Ephemeral volumes aren't raided in all cases

[hannes-ucsc]

If you use an instance type with two ephemeral drives, cgcloud will install mdadm during create -IT and then take the image. So mdadm is present in the image and will not need to be installed again when running recreate from the image. But if you use an instance type with 0 or 1 ephemeral drives for create -IT, mdadm will not be needed at that time and so must be installed during recreate. If, between create and recreate an update to the mdadm package is published, apt will fail to install mdadm during recreate because the on-disk package index is stale and refers to an non-existant .deb.

The solution is to install mdadm unconditionally during the regular setup. There still is a chance that the early installation of mdadm at boot time will fail during recreate if the package index on the base image is stale, too.

[hannes-ucsc]

Slight twist. CloudInitBox does actually install mdadm on every first generation at boot time. But that could fail and so we should still schedule mdadm installation at setup time, as a fallback.

[hannes-ucsc]

And, BTW, cloud-init will mount one of the drives later on in the boot process. At that time it's too late for The dockerbox init script to do its thing with moving /var/lib/docker to ephemeral. This timing crap is killing me.