Comment 12 for bug 1170103

Performance improvements in 5.5.30-30.2 is the main suspect. We've been trying to reproduce the memory leak on internal tests, but have not succeeded so far. So I have a request to everyone affected by this issue: if you can test a custom-built binary with additional diagnostics, please help us track down this problem.

The additional diagnostics is only available in a source branch currently, but we are going to provide a tarball shortly.

Instructions to build:

bzr branch lp:~percona-core/percona-server/5.5.30-30.2-memory-diagnostics
cd 5.5.30-30.2-memory-diagnostics/Percona-Server
cmake . -DBUILD_CONFIG=mysql_release -DFEATURE_SET=community -DWITH_EMBEDDED_SERVER=OFF -DEXTRA_VERSION="-30.2-memory-diagnostic"

make -j6

You can temporarily replace the installed mysqld binary with the newly built one located in sql/ (don't forget to backup the original binary). Additional diagnostics should not have any measurable overhead. It adds 2 status variables: innodb_read_views_memory and innodb_descriptors_memory.

So when the excessive memory usage can be observed after a few hours, please get the output of

SHOW GLOBAL STATUS LIKE 'innodb%memory';

Thank you.