I had a call from a new customer last week. They had issues with their MySQL server. Apparently, it was repeatingly crashing, every few hours. To have their production system kept alive, they used a script to periodically see if MySQL was still alive, and if not – start it.
I was first thinking in directions of old installations; wrong memory allocations (too little memory for large content; to large memory allocation for weak machine). When I reached the customer’s premises, I quickly reviewed general settings, while explaining some basic configuration guidelines. There were peculiarities (e.g. query_cache_limit being larger than query_cache_size), but nothing to obviously suggest a reason for crash.
I then observed the error log. Took some time to find it, since the log_error parameter appeared twice in the my.cnf file; first one appearing at the very beginning, the second (overwriting the first) was far into the file.
Sure enough, there were a lot of startup messages. And… shutdown messages. In fact, the log looked something like:
090923 17:38:15 mysqld started 090923 17:38:16 InnoDB: Started; log sequence number 0 3707779031 090923 17:38:16 [Note] /usr/local/mysql/bin/mysqld: ready for connections. Version: '5.0.77-log' socket: '/tmp/mysql50.sock' port: 3306 MySQL Community Server (GPL) 090923 19:53:41 [Note] /usr/local/mysql/bin/mysqld: Normal shutdown 090923 19:53:56 mysqld started 090923 19:53:56 InnoDB: Started; log sequence number 0 5288400927 090923 19:53:56 [Note] /usr/local/mysql/bin/mysqld: ready for connections. Version: '5.0.77-log' socket: '/tmp/mysql50.sock' port: 3306 MySQL Community Server (GPL) 090929 22:38:48 [Note] /usr/local/mysql/bin/mysqld: Normal shutdown 090923 22:38:58 mysqld started 090923 22:38:58 InnoDB: Started; log sequence number 0 7102832776 090923 22:38:58 [Note] /usr/local/mysql/bin/mysqld: ready for connections. Version: '5.0.77-log' socket: '/tmp/mysql50.sock' port: 3306 MySQL Community Server (GPL) ...
(The above is just a sample, not the original file).
Well, the log says there’s a lot of normal shutdowns.
Looking at the script
Next it was time to look at the script which was supposed to verify MySQL was up and running – else wise start it. And it went something like this:
while [ 1 ];
if [`ps aux | grep mysqld | wc -l` -lt 2]
then /etc/init.d/mysql restart
...
sleep 60
The script was looking for all processes, then grepping for mysqld, counting number of lines in output. It expected 2 lines: one for the mysqld process, one for the grep mysqld process itself.
If you don’t know what’s wrong with this, a very brief explanation about how pipelines work in unix work it at place:
Pipelined processes do not execute one after another, or one before another. They are all executed at once. So, “ps aux | grep mysqld | wc -l” immediately spawns ps, grep, wc, then sets the standard output of one to the standard input of the other (most simplified description I could think of).
It is likely that grep will outrun ps in the race for starting up. So grep is started, already waiting for input from ps, which then notices both mysqld is alive, but also grep mysqld, lists them both along with other processes, to be filtered by grep, to be counted by wc (returning two lines count).
But this is just because ps is heavier than grep. It doesn’t have to be like that.
The less common scenario
Every once and again, ps outruns grep in the race for starting up. It would list the active processes, and no “grep” would appear in the listing. grep would then run, filtering the result of ps, then to be counted by wc — oh, there is only one line now! The script assumes this means mysqld is down (since it assumed grep is always there), and restarts mysqld.
So, the script which was supposed to take care of MySQL crashes, was actually causing them (though no crash occurred).
Better ways to test that MySQL is alive
Look for the pid file. This is the standard (that’s how the mysql service works). Look for the unix socket file. Both require that you parse the my.cnf file to learn where these files are.
If you’re reluctant to read the configuration files, other options are at hand. OK, look for the process; but use pgrep mysqld. No need to list all processes, then grep them.
And best way, that will set your mind at ease even if you’re worried that “mysql is running but not responding; it is stuck”: connect to MySQL and issue SELECT 1, SELECT NOW(), SELECT something. This would be the ultimate test that MySQL is up, listening, responding and valid.
It all sounded very technical, your description of piped processes running simultaneously but in my experience earlier ones always feed into later ones. The issue was more likely a logic error somewhere or simply not testing for the right number. A cut and paste of the actual faulty script would have been good.
On my Linux box a normal mysql install runs 2 processes and the grep process would take the minimum count to 3. So they should perhaps have been testing for less than 3 and not 2.
However my main point is that I seriously doubt the accuracy of your pipe explanation as it does not tally with my many scripting ‘adventures’.
Incidentally a nice way to discount the grep would be
[ `ps -ef | grep [m]ysqld | wc -l` -lt 2 ]
Instead of connecting and running a query you can use mysqladmin ping then check the exit status (zero for alive, non-zero for dead.) That’s the “official” way to check it, won’t have the overhead and the potential problems of running a query, and will avoid false positives from other processes with names that include mysql.
If really using pgrep, at least do it correctly: pgrep “^mysql$” (or whatever the process is named, taking care of quoting any special meaning characters.)
Really the best way is the Query, When you are really old you will realize that even when a process is hung it still shows up in ps listings. Do the query because it confirms the process is handling requests and not overloaded.
I wrote a little C program & an article on how to do exactly this at my last company..
http://knowledgehub.zeus.com/articles/2006/02/27/custom_monitors_for_mysql_in_c
Might be of help to anyone who as read this.
Cheers.