Comment 12 for bug 1645422

This is a legitimate issue for us as well (currently contributing to descent into madness) - without this, failed nodes get re-numbered, and targeted placements of units aiming to be on those nodes get wonky despite the --map-machines flag on iterative overlays (up to 5 here for openstack ha with vault and a bunch of other stuff).
The problem also exists with LXD units - Juju has no way to retry those, and that's entirely within its scope of control.