[Openstack-operators] Milti-site Keystone & Galera

Jonathan Proulx jon at jonproulx.com
Tue Sep 8 21:22:18 UTC 2015


Thanks Jay & Matt,

That's basically what I thought, so I'll keep thinking it :)

We're not replicating glance DB because images will be stored in
different local Ceph storage on each side so the images won't be
directly available.  We thought about moving back to a file back end
and rsync'ing but RBD gets us lots of fun things we want to keep
(quick start, copy on write thin cloned ephemeral storage etc...) so
decided to live with making our users copy images around.

-Jon



On Tue, Sep 8, 2015 at 5:00 PM, Jay Pipes <jaypipes at gmail.com> wrote:
> On 09/08/2015 04:44 PM, Jonathan Proulx wrote:
>>
>> Hi All,
>>
>> I'm pretty close to opening a second region in my cloud at a second
>> physical location.
>>
>> The plan so far had been to only share keystone between the regions
>> (nova, glance, cinder etc would be distinct) and implement this by
>> using MariaDB with galera replication between sites with each site
>> having it's own gmcast_segment to minimize the long distance catter
>> plus a 3rd site with a galera arbitrator for the obvious reason.
>
>
> I would also strongly consider adding the Glance registry database to the
> same cross-WAN Galera cluster. At AT&T, we had such a setup for Keystone and
> Glance registry databases at 10+ deployment zones across 6+ datacenters
> across the nation. Besides adjusting the latency timeout for the Galera
> settings, we made no other modifications to our
> internal-to-an-availability-zone Nova database Galera cluster settings.
>
> The Keystone and Glance registry databases have a virtually identical read
> and write data access pattern: small record/row size, small number of
> INSERTs, virtually no UPDATE and DELETE calls, and heavy SELECT operations
> on a small data set. This data access pattern is an ideal fit for a
> WAN-replicated Galera cluster.
>
>> Today I was warned against using this in a multi writer setup. I'd planned
>>   on one writer per physical location.
>
>
> I don't know who warned you about this, but it's not an issue in the real
> world. We ran in full multi-writer mode, with each deployment zone writing
> to and reading from its nearest Galera cluster nodes. No issues.
>
> Best,
> -jay
>
>> I had been under the impression this was the 'done thing' for
>> geographically sepperate regions, was I wrong? Should I replicate just
>> for DR and always pick a single possible remote write site?
>>
>> site to site link is 2x10G (different physical paths), short link is
>> 2.2ms average latency (2.1ms low, 2.5ms high over 250 packets) long
>> link shouldn't be much longer but isn't yet complete to test.
>>
>> -Jon
>>
>> _______________________________________________
>> OpenStack-operators mailing list
>> OpenStack-operators at lists.openstack.org
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
>>
>
> _______________________________________________
> OpenStack-operators mailing list
> OpenStack-operators at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators



More information about the OpenStack-operators mailing list