Honour HPA Scale-Down Policy When Scaling Down to Zero Replicas

[subbu2k]

Proposal

Current KEDA behaviour during scale to zero is such that if the scale trigger remains inactive after the cooldownPeriod, KEDA will force the deployment to scale down to 0, regardless of the current number of replicas.

The proposal is to honour the scale-down policy defined in the Horizontal Pod Autoscaler (HPA) before enforcing the scale to zero. If the user configures the following in the ScaledObject:

behavior:
  scaleDown:
    policies:
    - type: Pods
      value: 2
      periodSeconds: 60

then, KEDA should first gradually scale down the replicas from N to 1 in steps of 2, and only then scale from 1 to 0.

Use-Case

After the trigger becomes inactive for the cooldownPeriod, KEDA's current behaviour forces the deployment to scale down to zero immediately. If there is a short burst of traffic right after the cooldownPeriod, the deployment will need to scale up from 0 to 1, increasing request latency.

If KEDA honours the HPA scale-down policy, replicas will scale down gradually in defined steps, ensuring that even during short traffic bursts, there are still pods available to serve requests, reducing latency and improving response times.

Is this a feature you are interested in implementing yourself?

No

Anything else?

No response

[JorTurFer]

Hello
I'm not sure if I'm understanding the issue. Let's say that you have 10 instances because you have a peak and then the load stops. KEDA will scale the workload from 10 to zero as soon as cooldownPeriod passes. You would like to scale it from to 1 using the regular HPA flow and only when the HPA is at minReplicas, scaling to zero, is that?

[subbu2k]

Hi @JorTurFer
Yes, you are correct. I would like the scaling from N to minReplicaCount to be handled by the regular HPA flow, adhering to the policy defined in the HPA spec. From minReplicaCount to 0, KEDA should wait for cooldownPeriod seconds before scaling down to 0 replicas.

Example: Suppose a deployment has 9 replicas. Using the HPA scale-down policy provided in HPA documentation in the ScaledObject

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
spec:
  cooldownPeriod:        60                              
  minReplicaCount:        1              
  maxReplicaCount:       9                                                  
  advanced:                                     
    horizontalPodAutoscalerConfig:                         
      name: {name-of-hpa-resource}                 
      behavior:                                           
        scaleDown:
          policies:
          - type: Pods
            value: 4
            periodSeconds: 60

As per the HPA documentation, the deployment should be scaled down from 9 to 5 after the first minute, and then from 5 to 1 after the second minute.

The first policy (Pods) allows at most 4 replicas to be scaled down in one minute.

However, with the current KEDA behavior, since the cooldownPeriod is 60 seconds, the deployment scales down from 9 to 0 immediately after the first minute if there is no traffic. Instead, I would like the HPA to handle scaling down from 9 to 1 during the first two minutes, and then have KEDA kick in to scale it down to zero in the third minute.

This approach prevents a sudden drop in backend capacity. If any load appears after the first minute, the backend will still be able to serve traffic until the third minute.

[JorTurFer]

It looks as a good feature to include 😄 Not replacing current behaviour but supporting both

@kedacore/keda-core-contributors @kedacore/keda-core-maintainers WDYT?

[rickbrouwer]

Yes, it thinks it's a good feature.

[rlanhellas]

Agree on this feat, and I'm willing to work on this. May you assign to me @JorTurFer / @rickbrouwer ?

[rlanhellas]

@JorTurFer , in order to keep supporting both behaviors, we need a new field on SO, something like a boolean coolDownOnlyAfterMinReplica meaning that coolDownPeriod will be evaluated only after min replica reached, with default equal to false, so keeping current behavior as default.

WDYT?