HetznerTerra/ansible/roles/observability/tasks/main.yml

---
- name: Check if Helm is installed
  command: helm version --short
  register: helm_check
  changed_when: false
  failed_when: false

- name: Install Helm
  shell: curl -fsSL https://raw.githubusercontent.com/helm/helm/main/scripts/get-helm-3 | bash
  when: helm_check.rc != 0
  changed_when: true

- name: Ensure observability namespace exists
  command: kubectl create namespace {{ observability_namespace }}
  register: create_observability_ns
  failed_when: create_observability_ns.rc != 0 and "AlreadyExists" not in create_observability_ns.stderr
  changed_when: create_observability_ns.rc == 0

- name: Set Grafana admin password
  set_fact:
    grafana_password_effective: "{{ grafana_admin_password if grafana_admin_password | length > 0 else lookup('password', '/dev/null length=32 chars=ascii_letters,digits') }}"

- name: Write kube-prometheus-stack values
  template:
    src: kube-prometheus-stack-values.yaml.j2
    dest: /tmp/kube-prometheus-stack-values.yaml
    mode: "0644"

- name: Add Prometheus Helm repo
  command: helm repo add prometheus-community https://prometheus-community.github.io/helm-charts
  register: add_prom_repo
  failed_when: add_prom_repo.rc != 0 and "already exists" not in add_prom_repo.stderr
  changed_when: add_prom_repo.rc == 0

- name: Add Grafana Helm repo
  command: helm repo add grafana https://grafana.github.io/helm-charts
  register: add_grafana_repo
  failed_when: add_grafana_repo.rc != 0 and "already exists" not in add_grafana_repo.stderr
  changed_when: add_grafana_repo.rc == 0

- name: Update Helm repos
  command: helm repo update
  changed_when: false

- name: Install kube-prometheus-stack
  command: >-
    helm upgrade --install kube-prometheus-stack prometheus-community/kube-prometheus-stack
    --namespace {{ observability_namespace }}
    --version {{ prometheus_chart_version }}
    --values /tmp/kube-prometheus-stack-values.yaml
    --wait
    --timeout 10m
  changed_when: true

- name: Write Loki values
  template:
    src: loki-values.yaml.j2
    dest: /tmp/loki-values.yaml
    mode: "0644"
  when: loki_enabled

- name: Validate Loki chart produces resources
  command: >-
    helm template loki grafana/loki
    --namespace {{ observability_namespace }}
    --version {{ loki_chart_version }}
    --values /tmp/loki-values.yaml
  register: loki_template
  changed_when: false
  failed_when: "loki_template.rc != 0 or 'kind: StatefulSet' not in loki_template.stdout"
  when: loki_enabled

- name: Remove legacy Loki resources
  command: >-
    kubectl -n {{ observability_namespace }} delete
    deployment/loki-gateway
    statefulset/loki
    statefulset/loki-chunks-cache
    statefulset/loki-results-cache
    statefulset/loki-backend
    statefulset/loki-read
    statefulset/loki-write
    poddisruptionbudget/loki-memcached-chunks-cache
    poddisruptionbudget/loki-memcached-results-cache
    --ignore-not-found=true
  changed_when: false
  failed_when: false
  when: loki_enabled

- name: Clear stuck Helm lock for Loki
  command: kubectl -n {{ observability_namespace }} delete secret sh.helm.release.v1.loki.v1 --ignore-not-found=true
  changed_when: false
  failed_when: false
  when: loki_enabled

- name: Uninstall failed Loki release (if stuck)
  command: helm uninstall loki -n {{ observability_namespace }}
  changed_when: false
  failed_when: false
  when: loki_enabled

- name: Install Loki
  command: >-
    helm upgrade --install loki grafana/loki
    --namespace {{ observability_namespace }}
    --version {{ loki_chart_version }}
    --values /tmp/loki-values.yaml
  register: loki_install
  changed_when: true
  when: loki_enabled

- name: Wait for Loki StatefulSet
  command: kubectl -n {{ observability_namespace }} rollout status statefulset/loki --timeout=10m
  register: loki_rollout
  changed_when: false
  when: loki_enabled

- name: Show Loki pod status
  command: kubectl -n {{ observability_namespace }} get pods -l app.kubernetes.io/name=loki -o wide
  register: loki_pods
  changed_when: false
  when: loki_enabled

- name: Debug Loki pods
  debug:
    msg: "{{ loki_pods.stdout }}"
  when: loki_enabled

- name: Write Promtail values
  template:
    src: promtail-values.yaml.j2
    dest: /tmp/promtail-values.yaml
    mode: "0644"
  when: loki_enabled

- name: Install Promtail
  command: >-
    helm upgrade --install promtail grafana/promtail
    --namespace {{ observability_namespace }}
    --version {{ promtail_chart_version }}
    --values /tmp/promtail-values.yaml
    --wait
    --timeout 10m
  changed_when: true
  when: loki_enabled

- name: Write Grafana Loki datasource manifest
  template:
    src: grafana-datasource-loki.yaml.j2
    dest: /tmp/grafana-datasource-loki.yaml
    mode: "0644"
  when: loki_enabled

- name: Create Grafana Loki datasource
  command: kubectl apply -f /tmp/grafana-datasource-loki.yaml
  changed_when: true
  when: loki_enabled

- name: Check Tailscale service readiness for Grafana
  command: kubectl -n {{ observability_namespace }} get svc kube-prometheus-stack-grafana -o jsonpath='{.status.conditions[?(@.type=="TailscaleProxyReady")].status}'
  register: grafana_tailscale_ready
  changed_when: false
  failed_when: false
  when:
    - observability_tailscale_expose | bool
    - tailscale_operator_ready | default(false) | bool

- name: Check Tailscale service readiness for Prometheus
  command: kubectl -n {{ observability_namespace }} get svc kube-prometheus-stack-prometheus -o jsonpath='{.status.conditions[?(@.type=="TailscaleProxyReady")].status}'
  register: prometheus_tailscale_ready
  changed_when: false
  failed_when: false
  when:
    - observability_tailscale_expose | bool
    - tailscale_operator_ready | default(false) | bool

- name: Check Tailscale endpoint (IP/hostname) for Grafana
  shell: >-
    kubectl -n {{ observability_namespace }} get svc kube-prometheus-stack-grafana
    -o go-template='{{"{{"}}range .status.loadBalancer.ingress{{"}}"}}{{"{{"}}if .ip{{"}}"}}{{"{{"}}.ip{{"}}"}}{{"{{"}}else{{"}}"}}{{"{{"}}.hostname{{"}}"}}{{"{{"}}end{{"}}"}}{{"{{"}}end{{"}}"}}'
  register: grafana_lb_ip
  changed_when: false
  failed_when: false
  when:
    - observability_tailscale_expose | bool
    - tailscale_operator_ready | default(false) | bool

- name: Check Tailscale endpoint (IP/hostname) for Prometheus
  shell: >-
    kubectl -n {{ observability_namespace }} get svc kube-prometheus-stack-prometheus
    -o go-template='{{"{{"}}range .status.loadBalancer.ingress{{"}}"}}{{"{{"}}if .ip{{"}}"}}{{"{{"}}.ip{{"}}"}}{{"{{"}}else{{"}}"}}{{"{{"}}.hostname{{"}}"}}{{"{{"}}end{{"}}"}}{{"{{"}}end{{"}}"}}'
  register: prometheus_lb_ip
  changed_when: false
  failed_when: false
  when:
    - observability_tailscale_expose | bool
    - tailscale_operator_ready | default(false) | bool

- name: Show Tailscale access details
  debug:
    msg: |
      Observability stack deployed with Tailscale access!

      Grafana: http://{{ grafana_tailscale_hostname }}{% if grafana_lb_ip.stdout | default('') | length > 0 %} (or http://{{ grafana_lb_ip.stdout }}){% endif %}
      Prometheus: http://{{ prometheus_tailscale_hostname }}{% if prometheus_lb_ip.stdout | default('') | length > 0 %} (or http://{{ prometheus_lb_ip.stdout }}){% endif %}

      Login: admin / {{ grafana_password_effective }}

      Tailscale readiness:
      - Grafana proxy ready: {{ grafana_tailscale_ready.stdout | default('pending') }}
      - Prometheus proxy ready: {{ prometheus_tailscale_ready.stdout | default('pending') }}

      Access via:
      - MagicDNS: http://{{ grafana_tailscale_hostname }} and http://{{ prometheus_tailscale_hostname }}
      - Tailnet FQDN: http://{{ grafana_tailscale_hostname }}.{{ tailscale_tailnet | default('tailnet.ts.net') }}
      - Direct endpoint: {% if grafana_lb_ip.stdout | default('') | length > 0 %}http://{{ grafana_lb_ip.stdout }}{% else %}(pending){% endif %} / {% if prometheus_lb_ip.stdout | default('') | length > 0 %}http://{{ prometheus_lb_ip.stdout }}{% else %}(pending){% endif %}
  when:
    - observability_tailscale_expose | bool
    - tailscale_operator_ready | default(false) | bool

- name: Show observability access details (fallback)
  debug:
    msg: |
      Observability stack deployed.
      Namespace: {{ observability_namespace }}
      Grafana (tailnet): kubectl -n {{ observability_namespace }} port-forward svc/kube-prometheus-stack-grafana 3000:80
      Prometheus (tailnet): kubectl -n {{ observability_namespace }} port-forward svc/kube-prometheus-stack-prometheus 9090:9090
      Grafana admin password: {{ grafana_password_effective }}
      {% if loki_enabled %}
      Loki: Enabled - logs available in Grafana
      {% else %}
      Loki: Disabled
      {% endif %}
  when:
    - not (observability_tailscale_expose | bool and (tailscale_operator_ready | default(false) | bool))