MySQL on Linux – Managing Multiple Processes

MySQL should include a number of starting point configuration files for my.cnf that can adjust the memory usage. The files are named my-small.cnf, my-medium.cnf, my-large.cnf, and my-huge.cnf.

There is an online MySQL memory usage calculator at http://www.mysqlcalculator.com/. By comparing the starting point files and using the memory usage calculator you should be able to adjust your my.cnf configuration file to accomodate your server's memory constraints.

For more information about what each parameter does, you can reference the MySQL documentation available online. Here is a link to the MySQL documentation for MySQL server 5.1: http://dev.mysql.com/doc/refman/5.1/en/server-system-variables.html

It seems theoretically possible that table could still dump properly if the corruption were in the indexes, which aren't dumped.

It should not be possible for anything in your configuration to cause MySQL to crash with a Signal 11, a segmentation fault.

I've been staring at this for a while, now, and I haven't come up with answers... just questions (in no particular order):

• have you run memory diagnostics on your server? You mentioned that you "tried to move hardware" but you also mention having not tried a restore of your dump, so I'm not clear exactly what you tried moving. Resist the temptation to think "it can't be that." Test the memory.
• is your system using any swap space at all? Hopefully not -- but if (and only if) it is, then you should reduce the innodb_buffer_pool_size to the point that it isn't ... because there's not really a point in buffering to memory that gets swapped, and the swap partition could be introducing problems. This one is a stretch, but worth eliminating, I think.
• is this a problem that occurred after an upgrade to 5.5.28 or is this a new application or deployment?
• if it's new, have you tried replicating the problem with MySQL 5.6?
• is partitioning involved? That means touching more code.
• are you using a binary distribution of MySQL that you downloaded from Oracle (tar/deb/rpm)? Or is it from Ubuntu (I always use generic tar binaries, so I don't know what the current version of MySQL 5.5 is, in 12.04LTS) or another source? Or compiled from source code?
• are you using any unusual plugins or UDFs?

This could be a bug, but when you hear the sound of hooves, suspect horses before zebras (at least where I come from).

update (from comments):

"Another" memory bug?

Checking the memory would be the first thing I would try, for sure.

The snapshots should be getting you a reliable backup, I agree, but if there's any kind of binary wierdness going on in your files, it would be perfectly replicated. It will take some time, but restoring to a fresh system using mysqldump files would be a better test, since all of the table structures would all be absolutely rebuilt from scratch. Since the table structures seem to be valid, it may be unlikely that this will change anything, but it feels like you're kind of at the point where every possibility needs to be pinned down... clearly, what you're seeing should not be happening.

For a new test system, though, I would install the server using the "Linux - Generic 2.6 (x86, 64-bit), Compressed TAR Archive" package from the download site. Download the tarball, verify it's md5 checksum, then tar xvzf it into /usr/local and symlink the resulting directory to /usr/local/mysql. (I think Ubuntu still puts it in /var/lib/mysql, so you can probably do this even without removing the distro version, as long as you don't have the other copy running). Then move the "data" directory from inside /usr/local/mysql to whatever partition it needs to live on (if different), and symlink it back into /usr/local/mysql/data. Put your config file at /usr/local/mysql/my.cnf and pass that as the first option ... using --defaults-file=/usr/local/mysql/my.cnf when using the install scripts and when starting the server -- this will cause any other my.cnf's (such as those in /etc) to not be read.

The rest of the setup is pretty straightforward. It's more work, but it completely eliminates the "black box" of using the package manager. The real motivation here, though, is that the disto packages may have been compiled from source, and the resulting binaries could have slight variations from the "official" Oracle binaries.