Openstack provider should work without object-store

[Martin Packman]

Currently the Openstack provider depends on a working object-store endpoint in the same region as compute. With all the uses for blob storage except state server tracking now implemented through other means, the provider should support deployments with no swift or ceph.

The way around this in pyjuju was with a hybrid provider with storage provided by S3 instead, and there has been discussion in the past of separating object store and general compute functions of providers.

The easiest path forward would be to use metadata to tag all state servers, and use that to implement the current out of band lookup that uses storage.

This is currently required to support Dreamhost, which doesn't offer object store.

[Eric Snow (ericsnowcurrently)]

FWIW, I believe the approach for which Martin is advocating is the one which we took with the GCE provider.

[Ian Booth (wallyworld)]

The basic building blocks are in place to allow this issue to be solved, especially with the recent work to support instance tagging landing in 1.24. The implementation work is (on paper) not complicated, but there's potential for lots of little issues to pop up in various tests etc that will require fixing along the way. The work to put in place upgrade steps for existing environments is also potentially non trivial. I would scope this bug as akin to a mini-feature taking to upgrade work etc into account.

[Andrew Wilkins (axwalk)]

Martin, if you have an account on Dreamhost I'd love to know if this works now. We should not be using Swift for anything except for upgrades now, so a bootstrap off master should be fine.

[Martin Packman (gz)]

Fix not complete at present, we're still requiring object-store to exist in the keystone catalogue.

[Andrew Wilkins (axwalk)]

No longer requiring the object-store service in keystone.