Issues using ignoreDifferences with HPA

[jgutschon]

I'm having some issues using ignoreDifferences on a few applications that are running Deployments with an HPA. I'm setting ignoreDifferences on the Application to ignore fields managed by kube-controller-manager:

ignoreDifferences:
    - group: apps
      kind: Deployment
      managedFieldsManagers:
        - kube-controller-manager

and I've set the sync policy as:

syncPolicy:
    syncOptions:
      - RespectIgnoreDifferences=true

My goal here is to ignore the replicas field on the deployment since it will be managed by kube-controller-manager with the HPA enabled for the deployment. I'm aware of the recommendation to omit this field entirely if using an HPA, however I've chosen to set it as the HPA's minReplicas to prevent scaling down to one pod if the Deployment is recreated in some scenarios.

This has worked fairly well in general, however, if the replicas field in the Deployment manifest is manually edited and applied in the ArgoCD UI, the kube-controller-manager managed field is overwritten and argocd-server becomes the manager:

- apiVersion: apps/v1
  fieldsType: FieldsV1
  fieldsV1:
    'f:spec':
      'f:replicas': {}
  manager: argocd-server
  operation: Update
  time: '2023-06-21T15:02:04Z'

After this change, the deployment becomes out of sync and shows a diff on the replicas field. Syncing again after this will then cause it to scale down to value in the manifest. After syncing, the replicas field becomes managed by argocd-controller. The HPA still functions normally and scales as needed but the deployment stays out of sync as the replicas value changes.

The only way I've been able to restore the managedFields for kube-controller-manager at this point has been to orphan-delete the deployment and recreate it, but any future manual edits to the replicas field reproduce this behavior. Not exactly sure if this is intended behavior so I'm wondering if anyone knows a way around this issue.

[ricardojdsilva87]

Hello @jgutschon have you tried to ignore the spec.replicas field on the Deployment?
https://blog.argoproj.io/new-sync-and-diff-strategies-in-argocd-44195d3f8b8c

It is working in my case, but I've checked a strange behaviour that for example the Deployment for argocd helm chart does not add the replicas field if autoscaling is enabled:
https://github.com/argoproj/argo-helm/blob/main/charts/argo-cd/templates/argocd-server/deployment.yaml#L19

We have ArgoCD managing itself and in this case with autoscaling enabled the replicas differences are not "seen" by ArgoCD

Tried to do the same on a custom deployment and it seems that a replicas: 1 field is added by default to the Deployment if not added by the Helm Chart. Not sure if this is because it's argo managing itself and it ignores this field by default.
We went for the example that I've posted here, because for some reason the managedFields do not appear on the Deployment even though we have an HPA fully working

[kevin-bates]

We simply remove the replicas entry from the Deployment and let the minReplicas entry of the HPA definition be the initial desired replicas value. This essentially tells argo that replicas is unmanaged.

[RickyDepop]

What if I am using an external Helm template I don't have control on?