[Openstack-operators] Two regions and so two metadata servers sharing the same VLAN

James Dempsey jamesd at catalyst.net.nz
Thu Nov 26 22:29:25 UTC 2015

On 27/11/15 03:49, gilles.mocellin at nuagelibre.org wrote:
> Hello stackers !
> Sorry, I also cross-posted that question here
> https://ask.openstack.org/en/question/85195/two-regions-and-so-two-metadata-servers-sharing-the-same-vlan/
> But I think I can reach a wider audience here.
> So here's my problem.
> I'm facing an non-conventional situation. We're building a two region
> Cloud to separate a VMware backend and a KVM one. But both regions share
> the same 2 VLANs where we connect all our instances.
> We don't use routers, private network, floating IPs... I've enabled
> enable_isolated_metadata, so the metadata IP is inside the dhcp
> namespace and there's a static route in the created instances to it via
> the dhcp's IP. The two DHCPs could have been a problem but we will use
> separate IP ranges, and as Neutron sets static leases with the instances
> MAC address, they should not interfere.
> The question I've been asked is whether we will have network problems
> with the metadata server IP, that will exist in 2
> namepaces on 2 neutron nodes but on the same VLAN. So they will send ARP
> packets with different MAC, and will perhaps perturb access to the
> metadata URL form the instances.

I think you will see periodic interruptions in service.  ARP tables will
have entries that for the metadata service IP which flap back and forth
as the MAC is expired/re-learned.  As is often the case with duplicate
addressing, it will work sometimes and be unhappy sometimes.  This might
not be a huge problem, if cloud-init is retrying enough during boot, but
keep in mind that other pieces of software also poll the metadata
service(puppet/facter, for example).

I think you understand the core issue: you have two instances of Neutron
working in the same L2 broadcast domain... I wouldn't want to support a
configuration like this in production.

> Tcpdump shows nothing wrong, but I can't really test now because we
> haven't got yet the two regions. What do you think ?
> Of course, the question is not about why we choose to have two regions.
> I would have chosen Host Agregates to separate VMware and KVM, but
> cinder glance should have been configure the same way. And with VMware,
> it's not so feasible.
> Also, if we can, we will try to have separate networks for each regions,
> but it involves a lot of bureaucracy here...
> _______________________________________________
> OpenStack-operators mailing list
> OpenStack-operators at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators

James Dempsey
Senior Cloud Engineer
Catalyst IT Limited

More information about the OpenStack-operators mailing list