Yay, due to some unknown fix arm64 instances have stopped auto-destructing themselves (https://launchpad.net/bugs/1531768
). So I took another attempt at moving armhf autopkgtesting from our manually maintained Calxeda nodes into Scalingstack.
I now have a cloud-init userdata script which sets up a large btrfs partition for LXD (as that's vastly faster than on ext4), configure the LXD bridge, cron job to regularly build autopkgtest images for all supporte releases etc., and it works quite nicely. I've been running a test loop on this over night using autopkgtest's remote lxd runner, without a single hiccup.
This also uses socat to forward the local Unix socket to the TCP port, instead of lxd's builtin SSL web server. autopkgtest sends a lot of "lxc exec" calls, and as each of those would have to go through the whole SSL negotiation, running tests is achingly slow. With that trick (thanks to +Tyler Hicks
) this now runs very fast and smooth, even though all the operations happen on a remote LXD server!
(Running remotely is necessary due to how our different clouds (Prodstack vs. Scalingstack) are organized and firewalled from each other.)