[openstack-dev] IPAM reference driver status and other stuff

Salvatore Orlando sorlando at nicira.com
Mon Mar 9 10:44:20 UTC 2015


Aloha everybody!

This is another long email, so here's the summary for people with 5-minute
or less attention span:

The IPAM reference driver is almost ready [1]. Unfortunately letting the
driver doing allocation pools required a few major changes, so the latest
patchset is rather big. I would like reviewers to advice on how they prefer
to split it out in multiple patches.

I also had to add an additional method to the subnet interface [2]

However, I am not sure we are doing the right thing wrt the introduction of
the driver. I think we might need to think about it more thoroughly, and
possibly introduce a concept of "allocator". Otherwise if we switch an
existing deployment to the driver, it's very likely things will just break.

So here's the detailed part.
There are only a few bits needed to complete the IPAM reference driver:
- A thorough unit test case for the newly introduced "subnet manager",
which provides db apis.
- Some more unit tests for driver behaviour
- Logic for handling allocation pools update (which pretty much needs to be
copied over from db base plugin with minor changes)
- Rework the synchronisation between the ipam object and the neutron db
object. The current one is very error prone.

While dealing with management of allocation pools, I found out that it's
not convenient for the IPAM driver to rely on Neutron's DB allocation
pools. This is a bit of a chicken and egg problem, since:
A) Neutron needs to call the driver to define allocation pools
B) The driver needs neutron to define allocation pools before creating
availability ranges
C) Neutron cannot create allocation pools if the driver has not defined them
I therefore decided, reluctantly, to add more data duplication -
introducing IPAM tables for allocation pools and availability ranges. The
latter is meant to replace Neutron's one when the IPAM driver is enabled,
whereas the former is pure data duplication. There is also an association
table between neutron subnets and ipam objects for the same subnet; I
decided to do so to not do further duplication.
I dislike this data separation; on the other hand the only viable
alternative would be to give the IPAM driver full control over neutron's
database tables for IP Allocation and Subnet allocation pools. While this
is feasible, I think it would be even worse to give code which can
potentially be 3rd party control over data structures used by the Neutron
API.

As a result, the patch is different and larger. I would like to split it to
make it simpler to review, but rather than just doing that of my own accord
I would like to ask reviewers how they would prefer to have it split. At
the end of the day I'm not the one who has to spend a lot of time reviewing
that code.

Nevertheless, I think I've realised our refactoring approach is kind of
flawed, since it might work for greenfields deployments, but I don't think
it will work for brownfields deployments. Also, switching between drivers
is pretty much impossible (but we were already aware of that and we agreed
to address this in the future).
The decorator approach currently taken in [3] allows to either use the
driver or not - which means that if an operator decides to switch to the
IPAM driver, then all allocation data for existing subnets would be simply
lost, unless we provide a solution for migrating data. This is feasible and
rather trivial, but implies an additional management step.

An alternative solution would be to introduce a concept of "allocator" for
subnets. Such information should be stored in the database. It could point
to an IPAM driver or to nothing. In the latter case it would simply
instruct to use the "usual" baked-in IPAM code. This would allow us to make
the driver work on existing deployments, and pave the way for multiple
drivers. Indeed in this way, the new IPAM driver will be used for subnets
created after enabling it, whereas existing subnets will keep using the
existing logic. This should also make safe cases in which operators revert
the decision of using the IPAM driver. Additionally, administrative APIs
might be provided to migrate existing subnets to/from the driver. However,
when adopting solutions like this, it is important to pay extra care in
ensuring that the database does not start relying on the current
configuration, and therefore we need to find a way to decouple the
allocator from the actual IPAM driver, which should represent its
realization.

Any feedback is very appreciated as usual.

Salvatore

[1] https://review.openstack.org/#/c/150485/
[2] https://review.openstack.org/#/c/150485/6/neutron/ipam/driver.py
[3]
https://review.openstack.org/#/c/153236/15/neutron/db/db_base_plugin_v2.py
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20150309/fc24aad2/attachment.html>


More information about the OpenStack-dev mailing list