SQL Tuning [Englisch] [Taschenbuch]

This book is refreshing. It doesn't waste time going over all of the stuff you learned years ago.

Knowing how to read an execution plan or when to pick a hash join over a nested loop join is not what this book is about. There are plenty of books on the market that cover basic, vendor specific, query tuning. I personally have about 20 of these books on my bookshelf here at home. (Over the years I've worked on Sybase, SQL Server, Informix XPS, & Oracle.)

Here's the deal...

Anyone who has worked with really big systems will eventually run into an optimization problem that seems to be unsolvable. You can try histograms, compressed key indexes, partitioning, pre-joined indexes, and materialized views, but you still can't get the performance that's being requested. For a DBA, it can be a very frustrating dilemma. This is especially true when you know from the data volume that you should be able to get there.

The truth is, optimizers can't always get the right solution, even with correct statistics. There are some good technical reasons why this is true, but that's out of scope for my review. In any case, that's where this book comes to the rescue. I feel that it gives you some insight into the optimization problem and tells you how to correct the problems that your optimizer can't figure out.

This book is NOT for use on 95% of your queries. Most optimizers will pick the correct access plan if the DBA does his/her job correctly and collects the appropriate statistics.

In my experience, I get two types of problems that I have trouble getting the optimizer to solve:

#1. Joining together a large number (8-14) of tables. At least 1 or 2 of the tables have over 30G of real data. By `real data', I mean that 30G of data is actually populated.

#2. Making high transaction queries read the fewest amount of buffers in order to get rid of latching problems.

After I read the first few diagramming chapters of this book, I thought I'd give it a try on a problem that was recently solved at work. I was surprised. It worked, and even with my clumsiness with the method, it only took me about 2 hours to get a solution. It took us about 3 days at work. Our trial and error solution was slightly better, than what I came up with using the author's method. However, it was so close that had I used his method, I most certainly would have gone on to another problem. (The author's solution was strange because I would never have solved the join order the way that he did. His solution actually had me pick the largest table, out of 10, to drive the query. I thought this was odd because my test query had some very good filters on smaller tables.)

These were the results per execution:

Optimizer - 100K buffers.
Trial & Error - 1700 buffers.
Book method (basic graphing) - 2000 buffers.

If I had known about this book, I could have solved this problem in a few hours. Instead it took several DBA's, a few day's time to come up with a solution that was only marginally better.

Since this method appeared to work so well for OLTP, I went back and looked at a very large OLAP query that I had worked on last year. I and several developers had spent days trying to get it to run faster. Even with the author's method, I still couldn't make it better, but I was happy to see that the join order picked by the optimizer was almost identical to what I had calculated using the book.

I plan on studying this guide over and over until I've memorized this method. I don't think it will solve every problem, but I think it gives you an edge over using experience alone.

Before you purchase this book...

Keep in mind, that SQL tuning is fundamentally *not a simple problem*, so readers should not buy the book expecting an easy list of simple tips and tricks. (If the problem was easy, the optimizer likely would have got it right in the first place, and you wouldn't be tuning!) The correct solution to the problem is fairly complex, as complex as it *needs* to be.

Also, it's not noted anywhere, but I gathered that the author assumes that the reader will have basic high school Algebra skills, and will have taken a basic statistics course in order to fully understand the reasoning behind the explanations. Although, Algebra and statistical knowledge are probably not necessary to learn the tuning method presented. I state this because I told a good friend about this book, and his first question was, "Does it have a lot of math to go through?" So, if statistics is not your thing, don't worry. The author only presents enough information to give you a level of comfort that his method is based on mathematics. He doesn't require that you know "graph theory" or anything like that.

By the way, our change from 100K buffers per transaction to 1700 buffers, dropped the CPU utilization for the website that was running the OLTP query by about 70%. That was on a 16 CPU, Sun/Solaris box. So proper SQL tuning can save you a lot of $$$.

The first four chapters of the book cover some of the subjects that are found in other SQL tuning books, and focus on the important points without delving too deeply into database internals. The really "good stuff" is covered in chapters 5 through 10, which includes how to "draw" a query diagram, and the procedures to follow to derive an optimal join order. The problems/examples presented and resolved in these chapters help the reader get a start on understanding the methodology, which can then be applied to actual tuning problems as they are encountered.

So far I have applied this technique to about 20 tuning problems I have encountered on the job using the Oracle RDBMS. I still consider myself a relative novice in using the methodology, but for each of the 20 problems I have been able to significantly reduce elapsed time and resource consumption (gets/reads). Also, these results were achieved in a matter or hours, rather than days. The query diagram has even facilitated finding "missing joins" in some of the queries without having a knowledge of the particular application.

This book is a "must have" for DBA's and developers and I highly recommend it.