Open Stack

On 31/08/2015 1:06 PM, Julien Danjou wrote:
> On Mon, Aug 31 2015, gord chung wrote:
>
>> for context, we have an open bug in Ceilometer where the value used to capture
>> cputime of an instance is being reset whenever the instance is restarted or
>> shutdown[1]. this seems to be the expected behaviour for libvirt driver[2] but
>> it conflicts with the fact that the meter is currently defined as 'cumulative'
>> -- which by definition, should not reset. it seems the correct meter type is a
>> 'gauge'.
> Nah, gauge is for things with discrete values, e.g. a number of objects
> in Swift.
> Cumulative are counters that increase and that can be reset.
> The problem is that the API still don't know to detect resets and
> normalize the values.
>
> That's what needs fixing. We could fix it with Gnocchi.
>
i think that's the reason the bug is raised... that's not what 
cumulative means. by definition it should increase over each successive 
value. it's cumulative from the pov of qemu process but not from 
instance pov. judging by how Hyper-V meter is built, it doesn't seem 
like Hyper-V's cputime meter would ever reset.

i'm not sure Gnocchi is where we should be fixing this as it really only 
(potentially) fixes it for Gnocchi and not for any of the other ways 
Ceilometer data can be consumed.

a proposed solution in bug and in a previous thread suggests that a 
'delta' meter be built from cputime value libvirt provides which would 
better handle the reset scenario. that said, is there another option to 
truly have a cumulative meter?

cheers,

-- 
gord