Comment 22 for bug 1654116

I was finally able to reproduce this issue with logging set to TRACE. The full logs of the deployment (basically /var/log/* from every unit and the bootstrap node) are available from https://private-fileshare.canonical.com/~fginther/juju/lp-1654116/unit-not-leader-with-trace-logs.tar.gz (it's nearly 50GB and only available within canonical, sorry). The logs for the bootstrap unit are under "landscape-0-inner-logs/bootstrap/var/log".

I did have to use machines with a higher core count to reproduce this issue (12 instead of 4 with 16GB of RAM). This helps support the theory that the unit is over-loaded. A 12-core system will spawn a lot more openstack API processes then a 4-core system due to the worker-multiplier option. This detail, combined with the fact that we are smooshing multiple services into the same physical machine with LXD containment means that a physical machine will be running multiple dozens of additional processes in the same memory footprint.