VPA assigns minAllowed.memory instead of actual recommendation when restarting Prometheus pod if Limitrange is set

[MatteoMori8]

Which component are you using?:

• vertical-pod-autoscaler

What version of the component are you using?:

• 1.2.1

Component version:

What k8s version are you using (kubectl version)?:

Server Version: v1.29.8-eks

What environment is this in?:

• AWS - eks

What did you expect to happen?:
I expected VPA's AdmissionController to patch my Prometheus pod with a request.memory value equal to the memory target recommendation

What happened instead?:
Instead, the new pod got deployed with the following:

• limit.memory got halved
• request.memory has been assigned the value that I specified as minAllowed.memory from the VPA object itself

On the other hand, CPU got updated correctly

How to reproduce it (as minimally and precisely as possible):

VPA object

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: prometheus-test
spec:
  resourcePolicy:
    containerPolicies:
    - containerName: prometheus
      controlledResources:
      - cpu
      - memory
      controlledValues: RequestsAndLimits
      maxAllowed:
        memory: 30Gi
      minAllowed:
        cpu: 100m
        memory: 500Mi
    - containerName: thanos-sidecar
      controlledResources:
      - cpu
      - memory
      controlledValues: RequestsAndLimits
      maxAllowed:
        memory: 20Gi
      minAllowed:
        memory: 100Mi
  targetRef:
    apiVersion: apps/v1
    kind: StatefulSet
    name: prometheus-test
  updatePolicy:
    updateMode: "Initial"

LimitRange

apiVersion: v1
kind: LimitRange
metadata:
  name: limits
spec:
  limits:
  - max:
      memory: 60Gi
    min:
      cpu: 1m
      memory: 50Mi
    type: Pod
  - default:
      memory: 512Mi
    defaultRequest:
      cpu: 20m
      memory: 256Mi
    min:
      memory: 50Mi
    type: Container

Anything else we need to know?:

We talked this out on the #sig-autoscaling Slack channel and it turns out that by deleting the LimitRange, VPA goes back on track and manages memory in the appropriate way.

Personally I do not know if this is a bug or an opportunity to enhance the docs but we thought it could be good to raise an Issue about it

[adrianmoisey]

/area vertical-pod-autoscaler

[adrianmoisey]

/assign

[adrianmoisey]

Does the Prometheus statefulset define any resources?

[adrianmoisey]

Instead, the new pod got deployed with the following:

• limit.memory got halved
• request.memory has been assigned the value that I specified as minAllowed.memory from the VPA object itself

Which container in the Pod got these set?