Open Stack

On 1/20/2017 9:00 AM, Matt Riedemann wrote:
> On 1/20/2017 4:53 AM, Eoghan Glynn wrote:
>>
>> Do we also need to be concerned about the placement API "warm-up" time?
>>
>> i.e. if a placement-less newton deployment is upgraded to placement-ful
>> ocata, then would there surely be a short period during which placement
>> is able to respond to the incoming queries from the scheduler, but only
>> with incomplete information since all the computes haven't yet triggered
>> their first reporting cycle?
>>
>> In that case, it wouldn't necessarily lead to a NoValidHost failure on a
>> instance boot request, but rather a potentially faulty placement
>> decision,
>> being based on incomplete information. I mean "faulty" there in the sense
>> of not strictly following the configured scheduling strategy.
>>
>> Is that a concern, or an acceptable short degradation of service?
>>
>> Cheers,
>> Eoghan
>>
>
> That's discussed a bit in this older thread [1]. If placement is up and
> running but there are no computes checking in yet, then it's going to be
> a NoValidHost from the filter scheduler because it's not going to
> fallback to the compute_nodes table.
>
> The nova-status command was written as an upgrade readiness check for
> this situation [2]. If there are compute nodes in the database (from
> Newton) but no resource providers in the placement service, then that
> upgrade check is going to fail.
>
> If it's a fresh install and there are 0 resource providers in the
> placement service and 0 compute nodes, then the upgrade check passes but
> provides a reminder about needing to make sure you get the computes
> configured and registered for placement as you bring them online.
>
> [1]
> http://lists.openstack.org/pipermail/openstack-dev/2016-December/109060.html
>
> [2]
> https://github.com/openstack/nova/blob/ae753d96281709397dcfe5dd4ff7d6db57f3683e/nova/cmd/status.py#L301
>
>

Also, before anyone mentions this:

"Well what if I've upgraded my controller services to Ocata, including 
placement and nova-scheduler, but my computes are all still Newton?!"

The answer is Newton computes can be configured to register with the 
placement service already, and we backported a fix from Dan Smith [1] to 
make sure the Newton computes keep trying to connect to placement.

So this means you can still configure your Newton computes to talk to 
the Ocata placement service, even before rolling out placement, and once 
it's up the computes will check in and you shouldn't have NoValidHost 
issues, or at least the window should be small.

[1] https://review.openstack.org/#/c/419217/

-- 

Thanks,

Matt Riedemann