KEDA does not scale up from 0 during rollout causing deployment failure

[Jijithnair]

We use KEDA (v2.18) with a Prometheus trigger based on Istio request metrics:

minReplicaCount: 0 (scale to zero)
cooldownPeriod: 3600
Trigger: istio_requests_total (RPS)

When the Deployment is already at 0 replicas and we deploy a new version, the rollout never starts because KEDA does not scale the workload back up. The reason is, Prometheus returns 0 during deploy (if there is no traffic), and KEDA never sees trigger “activity” therefore never scales 0→1.

kubectl rollout status/Helm --wait fails, due to which our deployment pipelines would complain that release was not successful and proceed to rollback, since there are no pods to help in asserting that the rollout was successful or healthy.

The service under scope is not a production workload and is something that needs to be scaled up only if there is traffic, hence the decision of scaling down to 0 and using Prometheus trigger for traffic detection via Istio metric.

Is there an existing or recommended approach in KEDA to ensure Deployment rollouts succeed when replicas are at 0, given that the Prometheus metric remains 0 until a pod exists? I would be looking at something like:

• Deployment would always scale to a configured number of replicas initially
• After the deployment, or after a threshold, KEDA takes over the scaling behaviour

If not, is this a known limitation and should we open a feature request?

[rickbrouwer]

Are you looking for something like this? You could add an annotation to pause a ScaledObject at a given number, which it will then scale to.

kubectl annotate scaledobject <name> autoscaling.keda.sh/paused-replicas="1" --overwrite

[Jijithnair]

@rickbrouwer This would mean the KEDA scaling decisions are paused? I presume, I would need to 'resume' this post deployment, is that what you are coming to? If that's the case, we would require our pipelines to orchestrate this 'pause' and 'resume'. I was expecting something in built from KEDA to achieve this.

[rickbrouwer]

That's right. But perhaps there's another solution, but I can't think of it right now.

[rickbrouwer]

What kind of solution did you expect in KEDA?

[Jijithnair]

@rickbrouwer, What I’m trying to achieve is simple:
What I guess is happening is, when a fresh deployment happens, Kubernetes starts with >0 replicas, but KEDA immediately scales it back to 0 if the metric for taking scaling decision returns a 0 value or nil (since our minReplica configuration is 0)

I was hoping for a KEDA-native option to keep an initial warm-up replica (or small initial replica count) after a fresh deploy, and then let KEDA take over scaling normally afterward.
Basically: don’t instantly scale a new deployment to 0, but still allow scale-to-zero later. Maybe KEDA can honour the 'coolDownPeriod'.

Or from another dimension, for users leveraging KEDA for 'scale-to-zero' behaviour, what is the expected and standard way to deal with the deployment pipelines that rely on pod readiness to mark a rollout as successful?