Scaling to 0 with a CPU scaler

[aryan9600]

Hello, I'd like to ask is it at all possible to scale the target deployment to 0 replicas when using a CPU scaler? From https://keda.sh/docs/2.6/scalers/cpu/

This scaler will never scale to 0 and even when user defines multiple scaler types (eg. Kafka + cpu/memory, or Prometheus + cpu/memory), the deployment will never scale to 0.

Is this true, even when spec.IdleReplicaCount is set to 0? If so, could I please understand the reason behind this functionality?

[tomkerkhove]

Scale to 0 for CPU or memory is not required because if it goes to 0, nothing is running.

If nothing is running, CPU nor memory would ever go up because there is nothing running anymore.

Makes sense?

[aryan9600]

Yeah, but a use case might arise where there are other controllers that can scale the deployment from 0 to 1 and vice versa. KEDA can scale the deployment between the specified min and max replica count based on the cpu consumption only when there's a minimum of one pod running. The HPA behaves in a similar manner, if I am not mistaken.

[tomkerkhove]

I'm not sure I'm getting the scenario - Sorry.

So you want to scale the workload with KEDA but have another controller scale it as well; is that it?

[aryan9600]

Kind of yes. KEDA would scale the the deployment only when there are 1 or more replicas, but it should leave the deployment alone if it has been scaled down to zero replicas. Maybe we could have an annotation that enables this, or maybe the CPU and memory scalers could infer that if idleReplicaCount is set to 0, then it shouldn't scale the deployment back to the minReplicaCount?

[tomkerkhove]

You might want to check #944 but we do not encourage people to mix autoscalers. We'd probably actually discourage you to do it.

May I ask why you'd want to use another autoscaler? If so, maybe #944 is what you need where that other controller just tells it to stop autoscaling and what replica count wants.

[aryan9600]

Yep, this looks apt. We want to add KEDA support for Flagger. Flagger scales the canary deployment to 1 during a canary analysis (which can be further autoscaled using a HPA) and then it scales the deployment down to 0 when the analysis is over. With KEDA, the deployment is scaled back up to the minReplicaCount when Flagger scales the deployment down to 0, which is unwanted.

[tomkerkhove]

Then I think using the above is ideal for you.

I'm not sure about Flagger's details, but if it has a CRD we can scale that resources as a scaleTarget as well.

[asobti]

For anyone who comes across this issue, this has been made possible since Keda v2.11
#4346