Open Stack

On December 13, 2014 at 3:26:34 AM, Henry (henry4hly at gmail.com) wrote:
Hi Morgan,

A good question about keystone.

In fact, keystone is naturally suitable for multi-region deployment. It has only REST service interface, and PKI based token greatly reduce the central service workload. So, unlike other openstack service, it would not be set to cascade mode.


I agree that Keystone is suitable for multi-region in some cases, I am still concerned from a security standpoint. The cascade examples all assert a *global* tenant_id / project_id in a lot of comments/documentation. The answer you gave me doesn’t quite address this issue nor the issue of a disparate deployment having a wildly different role-set or security profile. A PKI token is not (as of today) possible to use with a Keystone (or OpenStack deployment) that it didn’t come from. This is like this because Keystone needs to control the AuthZ for it’s local deployment (same design as the keystone-to-keystone federation). 

So I have to direct questions:

* Is there something specific you expect to happen with the cascading that makes resolving a project_id to something globally unique or am I mis-reading this as part of the design? 

* Does the cascade centralization just ask for Keystone tokens for each of the deployments or is there something else being done? Essentially how does one work with a Nova from cloud XXX and cloud YYY from an authorization standpoint?

You don’t need to answer these right away, but they are clarification points that need to be thought about as this design moves forward. There are a number of security / authorization questions I can expand on, but the above two are the really big ones to start with. As you scale up (or utilize deployments owned by different providers) it isn’t always possible to replicate the Keystone data around.

Cheers,
Morgan

Best regards
Henry

Sent from my iPad

On 2014-12-13, at 下午3:12, Morgan Fainberg <morgan.fainberg at gmail.com> wrote:



On Dec 12, 2014, at 10:30, Joe Gordon <joe.gordon0 at gmail.com> wrote:



On Fri, Dec 12, 2014 at 6:50 AM, Russell Bryant <rbryant at redhat.com> wrote:
On 12/11/2014 12:55 PM, Andrew Laski wrote:
> Cells can handle a single API on top of globally distributed DCs.  I
> have spoken with a group that is doing exactly that.  But it requires
> that the API is a trusted part of the OpenStack deployments in those
> distributed DCs.

And the way the rest of the components fit into that scenario is far
from clear to me.  Do you consider this more of a "if you can make it
work, good for you", or something we should aim to be more generally
supported over time?  Personally, I see the globally distributed
OpenStack under a single API case much more complex, and worth
considering out of scope for the short to medium term, at least.

For me, this discussion boils down to ...

1) Do we consider these use cases in scope at all?

2) If we consider it in scope, is it enough of a priority to warrant a
cross-OpenStack push in the near term to work on it?

3) If yes to #2, how would we do it?  Cascading, or something built
around cells?

I haven't worried about #3 much, because I consider #2 or maybe even #1
to be a show stopper here.

Agreed

I agree with Russell as well. I also am curious on how identity will work in these cases. As it stands identity provides authoritative information only for the deployment it runs. There is a lot of concern I have from a security standpoint when I start needing to address what the central api can do on the other providers. We have had this discussion a number of times in Keystone, specifically when designing the keystone-to-keystone identity federation, and we came to the conclusion that we needed to ensure that the keystone local to a given cloud is the only source of authoritative authz information. While it may, in some cases, accept authn from a source that is trusted, it still controls the local set of roles and grants. 

Second, we only guarantee that a tenan_id / project_id is unique within a single deployment of keystone (e.g. Shared/replicated backends such as a percona cluster, which cannot be when crossing between differing IAAS deployers/providers). If there is ever a tenant_id conflict (in theory possible with ldap assignment or an unlucky random uuid generation) between installations, you end up with potentially granting access that should not exist to a given user. 

With that in mind, how does Keystone fit into this conversation? What is expected of identity? What would keystone need to actually support to make this a reality?

I ask because I've only seen information on nova, glance, cinder, and ceilometer in the documentation. Based upon the above information I outlined, I would be concerned with an assumption that identity would "just work" without also being part of this conversation. 

Thanks,
Morgan 
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________  
OpenStack-dev mailing list  
OpenStack-dev at lists.openstack.org  
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20141213/141eafc1/attachment.html>