Bringing conversation for domains in Keystone to the broader mailing
lists.
On Oct 26, 2012, at 5:18 AM, Dolph Mathews <dolph.mathews@xxxxxxxxx
<mailto:dolph.mathews@xxxxxxxxx>> wrote:
I think this discussion would be great for both mailing lists.
-Dolph
On Fri, Oct 26, 2012 at 5:18 AM, Henry Nash <henry.nash@xxxxxxx
<mailto:henry.nash@xxxxxxx>> wrote:
Hi
<Not sure where best to have this discussion - here, as a comment
to the v3api doc, or elsewhere - appreciate some guidance and
will transfer this to the right place>
At the Summit we started a discussion on whether things like user
name, tenant name etc. should be globally unique or unique within
a domain. I'd like to widen that discussion to try and a) agree
a direction, b) agree some changes to our current spec. Here's my
view as an opening gambit:
- When a Keystone instance is first started, there is only one,
default, Domain. The Cloud Provider does not need to create any
new domains, all projects can exist in this default domain, as
will the users etc. There is one, global, name space. Clients
using the v2 API will work just fine.
+1
Very much what we were thinking for the initial implemenation and
rollout to make it backwards "compatible" with the V2 (non-domain)
core API
- If the Cloud Provider wants to provide their customers with
regions they can administer themselves and be self-contained,
then they create a Domain for each customer. It should be
possible for users/roles to be scoped to a Domain so that
(effectively) administrative duties can be delegated to some
users in that Domain. So far so good - all this can be done with
the v3 API.
Not clear on if you're referring to endpoint regions, or just
describing domain isolation?
I believe you're describing the key use cases behind the domains
mechanism to begin with - user and project partitioning to allow for
administration of those to be clearly "owned" and managed
appropriately.
- We still have work to do to make sure items in other OS
projects that reference tenants (e.g. Images) can take a Domain
or Project ID, but we'll get to that soon enough
Everything will continue to work with projects, but once middleware
starts providing a DOMAIN_ID and DOMAIN_NAME to the underlying
service, it'll be up to them to take advantage of it. Images per
domain is an excellent example use case.
- However, Cloud Providers want to start enabling enterprise
customers to run more and more of the workloads in OpenStack
clouds - over and above, the smaller sized companies that are
doing this today. For this to work, the encapsulation of a
Domain need, I think, to be able to be stricter - and this is
where the name space comes into play. I think we need to allow
for a Domain to have its own namespace (i.e. users, roles,
projects etc.) as an option. I see this as a first step to
allowing each Domain to have its own AuthZ/N service (.e.g
external ldap owned and hosted by the customer who will be using
the Domain)
Implementation:
- A simplistic version would just allow a flag to specified on
Domain creation that said whether this a "private" or "shared"
Domain. Shared would use the current global name space (and
probably be the default for compatibility reasons).
I like the direction of this -- need to digest implications :)
I like the idea conceptually - but let's be clear on the implications
to the end users:
Where we're starting is preserving a global name space for project
names and user names. Allowing a mix of segregated and global name
spaces imposes a burden of additional data being needed to uniquely
place authentication and authorization.
We've been keeping to 2 key pieces of info (username, password) to get
authenticated - and then (via CLI or Horizon dashboard) you can choose
from a list of protential projects and carry on. In most practical
circumstances, any user working primarily from the CLI is already
providing 3-4 pieces of information:
* username
* password
* tenant name
* auth_url
to access and use the cloud.
By allowing domains to be their own namespaces, we're adding another
element that will be absolutely required to identify the person
authenticating:
* domain name
implying a cascade of changes to the user experience all the way down
through horizon.
- A more flexible approach would be to allow the specification of
where the various sub-services of Keystone (e.g. AuthN/Z, Service
Catalogue, Resources (i.e Users, Projects)) are hosted. The
defaults would all point back to the default domain (i.e. are
global and shared), but instead could be specified as "self"
(I.e. the new Domain), or, in the future, some other external
location, e.g. for a remote ldap.
- As an aside, this multi-name space model could also allow the
Cloud Provider their own name space, separate from their
customers - i.e. they will have a need to create admins who can
just create domains and on-board customers into those domains.
These users & roles could exist in the default domain, while all
the customers' users/roles exist solely within their own domains.
- One potential problem I do see is with roles. Today, the role
name is, if I understand it correctly, a kind of shared secret
between, other services and Keystone - e.g. it is the actual name
of a given role, say "ProjectAdmin" , that must match in, say,
the Nova policy file and the role assignment in Keystone (please
correct me if I have this wrong).
You're 100% correct.
How would that work if the role names were not unique across
Domains?
Not that we would want admins to ever see Role ID's, or edit policy
files with role ID's, but they provide a potential solution.
The different role names would need to be accounted for in the policy
files the way they're set up today - the enforcement there is all at
the service level. There's no current provision for evaluating policy
differently based on domain. While that's possible, it sounds like a
tremendous cascade of additional complication, as the policy, and
roles, are all set up and managed by deployers.
I think this might be an interesting addition in the future, but want
to keep the initial implementation and roll-out of the policy
mechanisms and domains consistent and simple for a first roll-out
iteration.
What is the desired functionality for a Cloud Provider wanting to
give their enterprise customers this level of flexibility - will
they have dedicated Nova endpoints anyway? Sounds too rigid.
This might tie into another bp we are working on at IBM in terms
of using Availability zones to allow Cloud Providers to divide up
their compute resources in a more flexible way.
- Finally, I wanted to raise the subject of whether we should
make it a goal to remain compatible with the v2 API /once the
cloud provider starts creating additional domains/.
Joe and I briefly discussed this at the summit. As a migration to v3,
we'd obviously be creating the default domain and mapping all
existing users/projectse/etc to it. I'd be fine if the v2
implementation ONLY interacted with resources in that default domain;
i.e. if you want to use domains, upgrade to a v3 client.
As stated above, if just the default domain is being used, then
fine. And while I agree that, technically, the v2 API should
still work with the above if all the other domains point back to
the default domain for their sub-services - it feels overly
flexible (and maybe wrong conceptually) to support v2 semantics
across a multi-domain installation.
+1
Interested in everyone else's view.
Henry