Hello,
I’m actually revisiting a problem I originally just worked around… maybe 18 to 24 months ago. I just now did some searching of the list, but there were just pages and pages for most searches I tried so apologies if this is a solved thing and I missed it.
What I do now to deploy to cloud hosts is have the playbook call a role named rax/add_rax or ec2/add_ec2. In that role I use the appropriate cloud module to create the hosts (or discover they are already there, of course), then I run the add_host loop on the result. The next role that it runs is named gather_custom_facts and it detects what provider is in use and uses the appropriate “get_facts” type module to set facts named ‘private_ip’ and ‘public_ip’ for each host. (not to imply that hosts from different providers are in the same ansible-playbook run; that is still on a Someday list somewhere)
I do it this way so that plays work on both clouds with the same names for these IP variables (or facts in this case, I suppose). The problem is that I feel like it adds quite a bit of overhead (measured both in computing resources and to the learning curve for new team members). To just have a single ping: task you’re already two or three plays in just to have inventory.
It seems like the easiest thing would be to use set_fact in the cloud provider-specific create/“persist” role and have it still exist in the calling playbook. But, of course, the hosts aren’t actually in inventory yet (and even if they are the set_fact would still be a local_action). I’ve revisited this once or twice in the past and never really got anywhere so here I am. 
Any ideas?
Thank you,
Mark
P.S.: For entertainment value only, I was wondering also if it’d be possible to write an inventory script that simply calls ansible-playbook again and spits out some inventory-compatible json after running these two or three host gathering roles.