KEDA fallback does not scale from 0 to 1 when metrics cannot be fetched and minReplicaCount is 0

[roxannelandry]

Report

When using KEDA with the following ScaledObject configuration (Helm template below), fallback is triggered if metrics cannot be fetched. However, if minReplicaCount is set to 0, the deployment does not scale from 0 to 1 as expected when fallback is active.

fallback:
    failureThreshold: 1
    replicas: 1 
    behavior: currentReplicasIfHigher
minReplicaCount: 0
maxReplicaCount: 6
  triggers:
    - type: prometheus
      metricType: AverageValue
      metadata:
        serverAddress: http://prometheus-operated.monitoring.svc:9090
        metricName: redis_sidekiq_queue_backlog_total
        query: sum(redis_key_size{key=~"{{ $scaler.queueList }}",namespace="....",service="..."})
        threshold: "{{ $scaler.queueBacklog }}"

Expected Behavior

When fallback is triggered and replicas: 1 is set, deployment should scale from 0 to 1.

Actual Behavior

• Fallback is triggered (confirmed in events/logs).
• Replica count remains at 0, does not scale up to 1.

Steps to Reproduce the Problem

1. Deploy ScaledObject with minReplicaCount: 0 and fallback as above.
2. Cause metric source to fail (scale down prometheus).
3. Observe fallback is triggered but no scaling from 0 to 1.

Logs from KEDA operator

ERROR 2025-11-07T15:05:01.443182024Z 2025-11-07T15:05:01Z DEBUG scale_handler Getting metrics and activity from scaler {"scaledObject.Namespace": "....", "scaledObject.Name": "....", "scaler": "prometheusScaler", "metricName": "s0-prometheus", "metrics": [], "activity": false, "scalerError": "Get \"http://prometheus-operated.monitoring.svc:9090/api/v1/query?query=sum%28tasks_held%7Bprotocol%3D%22s3%22%7D%29&time=2025-11-07T15:05:01Z\": dial tcp: lookup prometheus-operated.monitoring.svc on 10.56.0.10:53: no such host"}
ERROR 2025-11-07T15:05:01.443185323Z 2025-11-07T15:05:01Z ERROR scale_handler error getting scale decision {"scaledObject.Namespace": "...", "scaledObject.Name": "....", "scaler": "prometheusScaler", "error": "Get \"http://prometheus-operated.monitoring.svc:9090/api/v1/query?query=sum%28tasks_held%7Bprotocol%3D%22s3%22%7D%29&time=2025-11-07T15:05:01Z\": dial tcp: lookup prometheus-operated.monitoring.svc on 10.56.0.10:53: no such host"}
ERROR 2025-11-07T15:05:01.443188010Z github.com/kedacore/keda/v2/pkg/scaling.(*scaleHandler).getScalerState
ERROR 2025-11-07T15:05:01.443214917Z /workspace/pkg/scaling/scale_handler.go:815
ERROR 2025-11-07T15:05:01.443217848Z github.com/kedacore/keda/v2/pkg/scaling.(*scaleHandler).getScaledObjectState.func1
ERROR 2025-11-07T15:05:01.443220922Z /workspace/pkg/scaling/scale_handler.go:662
ERROR 2025-11-07T15:05:01.443236562Z 2025-11-07T15:05:01Z DEBUG scale_handler Removing entry from ScalersCache {"key": "scaledobject....."}
ERROR 2025-11-07T15:05:01.443242155Z 2025-11-07T15:05:01Z DEBUG scale_handler scaler error encountered, clearing scaler cache {"scaledObject.Namespace": "....", "scaledObject.Name": "...."}
ERROR 2025-11-07T15:05:01.443247830Z 2025-11-07T15:05:01Z DEBUG events Get "http://prometheus-operated.monitoring.svc:9090/api/v1/query?query=sum%28tasks_held%7Bprotocol%3D%22s3%22%7D%29&time=2025-11-07T15:05:01Z": dial tcp: lookup prometheus-operated.monitoring.svc on 10.45.0.20:53: no such host {"type": "Warning", "object": {"kind":"ScaledObject","namespace":"...","name":"....","uid":"....","apiVersion":"keda.sh/v1alpha1","resourceVersion":"1762464509434479012"}, "reason": "KEDAScalerFailed"}
ERROR 2025-11-07T15:05:01.443252113Z 2025-11-07T15:05:01Z DEBUG scaleexecutor ScaleTarget will fallback to Fallback.Replicas after Fallback.FailureThreshold {"scaledobject.Name": ".....", "scaledObject.Namespace": "....", "scaleTarget.Name": "...."}
ERROR 2025-11-07T15:05:01.543416968Z github.com/kedacore/keda/v2/pkg/scalers.(*prometheusScaler).GetMetricsAndActivity

KEDA Version

2.18.1

Kubernetes Version

1.33

Platform

Google Cloud

Scaler Details

No response

Anything else?

We are currently unable to set minReplicaCount to 0 in our KEDA ScaledObject configuration. If we do set it to 0 and Prometheus goes down, our deployment can potentially remain at 0 replicas until Prometheus is available again—even if fallback is triggered. This means our application may not recover automatically when metrics cannot be fetched.

Thank you for your help and consideration!

[roxannelandry]

I saw this issue, but I'm not sure it has been addressed:
#965

[rickbrouwer]

When a ScaledObject is scaled to zero and the scaler encounters an error, the fallback mechanism will be only triggered during HPA metric requests via GetScaledObjectMetrics, but not during the polling loop in getScaledObjectState.

This caused the ScaledObject to remain at zero replicas indefinitely when errors occurred, as the polling loop would continuously fail without activating the fallback, preventing HPA from ever being consulted.

I think I have a possible solution, but I want to take a closer look at all this.