Open Stack

On 12 February 2015 at 19:57, John Belamaric <jbelamaric at infoblox.com>
wrote:

>
>
>   From: Salvatore Orlando <sorlando at nicira.com>
> Reply-To: "OpenStack Development Mailing List (not for usage questions)" <
> openstack-dev at lists.openstack.org>
> Date: Thursday, February 12, 2015 at 8:36 AM
> To: OpenStack Development Mailing List <openstack-dev at lists.openstack.org>
> Subject: [openstack-dev] [Neutron] Update on "DB" IPAM driver
>
>   Hi,
>
>  I have updated the patch; albeit not complete yet it's kind of closer to
> be an allocator decent enough to replace the built-in logic.
>
>  I will be unable to attend today's L3/IPAM meeting due to a conflict, so
> here are some highlights from me on which your feedback is more than
> welcome:
>
>  - I agree with Carl that the IPAM driver should not have explicit code
> paths for autoaddress subnets, such as DHCPv6 stateless ones. In that case,
> the consumer of the driver will generate the address and then to the IPAM
> driver that would just be allocation of a specific address. However, I have
> the impression the driver still needs to be aware of whether the subnet has
> an automatic address mode or not - since in this case 'any' address
> allocation won't be possible. There already comments about this in the
> review [1]
>
>
>  I think the auto-generated case should be a base class as you described
> in [1], but each subclass would implement the specific auto-generation. See
> the discussion at line 468 in [2] and see what you think. Of course for
> addresses that come from RA there would be no IPAM.
>

I think this makes sense.


>
>  [1] https://review.openstack.org/#/c/150485/
>  [2]
> https://review.openstack.org/#/c/153236/2/neutron/db/db_base_plugin_v2.py,unified
>
>
>  - We had a discussion last week on whether the IPAM driver and neutron
> should 'share' database tables. I went back and forth a lot of time, but
> now to me it seems the best thing to do is to have the IPAM driver maintain
> an 'ip_requests' tables, where it stores allocation info. This table
> partially duplicates data in IPAllocation, but on the plus side it makes
> the IPAM driver self sufficient. The next step would be to decide whether
> we want to go a step further and also assume the driver should not access
> at all Neutron's DB, but I would defer that discussion to the next
> iteration (for both the driver and the IPAM interface)
>
>  - I promised a non blocking algorithm for IP allocation. The one I was
> developing was based on specifying the primary key on the ip_requests table
> in a way that it would prevent two concurrent requests from getting the
> same address, and would just retry getting an address until the primary key
> constraint was satisfied. However, recent information emerged on MySQL
> galera's (*) data set [2] certification  clarified that this kind of
> algorithm would still result in a deadlock error from failed data set
> certification. It is worth noting that in this case a solution based on
> traditional compare-and-swap is not possible because concurrent requests
> would be inserting data at the same time. I am now working on an
> alternative solution, and I would like to first implement a PoC for it (so
> that I can prove it works).
>
>  - The db base refactoring being performed by Pavel is under way [3]. It
> is worth noting that this is a non-negligible change to some of Neutron's
> basic and more critical workflows. We should expect pushback from the
> community regarding the introduction of this change in the 3rd milestone.
> At this stage I would suggest either:
> A) consider a strategy for running pluggable IPAM as optional
> B) consider delaying to Liberty.
> (and that's where I get virtually jeered and pelted with rotten tomatoes)
>
>
>  I wish I had some old tomatoes! Seriously, I think "A" is a reasonable
> approach. To make this really explicit we may want to basically replace the
> DB plugin class with a shim that delegates to either the current
> implementation or the new implementation, depending on the flag.
>

The fact that the current implementation is pretty much a bunch of private
methods in the db base plugin class executed within a transaction for
creating a port makes the situation a wee bit more complex. I'm not sure we
can replace the db plugin class with a shim so easily, because we need to
consider the impact on plugins which inherit from this base class. For
instance some plugins override methods from the base class, and this would
be a problem. For those plugins we must ensure old-style IPAM is performed.
A transitory solution might be to have, for the relevant methods 2 versions
- one would be the current one, and the other one would be the one
leveraging pluggable IPAM. During plugin initialisation, the plugin itself
will decide whether use or not the latter methods. This might be tuneable
with a configuration parameter too. The downside of this approach is that
it will not allow us to remove "old" baked in IPAM code, and will have an
impact on code maintainability which ultimately will result in accumulating
even more technical debt. However, I might be missing some better
alternative, so if you have any proposal just let me know.

Salvatore



>
>
>
>  Thanks for reading this post,
> Salvatore
>
>  [1] https://review.openstack.org/#/c/150485/
> [2]
> http://lists.openstack.org/pipermail/openstack-dev/2015-February/056007.html
> [3] https://review.openstack.org/#/c/153236/
>
>
> __________________________________________________________________________
> 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/20150213/eb590513/attachment.html>