Some time Nodepool fails to keep track of an instance due to random errors. The instance is spawned in OpenStack but is not taken in account in Nodepool internal database. There is then a discrepancy between the OpenStack quota and the vision of Nodepool.
Turns out upstream has a patch to delete the leaked instances https://review.openstack.org/#/c/190827/ THAT PATCH HAS FOLLOW UP.
Should look at the serie of patches and see whether we can backport them to our version.