[openstack-dev] [Fuel][Plugins][Orchestration] Unclear handling of primary-controler and controller roles

Andrew Woodward xarses at gmail.com
Wed Jan 28 22:05:11 UTC 2015


On Wed, Jan 28, 2015 at 3:06 AM, Evgeniy L <eli at mirantis.com> wrote:
> Hi,
>
> +1 for having primary-controller role in terms of deployment.

Yes, we need to continue to be able to differentiate the difference
between the first node in a set of roles, and all the others.

For controllers we have logic around how the services start, and if we
attempt to create resources. This allows the deployment to run more
smoothly.
For mongo the logic is used to setup the primary vs backup data nodes.
For plugins I would expect to continue to see this kind of need and
would need to be able to expose a similar logic when adding roles /
tasks

I'm however not sure that we need to do this with some kind of role,
this could simply be some parameter that we then use to set the
conditional that we use to apply primary logic already. Alternately,
this could cause the inclusion of 'primary' or 'first node' tasks that
would do these specific work with out the presence of the conditional
to run this testing

> In our tasks user should be able to run specific task on primary-controller.
> But I agree that it can be tricky because after the cluster is deployed, we
> cannot say who is really primary, is there a case when it's important to
> know
> who is really primary after deployment is done?

for mongo, its important to find out who is currently the primary
prior to deployment starting (which may not have been the primary that
the deployment started with) So it may be special in it's case.

for controller, its irrelevant as long as it's not set to a newly
added node (a node with a lower node.id will cause this and create
problems)

> Also I would like to mention that in plugins user currently can write
> 'roles': ['controller'],
> which means that the task will be applied on 'controller' and
> 'primary-controller' nodes.
> Plugin developer can get this information from astute.yaml file. But I'm
> curious if we
> should change this behaviour for plugins (with backward compatibility of
> course)?
>

writing roles: ['controller'] should apply to all controllers as
expected, with the addition of roles: ['primary-controller'] only
applying to the primary controller.
> Thanks,
>
>
> On Wed, Jan 28, 2015 at 1:07 PM, Aleksandr Didenko <adidenko at mirantis.com>
> wrote:
>>
>> Hi,
>>
>> we definitely need such separation on orchestration layer.
>>
>> > Is it possible to have significantly different sets of tasks for
>> > controller and primary-controller?
>>
>> Right now we already do different things on primary and secondary
>> controllers, but it's all conducted in the same manifest and controlled by
>> conditionals inside the manifest. So when we split our tasks into smaller
>> ones, we may want/need to separate them for primary and secondary
>> controllers.
>>
>> > I wouldn't differentiate tasks for primary and other controllers.
>> > "Primary-controller" logic should be controlled by task itself. That will
>> > allow to have elegant and tiny task framework
>>
>> Sergii, we still need this separation on the orchestration layer and, as
>> you know, our deployment process is based on it. Currently we already have
>> separate task groups for primary and secondary controller roles. So it will
>> be up to the task developer how to handle some particular task for different
>> roles: developer can write 2 different tasks (one for 'primary-controller'
>> and the other one for 'controller'), or he can write the same task for both
>> groups and handle differences inside the task.
>>
>> --
>> Regards,
>> Aleksandr Didenko
>>
>>
>> On Wed, Jan 28, 2015 at 11:25 AM, Dmitriy Shulyak <dshulyak at mirantis.com>
>> wrote:
>>>
>>> But without this separation on orchestration layer, we are unable to
>>> differentiate between nodes.
>>> What i mean is - we need to run subset of tasks on primary first and then
>>> on all others, and we are using role as mapper
>>> to node identities (and this mechanism was hardcoded in nailgun for a
>>> long time).
>>>
>>> Lets say we have task A that is mapped to primary-controller and B that
>>> is mapped to "secondary" controller, task B requires task A.
>>> If there is no primary in mapping - we will execute task A on all
>>> controllers and then task B on all controllers.
>>>
>>> And how in such case deployment code will know that it should not execute
>>> commands in task A for "secondary" controllers and
>>> in task B on "primary" ?
>>>
>>> On Wed, Jan 28, 2015 at 10:44 AM, Sergii Golovatiuk
>>> <sgolovatiuk at mirantis.com> wrote:
>>>>
>>>> Hi,
>>>>
>>>> But with introduction of plugins and granular deployment, in my opinion,
>>>> we need to be able
>>>> to specify that task should run specifically on primary, or on
>>>> secondaries. Alternative to this approach would be - always run task on all
>>>> controllers, and let task itself to verify that it is  executed on primary
>>>> or not.
>>>>
>>>> I wouldn't differentiate tasks for primary and other controllers.
>>>> "Primary-controller" logic should be controlled by task itself. That will
>>>> allow to have elegant and tiny task framework ...
>>>>
>>>> --
>>>> Best regards,
>>>> Sergii Golovatiuk,
>>>> Skype #golserge
>>>> IRC #holser
>>>>
>>>> On Tue, Jan 27, 2015 at 11:35 PM, Dmitriy Shulyak
>>>> <dshulyak at mirantis.com> wrote:
>>>>>
>>>>> Hello all,
>>>>>
>>>>> You may know that for deployment configuration we are serializing
>>>>> additional prefix for controller role (primary), with the goal of deployment
>>>>> order control (primary-controller always should be deployed before
>>>>> secondaries) and some condiions in fuel-library code.
>>>>>
>>>>> However, we cannot guarantee that primary controller will be always the
>>>>> same node, because it is not business of nailgun to control elections of
>>>>> primary. Essentially user should not rely on nailgun
>>>>> information to find primary, but we need to persist node elected as
>>>>> primary in first deployment
>>>>> to resolve orchestration issues (when new node added to cluster we
>>>>> should not mark it as primary).
>>>>>
>>>>> So we called primary-controller - "internal" role, which means that it
>>>>> is not exposed to users (or external developers).
>>>>> But with introduction of plugins and granular deployment, in my
>>>>> opinion, we need to be able
>>>>> to specify that task should run specifically on primary, or on
>>>>> secondaries. Alternative to this approach would be - always run task on all
>>>>> controllers, and let task itself to verify that it is  executed on primary
>>>>> or not.
>>>>>
>>>>> Is it possible to have significantly different sets of tasks for
>>>>> controller and primary-controller?
>>>>> And same goes for mongo, and i think we had primary for swift also.
>>>>>
>>>>>
>>>>> __________________________________________________________________________
>>>>> 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
>>>>>
>>>>
>>>>
>>>>
>>>> __________________________________________________________________________
>>>> 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
>>>>
>>>
>>>
>>>
>>> __________________________________________________________________________
>>> 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
>>>
>>
>>
>> __________________________________________________________________________
>> 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
>>
>
>
> __________________________________________________________________________
> 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
>



-- 
Andrew
Mirantis
Fuel community ambassador
Ceph community



More information about the OpenStack-dev mailing list