Comment 2 for bug 1549219

I proposed a revert of the fix because it partially serialises resource creation on non-convergence stacks: https://review.openstack.org/#/c/313122/

At the summit we discussed a possible solution by allowing tasks to indicate to the scheduler how long to wait before polling them again, by doing e.g. "yield 5" (to sleep for 5s) instead of just "yield". This could be useful for many resource types, and of course it could vary dynamically so we could do things like exponential backoffs.