[openstack-dev] [ceilometer][aodh][vitrage] Raising custom alarms in AODH

AFEK, Ifat (Ifat) ifat.afek at alcatel-lucent.com
Wed Dec 2 17:16:04 UTC 2015

Hi Julien,

Please see our questions below.

Ifat and Elisha.

> -----Original Message-----
> From: Julien Danjou [mailto:julien at danjou.info]
> On Wed, Dec 02 2015, ROSENSWEIG, ELISHA (ELISHA) wrote:
> > Regarding the second point: Say we have 30 different types of alarms
> > we might want to raise on an OpenStack instance (VM). What I
> > understand from your explanation is that when we create a new
> > instance, we need to create 30 new alarms in Aodh that can be
> > triggered some time in the future. If we have 100 instances, we will
> > effectively have 3,000 alarms created in Aodh, and so on with more
> instances.
> Not necessarily. You can create one alarm that has conditions large
> enough to match e.g. all your VMs, and an alarm action that can be
> generic enough so that it will do the right thing for each VM.

As we understand it, if we take the first approach you describe, then we can have an alarm refer to all the VMs in the system, but then if the alarm is triggered by one VM or by five VMs, the result will be the same - only one alarm will be active. What we want is to be able to distinguish between the different VMs - to know which alarms were triggered on each specific VM.

One of the motivations for this is that in Horizon we would like to display all the alarms, where we would like to be able to see that a problem occurred on instance1, instance2 and instance8, not just that there was a problem on some VMs out of a group. 

Can this be supported without defining an alarm for every VM separately?

> The alarm system provided by Aodh is really a simple event -> trigger
> system in this area. How precise or large is your event really depends
> on the granularity that your trigger (which is usually a Web hook) can
> handle.
> > A different approach might be to create a new alarm in Aodh on-the-
> fly.
> > However, we are under the impression that the creation time can be up
> > to one minute, which will cause a large delay. Is there any way to
> shorten this?
> Creation time of an alarm of one minute? That's not normal. It should
> consist of just a record in the database so it should be pretty fast.

This is what Ryota Mibu wrote us:

> The reason is that aodh evaluator may not be aware of new alarm definitions and won't send notification until its alarm definition > cache is refreshed in less than 60 sec (default value).

Did we misunderstand?

More information about the OpenStack-dev mailing list