The subtlety you are referring to in #2 is that we moved from puppet for deployments to puppet to configure template nodes. We baked AMIs from those template nodes and deployed those, drastically shortening our time from boot to serving traffic.