I don't know how to solve this problem in aodh, but it is possible to use Prometheus to aggregate CPU utilization and trigger scaling. I wrote up how to do this with Senlin and Prometheus here: https://medium.com/@dkt26111/auto-scaling-openstack-instances-with-senlin-and-prometheus-46100a9a14e1?source=friends_link&sk=5c0a2aa9e541e8c350963e7ec72bcbb5 You can probably do something similar with Heat and Prometheus. On Sun, Aug 4, 2019 at 12:52 AM Bernd Bausch <berndbausch@gmail.com> wrote:
Prior to Stein, Ceilometer issued a metric named cpu_util, which I could use to trigger alarms and autoscaling when CPU utilization was too high.
cpu_util doesn't exist anymore. Instead, we are asked to use Gnocchi's rate feature. However, when using rates, alarms on a group of resources require more parameters than just one metric: Both an aggregation and a reaggregation method are needed.
For example, a group of instances that implement "myapp":
gnocchi measures aggregation -m cpu --reaggregation mean --aggregation rate:mean --query server_group=myapp --resource-type instance
Actually, this command uses a deprecated API (but from what I can see, Aodh still uses it). The new way is like this:
gnocchi aggregates --resource-type instance '(aggregate rate:mean (metric cpu mean))' server_group=myapp
If rate:mean is in the archive policy, it also works the other way around:
gnocchi aggregates --resource-type instance '(aggregate mean (metric cpu rate:mean))' server_group=myapp
Without reaggregation, I get quite unexpected numbers, including negative CPU rates. If you want to understand why, see this discussion with one of the Gnocchi maintainers [1].
My problem: Aodh allows me to set an aggregation method, but not a reaggregation method. How can I create alarms based on rates? The problem extends to Heat and autoscaling.
Thanks much,
Bernd.