Open Stack

On Fri, Aug 14, 2015 at 5:36 PM, Adrian Otto <adrian.otto at rackspace.com>
wrote:

> That's clear, thanks Gal. The feature benchmark should be parity with how
> the majority of other (complete) remote drivers for libnetwork behave. From
> a Magnum perspective we value consistency from an end user perspective when
> you use containers on OpenStack compared to when you run them outside of an
> OpenStack cloud environment.
>


When I previously evaluated the port mapping question, the most valid
reasons I could see of why an operator
that uses Neutron to provide container networking would want port mapping
was:
- density and availability of floating IPs, i.e. how many endpoints I could
expose, which instead of being
  determined by the amount of free FIPs it would be #FIPs times ports.
- current restrictions in the container orchestration software to use.


>
> If we offer extra features that's okay, but we do need to be careful not
> to to leave feature gaps where things hat work elsewhere don't work on
> OpenStack. Do we know if there are other libnetwork remote drivers that
> support port mapping, or have a statement of intent to implement it? That
> should help guide us here. Is it too early to know?
>

I would imagine that those that have a network model in which the container
networking is accessible/routable from the host
networking namespace will find a way to accommodate the port mapping use
case and will always use the host IP as the address to
reach the containers it contains.

The last status I saw for Docker's own overlay driver was that they did not
have a way to provide port mapping and that one
option that they had considered for software that depended on it was to
have one veth on the overlay and one veth for port
mapping.

Under the Neutron networking implementations the networks in which the
instances running in a compute node take part
are not necessarily routable/reachable from the default host namespace.
This is something that would break some
assumptions made by container orchestration that was developed when there
was only the typical docker0 bridge
scenario.

In Kubernetes, for example one would normally target the minions with
external load balancers and then the port
mapping would get the data from the host to the correct pod. In Neutron, to
serve this case, the usual thing would be
to use Neutron's LBaaS and assign a FIP to the LBaaS. Of course, that means
that you get a whole FIP for a single
service and that may be expensive.

We have to work with the current bay model providers and see how many
assumptions need to be accommodated
(either by Neutron API usage patterns or extensions) and which could be
smoothened out by means of plugins in
the orchestrator.



> --
> Adrian
>
> On Aug 14, 2015, at 8:09 AM, Gal Sagie <gal.sagie at gmail.com> wrote:
>
> Hi Adrian,
>
> Thanks for the explanation, i agree with you that we shouldn't break
> anything useful in docker, but from what i understand
> (and please correct me if i am wrong) you are describing an implementation
> detail of docker networking (at its default current state).
>
> Kuryr is not an implementation of containers networking, it is meant to
> allow docker networking to be
> constructed using Neutron plugins and solutions.
>
> For the point i was trying to make, lets take the simple case of
> connecting containers in a host (not the nested VM case), assuming
> we use with Kuryr the L2 OVS agent and L3 agent (Neutron reference
> implementation) and we are able to plug
> containers to a neutron network and define a floating ip to it, why would
> we need port mapping? (the iptables
> translation is already happening as we have NAT)
>
> Hope that make sense, and please correct me if i am saying nonsense or i
> didn't grasp the full
> use case of port mapping.
> But none the less, we will need to allow anything that docker allows and
> keep compatibility with all the available tooling
> that depends on it as you mention (and of course be flexible to use Kuryr
> in the same environment with other
> docker remote drivers)
>
> Gal
>
> On Fri, Aug 14, 2015 at 5:26 PM, Adrian Otto <adrian.otto at rackspace.com>
> wrote:
>
>> The port mapping feature is the -p flag on the "docker run" command. It
>> determines which ports in the network namespace of the container are
>> exposed to the root namespace. It configures iptables rules and docker
>> proxy capabilities to achieve the desired result. This feature is
>> essential, so we must not break it.
>>
>> In other words, this feature is what allows a network port within a
>> container to be externally accessed, and on what IP address(es) and port
>> number(s) on the host.
>>
>> Example:
>>
>> docker run -d -p 12.34.56.7:8000:80 nginx:latest
>>
>> This runs the nginx container and exposes top port 80 from inside the
>> container to tcp port 8000 on 12.34.56.7 on the host. Without this feature
>> docker is rather useless for running network services unless you use -net
>> host or an equivalent workaround. This could break a lot of tooling that
>> depends on -p.
>>
>> --
>> Adrian
>>
>> On Aug 14, 2015, at 6:57 AM, Gal Sagie <gal.sagie at gmail.com> wrote:
>>
>> Hi feisky,
>>
>> I think thats a great question, not because of port-mapping in particular
>> :) but because
>> we need to think on a feature by feature basis and map all the features
>> the dockers API allow which
>> we cannot support directly with Neutron API or its services sub-projects
>> API.
>> (apuimedo, maybe we need to set this as a future task for us)
>>
>> But we also need to understand the use cases for supporting these API's
>> so we can address them
>> and give them priorities (and this is something we as a community need to
>> decide how to handle).
>>
>> For your question, given that we have network isolation and security from
>> neutron API's and given
>> we have NAT support (by Neutron API and the plugins implementing the
>> network) , what do you see as a use case to use the port-mapping ?
>>
>> I welcome you and everyone else to raise and describe these use cases so
>> the Neutron/Kuryr community can think
>> how to solve and help, and if needed also adjust or add extensions for
>> support.
>>
>> Thanks
>> Gal.
>>
>>
>> On Fri, Aug 14, 2015 at 7:28 AM, feisky <feiskyer at gmail.com> wrote:
>>
>>> Will Kuryr supports docker's port-mapping?
>>>
>>>
>>>
>>> --
>>> View this message in context:
>>> http://openstack.10931.n7.nabble.com/neutron-kuryr-magnum-Design-Specification-for-Kuryr-tp82256p82299.html
>>> Sent from the Developer mailing list archive at Nabble.com.
>>>
>>>
>>> __________________________________________________________________________
>>> OpenStack Development Mailing List (not for usage questions)
>>> Unsubscribe:
>>> OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>
>>
>>
>>
>> --
>> Best Regards ,
>>
>> The G.
>>
>> __________________________________________________________________________
>> OpenStack Development Mailing List (not for usage questions)
>> Unsubscribe: OpenStack-dev-request at lists.openstack.org
>> ?subject:unsubscribe
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>
>>
>> __________________________________________________________________________
>> OpenStack Development Mailing List (not for usage questions)
>> Unsubscribe:
>> OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>
>>
>
>
> --
> Best Regards ,
>
> The G.
>
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> 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/20150816/7a900a39/attachment.html>