I can remember the first time that I worked on data paging code. I had to page through web site search results 20 at a time on a given web page. My task was to understand how it had been written and to do some bug fixing. After reviewing about 2,000 lines of code, and reviewing the seven different variables that were being used to know the current page, the next page, the previous page, the page size, the row at the top of the page, the row at the bottom of the page, and the number of pages, I finally understood what the code was intended to do. What a mess that was, but beyond the mess, the performance was horrible. The way the page worked was that based on the page you were on, all of the rows would be queried, then a loop would read through all of the rows before the current page, then loop through the rows on the current page displaying them on the page, and finally it would ignore the results after the current page. So page 1 was slow, page 2 was slower than page 1, page 3 was slower than page 2 and on and on.
Since that point I have implemented several different data paging algorithms myself, all better than the original implementation but none as elegant as the CTE way of doing data paging. I used to look at data paging as a painful task, but thanks to the SQL Server implementation of CTEs there is no more pain.
Steve also covers
FETCH. This technique won’t be great with enormous data sets, but for moderate-sized data sets which query quickly, it works. This is one area which is quite painful, and the best (and wackiest) solution I’ve come up with in SQL Server when the initial query is quite expensive is to create tables with random names to store results and row numbers, populate a table the first time a query is run, and query that table on subsequent runs, using the RETURN value in a stored procedure to pass along the name of the table to access. Granted, that solution works best with static data and you’d want to have a method to clean up those tables after they’re no longer in use (like storing a list of those tables and their last access dates and times). So it’s a mess.