Yep. It's fun when your entire etcd cluster dies at the same time (because you started a global unit that pulls a huge docker container).

You'd think they have safeguards to automatically rebalance on certain conditions (metadata filling up, etc) but no, you have to babysit btrfs.