Open Stack

Hi Eugene,

Thanks for the provided detail. See my comments below.

The point is to be able to share IP address, it really means that two VIPs(as we understand them in current model) need to reside within same backend (technically they need to share neutron port).

Aren't we leaking some implementation detail here? Why is it that 2 VIPs using the same IP address have to be implemented on the same backend? Isn't this a driver/technology capability? If a certain driver *requires* that VIPs sharing the same IP address have to be on the same "backend" (whatever a "backend" means), it just needs to ensure that this is the case, but another driver might be able to support VIPs sharing the same IP to be on different backends. The user really shouldn't care. Did I miss some important detail? It feels like it, so please be patient with me :)

I'm sorry this all creates so much confusion.
In order to understand why we need additional entity, you need to keep in mind the following things:
 1) We have a notion of root object. From user perspective it represents logical instance, from implementation perspective it also represents how that instance is mapped to a backend (agent, device), which flavor/provider/driver it has, etc
 2) We're trying to change vip-pool relationship to m:n, if vip or pool remain the root object, that creates inconsistency because root object can be connected to another root object with different parameters.

I'm not against introducing a wrapper entity that correlates the different config objects that logically make up one LB config, but I don't think it is needed from the logical object model pov IMO. Yes, it might make the implementation of the object model for some drivers easier, and I'm OK with having it, if it helps. But strictly speaking it is not needed, because a driver doesn't have to choose a backend when the pool is created or when a vip is created, if it doesn't have enough info yet. It can wait until a vip/pool are created and attached to each other, then it would have a clearer idea of the backends eligible to host that whole LB configuration. Another driver though, might be able to perform the configuration on its "backend" straight-away on each API call, and still be able to comply with the object model.

Youcef

On Thu, Feb 27, 2014 at 5:20 AM, Youcef Laribi <Youcef.Laribi at citrix.com<mailto:Youcef.Laribi at citrix.com>> wrote:
Hi Eugene,

1) In order to allow real multiple 'vips' per pool feature, we need the listener concept.
It's not just a different tcp port, but also a protocol, so session persistence and all ssl-related parameters should move to listener.

Why do we need a new 'listener' concept? Since as Sam pointed out, we are removing the reference to a pool from the VIP in the current model, isn't this enough by itself to allow the model to support multiple VIPs per pool now?

lb-pool-create  .... -->$POOL-1
lb-vip-create .....$VIP_ADDRESS,$TCP_PORT, default_pool=$POOL-1... --> $VIP-1
lb-vip-create .....$VIP_ADDRESS,$TCP_PORT, default_pool=$POOL-1... --> $VIP-2


Youcef





From: Eugene Nikanorov [mailto:enikanorov at mirantis.com<mailto:enikanorov at mirantis.com>]
Sent: Wednesday, February 26, 2014 1:26 PM
To: Samuel Bercovici
Cc: OpenStack Development Mailing List (not for usage questions)

Subject: Re: [openstack-dev] [Neutron][LBaaS] Object Model discussion

Hi Sam,

I've looked over the document, couple of notes:

1) In order to allow real multiple 'vips' per pool feature, we need the listener concept.
It's not just a different tcp port, but also a protocol, so session persistence and all ssl-related parameters should move to listener.

2) ProviderResourceAssociation - remains on the instance object (our instance object is VIP) as a relation attribute.
Though it is removed from public API, so it could not be specified on creation.
Remember provider is needed for REST call dispatching. The value of provider attribute (e.g. ProviderResourceAssociation) is result of scheduling.

3) As we discussed before, pool->vip relation will be removed, but pool reuse by different vips (e.g. different backends) will be forbidden for implementation simplicity, because this is definitely not a priority right now.
I think it's a fair limitation that can be removed later.

On workflows:
WFs #2 and #3 are problematic. First off, sharing the same IP is not possible for other vip for the following reason:
vip is created (with new model) with flavor (or provider) and scheduled to a provider (and then to a particular backend), doing so for 2 vips makes address reuse impossible if we want to maintain logical API, or otherwise we would need to expose implementation details that will allow us to connect two vips to the same backend.

On the open discussion questions:
I think most of them are resolved by following existing API expectations about status fields, etc.
Main thing that allows to go with existing API expectations is the notion of 'root object'.
Root object is the object which status and admin_state show real operability of the configuration. While from implementation perspective it is a mounting point between logical config and the backend.

The real challenge of model #3 is ability to share pools between different VIPs, e.g. between different flavors/providers/backends.
User may be unaware of it, but it requires really complex logic to handle statistics, healthchecks, etc.
I think while me may leave this ability at object model and API level, we will limit it, as I said previously.

Thanks,
Eugene.


On Wed, Feb 26, 2014 at 9:06 PM, Samuel Bercovici <SamuelB at radware.com<mailto:SamuelB at radware.com>> wrote:
Hi,

I have added to the wiki page: https://wiki.openstack.org/wiki/Neutron/LBaaS/LoadbalancerInstance/Discussion#1.1_Turning_existing_model_to_logical_model that points to a document that includes the current model + L7 + SSL.
Please review.

Regards,
                -Sam.


From: Samuel Bercovici
Sent: Monday, February 24, 2014 7:36 PM

To: OpenStack Development Mailing List (not for usage questions)
Cc: Samuel Bercovici
Subject: RE: [openstack-dev] [Neutron][LBaaS] Object Model discussion

Hi,

I also agree that the model should be pure logical.
I think that the existing model is almost correct but the pool should be made pure logical. This means that the vip <---->pool relationships needs also to become any to any.
Eugene, has rightfully pointed that the current "state" management will not handle such relationship well.
To me this means that the "state" management is broken and not the model.
I will propose an update to the state management in the next few days.

Regards,
                -Sam.




From: Mark McClain [mailto:mmcclain at yahoo-inc.com]
Sent: Monday, February 24, 2014 6:32 PM
To: OpenStack Development Mailing List (not for usage questions)
Subject: Re: [openstack-dev] [Neutron][LBaaS] Object Model discussion


On Feb 21, 2014, at 1:29 PM, Jay Pipes <jaypipes at gmail.com<mailto:jaypipes at gmail.com>> wrote:

I disagree on this point. I believe that the more implementation details
bleed into the API, the harder the API is to evolve and improve, and the
less flexible the API becomes.

I'd personally love to see the next version of the LBaaS API be a
complete breakaway from any implementation specifics and refocus itself
to be a control plane API that is written from the perspective of the
*user* of a load balancing service, not the perspective of developers of
load balancer products.

I agree with Jay.  We the API needs to be user centric and free of implementation details.  One of my concerns I've voiced in some of the IRC discussions is that too many implementation details are exposed to the user.

mark


_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org<mailto: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/20140227/eb945f54/attachment.html>