Open Stack

tl;dr We want to implement a new system which we’re calling an Aggregator which is based on Horizon and Keystone, and that can provide access to virtual Regions from multiple independent OpenStack providers. We plan on developing this system as a project in Stackforge, but we need help right now in identifying any unexpected dependencies.
 
 
 
For the last 6-7 years, there has been great interest in the potential for various business models involving multiple clouds and/or cloud providers. These business models include but are not limited to, federation, reseller, broker, cloud-bursting, hybrid and intercloud. The core concept of this initiative is to go beyond the simple dyadic relationship between a cloud service provider and a cloud service consumer to a more sophisticated “supply chain” of cloud services, dynamically configured, and operated by different business entities. This is an ambitious goal, but there is a general sense that OpenStack is becoming stable and mature enough to support such an undertaking.
 
Until now, OpenStack has focused on the logical abstraction of a Region as the basis for cloud service consumption. A user interacts with Horizon and Keystone instances for a Region, and through them gains access to the services and resources within the specified Region. A recent extension of this model has been to share Horizon and Keystone instances between several Regions. This simplifies the user experience and enables single sign on to a “single pane of glass”. However, in this configuration all of the services, shared or otherwise, are still administered by a single entity, and the configuration of the whole system is essentially static and centralized.
 
We’re proposing that the first step in realizing the Cloud Service Federation use cases is to enable the administrative separation of interface and region: to create a new system which provides the same user interface as today - Horizon, Keystone - but which is administratively separate from the Region(s) which provide the actual IaaS resources. We don’t yet have a good name for this system; we’ve been referring to it as the “Aggregator”. It includes slightly-modified Horizon and Keystone services, together with a subsystem which configures these services to implement the mapping of “Aggregator Regions” to multiple, administratively independent, “Provider Regions”. Just as the User-Provider relationship in OpenStack is “on demand”, we want the Aggregator-Provider mappings to be dynamic, established by APIs, rather than statically configured. We want to achieve this without substantially changing the user experience, and with no changes to applications or to core OpenStack services. The Aggregator represents an additional way of accessing a cloud; it does not replace the existing Horizon and Keystone.
 
The functionality and workflow is as follows: A user, X, logs into the Horizon interface provided by Aggregator A. The user sees two Regions, V1 and V2, and selects V1. This Region is actually provided by OpenStack service provider P; it’s the Region which P knows as R4.  X now creates a new tenant project, T. Leveraging the Hierarchical Multitenancy work in Kilo, T is actually instantiated as a subproject of a Domain in R4, thus providing namespace isolation and quota management. Now X can deploy and operate her project T as she is used to, using Horizon, CLI, or other client-side tools. UI and API requests are forwarded by the Aggregator to P’s Region R4. [I’ll transfer this to the wiki and add diagrams.]
 
As noted, the high-level workflow is relatively straightforward, but we need to understand two important concepts. First, how does P make R4 available for use by A? Are all of the services and resources in R4 available to A, or can P restrict things in some way? What’s the lifecycle of the relationship? Secondly, how do we handle identity? Can we arrange that same identity provider is used in the Aggregator and in the relevant domain within R4? One answer to these issues is to introduce what Mark Shuttleworth called “virtual Regions” at his talk in Paris; add a layer which exposes a Domain within a Region (with associated IDM, quotas, and other policies) as a browsable, consumable resource aggregate. To implement this, P can add a new service to R4, the Virtual Region Manager, with the twin roles of defining Virtual Regions in terms of physical Region resources, and managing the service provider side of the negotiation with the Aggregator when setting up Aggregator-to-provider mappings. The intention is that the Virtual Region Manager will be a non-disruptive add-on to an existing OpenStack deployment.
 
Obviously there are many more issues to be solved, both within OpenStack and outside (especially in the areas of OSS and BSS). However, we have the beginnings of an architecture which seems to address many of the interesting use cases. The immediate question is how to implement it within the OpenStack process. It’s too complicated for any one of the existing OpenStack projects to take it on; large-scale proposals rarely do well in this community. So we intend to start this work as a new Stackforge project, with the objective of completing a first version during the Liberty cycle. To meet this goal, we must identify all of the features or fixes that we’ll need in other OpenStack projects, and submit them for the Liberty cycle. (This is time critical!) Hopefully each of these changes will be small enough that it can be accepted without too much debate. The two projects most affected will be Keystone and Horizon; in many cases, we will need to replace a static configuration with something more dynamic. 
 
We think the time is right to begin this work. The Keystone team paved the way during Kilo with their work on Hierarchical Multitenancy, and during the Liberty cycle we expect to see work in other projects that will build on this. (Hierarchical quotas, aggregated Ceilometer queries, that kind of thing). By starting the Cloud Service Federation project within Stackforge, we hope to avoid the “complexity antibodies”. However we really need to get this proposal in front of OpenStack developers, because it’s critically important to identify unexpected dependencies as soon as possible. For this reason, we’d like to discuss it in Vancouver as part of the cross-project track in the Design Summit.
 
Geoff Arnold
Cisco Cloud Services
geoff(at)geoffarnold.com
geoarnol(at)cisco.com
@geoffarnold