Kyle, What happened to the long-term potential goal of ML2 driver APIs becoming neutron's core APIs? Do we really want to encourage new monolithic plugins?

ML2 is not a control plane - its really just an integration point for control planes. Although co-existence of multiple mechanism drivers is possible, and sometimes very useful, the single-driver case is fully supported. Even with hierarchical bindings, its not really ML2 that controls what happens - its the drivers within the framework. I don't think ML2 really limits what drivers can do, as long as a virtual network can be described as a set of static and possibly dynamic network segments. ML2 is intended to impose as few constraints on drivers as possible.

My recommendation would be to implement an ML2 mechanism driver for OVN, along with any needed new type drivers or extension drivers. I believe this will result in a lot less new code to write and maintain.

Also, keep in mind that even if multiple driver co-existence doesn't sound immediately useful, there are several potential use cases to consider. One is that it allows new technology to be introduced into an existing cloud alongside what previously existed. Migration from one ML2 driver to another may be a lot simpler (and/or flexible) than migration from one plugin to another. Another is that additional drivers can support special cases, such as bare metal, appliances, etc..


On 2/24/15 11:11 AM, Kyle Mestery wrote:
On Tue, Feb 24, 2015 at 3:19 AM, Salvatore Orlando < <>> wrote:

    On 24 February 2015 at 01:34, Kyle Mestery <
    <>> wrote:

        Russel and I have already merged the initial ML2 skeleton
        driver [1].

        The thinking is that we can always revert to a non-ML2 driver
        if needed.

    If nothing else an authoritative decision on a design direction
    saves us the hassle of going through iterations and discussions.
    The integration through ML2 is definitely viable. My opinion
    however is that since OVN implements a full control plane, the
    control plane bits provided by ML2 are not necessary, and a plugin
    which provides only management layer capabilities might be the
    best solution. Note: this does not mean it has to be monolithic.
    We can still do L3 with a service plugin.
    However, since the same kind of approach has been adopted for ODL
    I guess this provides some sort of validation.

To be honest, after thinking about this last night, I'm now leaning towards doing this as a full plugin. I don't really envision OVN running with other plugins, as OVN is implementing it's own control plane, as you say. So the value of using ML2 is quesitonable.

        I'm not sure how useful having using OVN with other drivers
        will be, and that was my initial concern with doing ML2 vs.
        full plugin. With the HW VTEP support in OVN+OVS, you can tie
        in physical devices this way. Anyways, this is where we're at
        for now. Comments welcome, of course.

    That was also kind of my point regarding the control plane bits
    provided by ML2 which OVN does not need. Still, the fact that we
    do not use a function does not make any harm.
    Also i'm not sure if OVN needs at all a type manager. If not, we
    can always implement a no-op type manager, I guess.

See above. I'd like to propose we move OVN to a full plugin instead of an ML2 MechanismDriver.





        On Mon, Feb 23, 2015 at 4:09 PM, Kevin Benton
        < <>> wrote:

            I want to emphasize Salvatore's last two points a bit
            more. If you go with a monolithic plugin, you eliminate
            the possibility of heterogenous deployments.

            One example of this that is common now is having the
            current OVS driver responsible for setting up the vswitch
            and then having a ToR driver (e.g. Big Switch, Arista,
            etc) responsible for setting up the fabric. Additionally,
            there is a separate L3 plugin (e.g. the reference one,
            Vyatta, etc) for providing routing.

            I suppose with an overlay it's easier to take the route
            that you don't want to be compatible with other networking
            stuff at the Neutron layer (e.g. integration with the 3rd
            parties is orchestrated somewhere else). In that case, the
            above scenario wouldn't make much sense to support, but
            it's worth keeping in mind.

            On Mon, Feb 23, 2015 at 10:28 AM, Salvatore Orlando
            < <>> wrote:

                I think there are a few factors which influence the
                ML2 driver vs "monolithic" plugin debate, and they
                mostly depend on OVN rather than Neutron.
                From a Neutron perspective both plugins and drivers
                (as long at they live in their own tree) will be
                supported in the foreseeable future. If a ML2 mech
                driver is not the best option for OVN that would be ok
                - I don't think the Neutron community advices
                development of a ML2 driver as the preferred way for
                integrating with new backends.

                The ML2 framework provides a long list of benefits
                that mechanism driver developer can leverage.
                Among those:
                - The ability of leveraging Type drivers which
                relieves driver developers from dealing with network
                segment allocation
                - Post-commit and (for most operations) pre-commit
                hooks for performing operation on the backend
                - The ability to leverage some of the features offered
                by Neutron's built-in control-plane such as L2-population
                - A flexible mechanism for enabling driver-specific
                API extensions
                - Promotes modular development and integration with
                higher-layer services, such as L3. For instance OVN
                could provide the L2 support for Neutron's built-in L3
                control plane
                - The (potential afaict) ability of interacting with
                other mechanism driver such as those operating on
                physical appliances on the data center
                - <add your benefit here>

                In my opinion OVN developers should look at ML2
                benefits, and check which ones apply to this specific
                platform. I'd say that if there are 1 or 2 checks in
                the above list, maybe it would be the case to look at
                developing a ML2 mechanism driver, and perhaps a L3
                service plugin.
                It is worth nothing that ML2, thanks to its type and
                mechanism driver provides also some control plane
                capabilities. If those capabilities are however on
                OVN's roadmap it might be instead worth looking at a
                "monolithic" plugin, which can also be easily
                implemented by inheriting from
                neutron.db.db_base_plugin_v2.NeutronDbPluginV2, and
                then adding all the python mixins for the extensions
                the plugin needs to support.


                On 23 February 2015 at 18:32, Ben Pfaff
                < <>> wrote:

                    [branching off a discussion on ovs-dev at this point:

                    On Fri, Feb 20, 2015 at 6:56 PM, Kyle Mestery
                    < <>>
                    > One thing to keep in mind, this ties somewhat
                    into my response to Russell
                    > earlier on the decision around ML2 vs. core
                    plugin. If we do ML2, there are
                    > type drivers for VLAN, VXLAN, and GRE tunnels.
                    There is no TypeDriver for
                    > STT tunnels upstream now. It's just an item we
                    need on the TODO list if we
                    > go down the STT tunnel path.

                    It was suggested to me off-list that this part of
                    the discussion should be on
                    openstack-dev, so here it is ;-)

                    OpenStack Development Mailing List (not for usage

                OpenStack Development Mailing List (not for usage

-- Kevin Benton

            OpenStack Development Mailing List (not for usage questions)

        OpenStack Development Mailing List (not for usage questions)

    OpenStack Development Mailing List (not for usage questions)

OpenStack Development Mailing List (not for usage questions)

OpenStack Development Mailing List (not for usage questions)

Reply via email to