Open Stack

Hi Salvatore,
Thanks for your hyperlink. It's really a monster thread that contains
everyone's opinion. But it's useful to me.
So, Before we focus on the Neutron core itself, we should firstly release a
suite standardized APIs and a framework for vendors' codes.
About this job, I think most of it is already OK. We have 20+ monolithic
plugins following NB API and plugin framework.
We need publish an API doc for internal interface(I prefer to call it SB
API, stand on the Neutron core's point to consider, vendors' codes
do not belong to core.) and other things unsuitable now.

In my opinion, the Neutron core's main responsibility is data model and DB,
schedule and dispatch, API and validation, framework and workflow.

Some more comments inline.


This is a very important discussion - very closely related to the one going
on in this other thread
http://lists.openstack.org/pipermail/openstack-dev/2014-September/045768.html
.
Unfortunately it is also a discussion that tends to easily fragment and
move in a thousand different directions.
A few months ago I was too of the opinion that vendor plugins and drivers
were the main reason of unnecessary load for the core team. I still think
that they're an unnecessary heavy load, but I reckon the problem does not
really lies with open source versus vendor code. It lies in matching
people's competencies with subsystems and proper interface across them - as
already pointed out in this thread.
Yes, it's really important.

I have some more comments inline, but unless growing another monster thread
I'd rather start a different, cross-project discussion (which will
hopefully not become just a cross-project monster thread!)

Salvatore

On 15 September 2014 08:29, Germy Lure <germy.lure at gmail.com> wrote:

> Obviously, to a vendor's plugin/driver, the most important thing is
> API.Yes?
> NB API for a monolithic plugin or a service plugin and SB API for a
> service driver or agent, even MD. That's the basic.
> Now we have released a set of NB APIs with relative stability. The SB
> APIs' standardization are needed.
>

The internal interface between the API and the plugins is standardized at
the moment through use of classes like [1]. A similar interface exists for
ML2 drivers [2].
To the monolithic plugins, [1] is useful. Vendors can implement those APIs
and keep their codes locally.

At the moment the dispatch of an API call to the plugin or from a plugin to
a ML2 driver is purely a local call so these interfaces are working fairly
well at the moment. I don't know yet however whether they will be
sufficient in case plugins are split into different repos. ML2 Driver
maintainers have however been warned in the past that the driver interface
is to be considered internal and can be changed at any time. This does not
apply to the plugin interface which has been conceived in this way to
facilitate the development of out of tree plugins.
Indeed, it's difficult to split MDs from ML2 plugin framework. I think it
need some adaption.

On the other hand, if by SB interfaces you are referring to the RPC
interfaces for communicating between the servers and the various plugin, I
would say that they should be considered internal at the moment.

[1]
https://github.com/openstack/neutron/blob/master/neutron/neutron_plugin_base_v2.py#L28
[2]
https://github.com/openstack/neutron/blob/master/neutron/plugins/ml2/driver_api.py

>
> Some comments inline.
>
>
>
> On Fri, Sep 12, 2014 at 5:18 PM, Kevin Benton <blak111 at gmail.com> wrote:
>
>> > So my suggestion is remove all vendors' plugins and drivers except
>> opensource as built-in.
>>
>> Yes, I think this is currently the view held by the PTL (Kyle) and some
>> of the other cores so what you're suggesting will definitely come up at the
>> summit.
>>
> Good!
>

The discussion however will not be that different from the one we're seeing
on that huge thread on splitting out drivers, which has become in my
opinion a frankenthread.
Nevertheless, that thread points out that this is far from being merely a
neutron topic (despite neutron being the project with the highest number of
drivers and plugins).


>
>>
>> > Why do we need a different repo to store vendors' codes? That's not the
>> community business.
>> > I think only a proper architecture and normal NB&SB API can bring "a
>> clear separation between plugins(or drivers) and core code", not a
>> different repo.
>>
>> The problem is that that architecture won't stay stable if there is no
>> shared community plugin depending on its stability. Let me ask you the
>> inverse question. Why do you think the reference driver should stay in the
>> core repo?
>>
>> A separate repo won't have an impact on what is packaged and released so
>> it should have no impact on "user experience", "complete versions",
>> "providing code examples",  or "developing new features". In fact, it will
>> likely help with the last two because it will provide a clear delineation
>> between what a plugin is responsible for vs. what the core API is
>> responsible for. And, because new cores can be added faster to the open
>> source plugins repo due to a smaller code base to learn, it will help with
>> developing new features by reducing reviewer load.
>>
> OK, the key point is that vendors' code should be kept by themselves NOT
> by the community. But in the same time, the community should provide
> some open source reference as standard examples for those new cores and
> vendors.
> U are right, "A separate repo won't have an impact on what is packaged and
> released". The open source can stays in the core repo or a different one.
> In any case, we need them there for referencing and version releasing.
> Any vendor would not maintain the open source codes, the community only.
>

I think that we are probably focusing too much on the "separate repo"
issue, which is probably being seen as punitive for drivers and plugins.
The separate repo would be just a possible tool for achieving the goal of
reducing the review load imposed by drivers on the core team while keeping
them part of the integrated release.
I don't think it will be seen as punitive. Vendors can write their plugins
or drivers when a deal occurs and they do not need to submit code to
community
and wait for approving. The community dose not need to waste energy to
review those codes.
If we must give a cause to keep vendors codes in the same or different
repos, I think just installation-automatic. Similar to Linux kernel, user
installs system on different hardware platform, and the system discovers
the type of CPU, NIC, etc. and loads the right driver for that device. But
OpenStack is not Linux, they are different.

Regarding the  "open source reference" solution... first there's no such
thing like this. The fact that the upstream gate tests ML2 + OVS mech
driver implicitly seem to make this the "reference", but this has not been
sanctioned anywhere; second, the Neutron source tree does not just have a
"reference plugin". It also has the DHCP agent, the OVS agent, the L3
agent, which implement also the management plane of a network
virtualization system which uses OVS, iptables and other tools as its
datapath. This is just for saying that when we talk about "splitting stuff
from the main repo" there is more than just vendor plugin and drivers. It's
more about giving control of subsystems to the people which are truly
experts of that subsystem. I don't know whether splitting repositories
would be the way to go.
Splitting into different repos is helpful to standardizing APIs. We can
also split those agents into separate repo as sub modules. This is, I
think, the next stage after vendors' codes splitting.

BR,
Germy
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140918/79c827ef/attachment.html>