(SalesFact.SalesDollars) AS SumOfSalesDollars
FROM TimeDimension INNER JOIN (ProductDimension INNER JOIN
SalesFact ON ProductDimension.ProductID = SalesFact.ProductID)
ON TimeDimension.TimeID = SalesFact.TimeID
WHERE ProductDimension.Category=’Brass Goods’ AND TimeDimension.Month=3
To drill down to a subcategory, the SQL would change to look like this:
SELECT Sum(SalesFact.SalesDollars) AS SumOfSalesDollars
FROM TimeDimension INNER JOIN (ProductDimension INNER JOIN
SalesFact ON ProductDimension.ProductID = SalesFact.ProductID)
ON TimeDimension.TimeID = SalesFact.TimeID
WHERE ProductDimension.SubCategory=’Widgets’ AND TimeDimension.Month=3
Sometimes, the dimension tables have the hierarchies broken out into separate tables. This is a more normalized structure, but leads to more difficult queries and slower response times.
Figure 9 represents the beginning of the snowflake process. The category hierarchy is being broken out of the ProductDimension table. This structure increases the number of joins and can slow queries. Since the purpose of an OLAP system is to speed queries, snowflaking is usually not productive. Some people try to normalize the dimension tables to save space. However, in the overall scheme of the data warehouse, the dimension tables usually only account for about 1% of the total storage. Therefore, any space savings from normalizing, or snowflaking, are negligible.
Building the Fact Table
The Fact Table holds the measures, or facts. The measures are numeric and additive across some or all of the dimensions. For example, sales are numeric and users can look at total sales for a product, or category, or subcategory, and by any time period. The sales figures are valid no matter how the data is sliced.
While the dimension tables are short and fat, the fact tables are generally long and skinny. They are long because they can hold the number of records represented by the product of the counts in all the dimension tables.
For example, take the following simplified star schema:
In this schema, there are product, time and store dimensions. With ten years of daily data, 200 stores, and 500 products, there is a potential of 365,000,000 records (3650 days * 200 stores * 500 products). This large number of records makes the fact table long. Adding another dimension, such as a dimension of 10,000 customers, can increase the number of records by up to 10,000 times.
The fact table is skinny because of the fields it holds. The primary key is made up of foreign keys that have migrated from the dimension tables. These fields are typically integer values. In addition, the measures are also numeric. Therefore, the size of each record is generally much narrower than those in the dimension tables. However, there are many, many more records in the fact table.
One of the most important decisions in building a star schema is the granularity of the fact table. The granularity, or frequency, of the data is determined by the lowest level of granularity of each dimension table, although developers often discuss just the time dimension and say a table has a daily or monthly grain. For example, a fact table may store weekly or monthly totals for individual products. The lower the granularity, the more records that will exist in the fact table. The granularity also determines how far users can drill down without returning to the base, transaction-level data.
One of the major benefits of the star schema is that the low-level transactions may be summarized to the fact table grain. This greatly speeds the queries performed as part of the decision support process. The aggregation or summarization of the fact table is not always done if cubes are being built, however.
Fact Table Size
The previous section discussed how 500 products sold in 200 stores and tracked for 10 years could produce 365,000,000 records in a fact table with a daily grain. This, however, is the maximum size for the table. Most of the time, there are far fewer records in the fact table. Star schemas do not store zero values unless zero has some significance. So, if a product did not sell at a particular store for a particular day, the system would not store a zero value. The fact table contains only the records that have a value. Therefore, the fact table is often sparsely populated.
Even though the fact table is sparsely populated, it still holds the vast majority of the records in the database and is responsible for almost all of the disk space used. The lower the granularity, the larger the fact table. In the previous example, moving from a daily to weekly grain would reduce the potential number of records to only slightly more than 52,000,000 records.
The data types for the fields in the fact table do help keep it as small as possible. In most fact tables, all of the fields are numeric, which can require less storage space than the long descriptions we find in the dimension tables.
Finally, be aware that each added dimension can greatly increase the size of the fact table. If just one dimension was added to the previous example that included 20 possible values, the potential number of records would reach 7.3 billion.
One of the greatest challenges in a star schema is the problem of changing attributes. As an example, examine the simplified star schema in Figure 10. In the StoreDimension table, each store is located in a particular region, territory, and zone. Some companies realign their sales regions, territories, and zones occasionally to reflect changing business conditions. However, if the company simply updates the table to reflect the changes, and users then try to look at historical sales for a region, the numbers will not be accurate. By simply updating the region for a store, the total sales for that region will appear as if the current structure has always been true. The business has “lost” history.
In some cases, the loss of history is fine. In fact, the company might want to see what the sales would have been had this store been in that other region in prior years. More often, however, businesses do not want to change the historical data. In this case, the typical approach is to create a new record for the store. This new record contains the new region, but leaves the old store record, and therefore the old regional sales data, intact. This approach, however, prevents companies from comparing this stores current sales to its historical sales unless the previous StoreID is preserved. In most cases the answer it to keep the existing StoreName (the primary key from the source system) on both records but add BeginDate and EndDate fields to indicate when each record is active. The StoreID is a surrogate key so each record has a different StoreID but the same StoreName, allowing data to be examined for the store across time regardless of its reporting structure.
This particular problem is usually called a “slowly-changing dimension” and there are various methods for handling it. The actual implementation is beyond the scope of this paper.
There are no right and wrong answers. Each case will require a different solution to handle changing attributes.
The data in the fact table is already aggregated to the fact table’s grain. However, users often ask for aggregated values at higher levels. For example, they may want to sum sales to a monthly or quarterly number. In addition, users may be looking for a total at a product or category levels.
These numbers can be calculated on the fly using a standard SQL statement. This calculation takes time, and therefore some people will want to decrease the time required to retrieve higher-level aggregations.
Some people store higher-level aggregations in the database by pre-calculating them and storing them in the the fact table. This requires that the lowest-level records have special values put in them. For example, a TimeDimension record that actually holds weekly totals might have a 9 in the DayOfWeek field to indicate that this particular record holds the total for the week.
A second approach is to build another fact table but at the weekly grain. All data is summarized to the weekly level and stored there. This works well for storing data summarized at various levels, but the problem comes into play when examining the number of possible tables needed. To summarize at the weekly, monthly, quarterly, and yearly levels by product, four tables are needed in addition to the “real”, or daily, fact table. However, what about weekly totals by product subcategory? And monthly totals by store? And quarterly totals by product category and territory? Each combination would require its own table.
This approach has been used in the past, but better alternatives exist. These alternatives usually consist of building a cube structure to hold pre-calculated values. Cubes were designed to address the issues of calculating aggregations at a variety of levels and respond to queries quickly.
The star schema, also called a relational data warehouse or dimensional model, is a consolidated, consistent, historical, read-only database storing data from one or more systems. The data often comes from OLTP systems but may also come from spreadsheets, flat files, and other sources. The data is formatted in a way to provide fast response to queries. Star schemas provide fast response by denormalizing dimension tables and potentially through providing many indexes.
Star schemas may be the end of the data warehousing process, but often they are the source for a cube-building product. Different engines work in different ways, but most store the data in new binary formats for even quicker retrieval, and calculate aggregations at various levels of granularity. While most modern cube-building engines do not requires a star schema as their source, a star schema is still the best source as the data has already been consolidated and made consistent before the cube is built.
Note on version 1.1
I’ve been surprised by the popularity of this paper, which I wrote the night before a talk at my first SQL Server Connections conference back in 1999. I’ve finally gotten around to making some minor updates, which include a summary, fixing a few typos, and changing from first person to third person. I wouldn’t mind completely rewriting the paper so perhaps a version 2.0 will appear at some point. – Craig Utley, 17 July 2008