MySQL TMPDIR on tmpfs doens't work with ALTER TABLE on large tables

[Matthew Treinish]

Trove deploys mysql's tmpdir onto /tmp with a tmpfs however this causes issues with MySQL when you do an ALTER TABLE on a large table if you don't have enough ram the operation will fail. There is no way to workaround this because the end user doesn't know the fs layout of the instance. (I'm also not sure the trove api exposes this config option) My only way to workaround this issue has been to rewrite the sql to avoid using an alter table on these tables.

We should add an option to deploy the tmpdir on real disk so that people can run operations that require more space than you have free ram on the instance.

[Amrith Kumar (amrith)]

This is another manifestation of the fact that only one cinder volume is attached to an instance.

[Matthew Treinish (treinish)]

FWIW, you can still make this work with a single volume, you just need a subdir on the 1 volume to dedicate to tmp for mysql.

[Amrith Kumar (amrith)]

I've tried the change that Matt just suggested by faking the template. Turns out that the directory must exist for this to work. This will therefore need some changes to the guest; and since it must get configuration from the host, the ripple is non-trivial.

Working on a fix, it isn't just a change to the template.