Lack of flexibility to install addons specific to your OS

[raelyz]

{{- if .Values.agent.enabled }}
{{- $clusterName := .Values.clusterName | required ".Values.clusterName is required." -}}
{{- $region := .Values.region | required ".Values.region is required." -}}
{{- $config := `{"logs":{"metrics_collected":{"application_signals":{}}},"traces":{"traces_collected":{"application_signals":{}}}}` | fromJson -}}
apiVersion: cloudwatch.aws.amazon.com/v1alpha1
kind: AmazonCloudWatchAgent
metadata:
  name: {{ template "cloudwatch-agent.name" . }}-windows
  namespace: {{ .Release.Namespace }}
spec:
  podSecurityContext:
    windowsOptions:
      runAsUserName: "NT AUTHORITY\\System"
  image: {{ template "cloudwatch-agent.image" . }}
  mode: daemonset
  serviceAccount: {{ template "cloudwatch-agent.serviceAccountName" . }}
  nodeSelector:
    kubernetes.io/os: windows
  config: {{ include "cloudwatch-agent.modify-config" (merge (dict "Config" $config) .) }}
  {{- with .Values.agent.resources }}
  resources: {{- toYaml . | nindent 4}}
  {{- end }}
  env:
    - name: K8S_NODE_NAME
      valueFrom:
        fieldRef:
          fieldPath: spec.nodeName
    - name: HOST_IP
      valueFrom:
        fieldRef:
          fieldPath: status.hostIP
    - name: HOST_NAME
      valueFrom:
        fieldRef:
          fieldPath: spec.nodeName
    - name: K8S_NAMESPACE
      valueFrom:
        fieldRef:
          fieldPath: metadata.namespace
  {{- with .Values.tolerations }}
  tolerations: {{- toYaml . | nindent 2}}
  {{- end }}
{{- end }}

If i'm not wrong all the daemonsets are enabled through the agent.enabled flag, this causes my helm installation on EKS to include a lot of bloat.

Is this intentional? I only wanted to install those relevant for Linux for example but I get the windows CRDs installed to my cluster as an unintended side effect.

[sky333999]

The main reason its structured the way its currently is due to the fact that we cannot pre-determine what kind of nodes would be added to a cluster from the helm chart. Hence we create both a linux and a windows daemonset with corresponding nodeSelectors enforcing only the appropriate pod runs on a particular node.

So the bloat here is really the additional CRD/daemonset resources on API server/etcd, but note that it wont spin up any unnecessary pods nor consume any actual resources.

Perhaps we could expose a helm values override to explicitly specify the OS if someone wants to avoid even creating these resources.