Kafka scaler - Need for clarification about the lagThreshold

[Vince-Chenal]

Hey,

I'm running a POC with the kafka scaler and trying to understand the way it's behaving.

Here's the content of the ScaledObject I'm using:

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: keda-kafka-poc
spec:
  scaleTargetRef:
    name: kafka-keda-poc-consumer
  minReplicaCount: 0
  maxReplicaCount: 20
  triggers:
  - type: kafka
    metadata:
      bootstrapServers: mykafka:9092
      consumerGroup: consumergroup
      topic: topic
      lagThreshold: "2000"
      offsetResetPolicy: earliest

Here's a screenshot of the metrics I'm getting when running my POC.
In order of appearance:

• The number of replicas of my consumer,
• The message rate of incoming message into the topic (msgs/sec)
• The lag computed by Kafka Minion
• The metric exported by the keda-operator-metrics-apiserver

When I look at the HPA I see that the lagThreshold is the target, but the value I'm getting for the "current value" is not the same "scale"as the Keda metric I have in my graph.

For example, when I have about 4k in the Keda metric, I see 470300m (470) in the current value:

NAME                      REFERENCE                            TARGETS              MINPODS   MAXPODS   REPLICAS   AGE
keda-hpa-keda-kafka-poc   Deployment/kafka-keda-poc-consumer   470300m/2k (avg)   1         20        10          27h

I have the impression that the current value is "per replica" (Keda metric / Nb Replicas). Am I missing something ?

Can someone clarify if the lagThreshold relates to the "total lag" or some kind of "lag per replica" ?

Thanks in advance

[zroubalik]

KEDA is feeding the metrics to HPA, which does the scaling then. You can check the algorithm explanation (and the current value calculation) in the HPA docs: https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/#algorithm-details