One of the many things you should try to keep a continuous watch for during the life of any of your applications are the performance of your SQL queries. You might be caching the hell out of your database layer, but some time you’ll have to hit the database server to retrieve data. And if that starts to happen often enough while you’re growing, you will see your SQL daemon taking up the largest part of your disk io and your CPU time. This might not be a problem for the load you’re seeing now, but could you handle a 10 fold increase in traffic? .. or how about 100x? (which, if I remember correctly, is what Google uses as the scale factor when developing applications)
Indexes Are Your Friend
During the christmas holiday I got around to taking a look at some of the queries running at one of my longest living, most active sites: pwned.no. Pwned is a tournament engine running on top of PHP and MySQL, containing about 40.000 tournaments, 450.000 matches and several other database structures. The site has performed well over the years and there hasn’t been any performance issues other than a few attempts at DoS-ing the site with TCP open requests (the last one during the holiday, actually).
Two weeks ago the server suddenly showed loads well above 30 – while it usually hovers around 0.3 – 0.4 at the busiest hours of the day. The reason? One of the previously less used functions of the tournament engine, using a group stage in your tournament, had suddenly become popular in at least one high traffic tournament. This part of the code had never been used much before, but when the traffic spike happened everything went bananas (B-A-N-A-N-A-S. Now that’s stuck in your head. No problem.) The reason: the query used a couple of columns in a WHERE-statement that wasn’t indexed, and the query ran against the table containing the matches for the tournament. This meant that over 400.000 rows were scanned each time the query ran, meaning that mysqld started hogging every resource it could. The Apache childs then had to wait, making the load a bit too much for my liking. Two CREATE INDEX-calls later the load went back down and everything chugged along nicely again.
My strategy for discovering queries that might need a better index scheme (or if “impossible”, a proper caching layer in front of it):
- Run your development server with slow-query-log=1, log-queries-not-using-indexes=1 and long-query-time=<an appropriately low value, such as 0.05 – depends on your setup>. You can also provide a log file name with the log-slow-queries=/var/log/mysql/… in your my.cnf-file for MySQL. This will log all potential queries for optimizing to the log file (this will not necessarily provide you with a complete list of good queries to optimize, but it might provide a few good hints). Be sure to use actual data from your site when working on your development version, as you might start seeing issues when the size of the data set reaches a certain size – such as 400.000 rows in the example mentioned above)
- Connect to your MySQL server and issue
SHOW FULL PROCESSLIST
statements every now and then. This will let you see any queries that run often and way too long (but they’ll have to run when you issue the command). You might not catch the real culprit, but if you’re seing MySQL chugging along with 100% CPU and are wondering what’s happening, try to check out what the threads are doing. You’ll hopefully see just which query is wreaking havoc with your server.
- Add a statistics layer in front of your MySQL calls in your application. If you’re using PDO you can subclass it to keep a bit of statistics about your queries around. The number of times each query is run, the time it took in total running the query and other interesting values. We’re using a version of this in the development version of Gamer.no and I’ll probably upload the class to my github repository as soon as I get a bit of free time in the new year.
Not sure what I’ll take a closer look at tomorrow, but hopefully I’ll decide before everything collapses!
What are your strategy for indexes? What methods do you use for finding queries that need a bit more love? Leave a comment below!