prole/monitoring/kps-cnpg-values.yaml

# kube-prometheus-stack (kps) Helm values — DB cluster (knoe-dev-cnpg-0).
#
# This file scopes a SECOND kube-prometheus-stack install — companion to
# `monitoring/kps-values-gke.yaml` (the app-cluster install on knoe-dev-0).
#
# Why two stacks: Kubernetes service discovery is cluster-local. The app-
# cluster Prometheus can't see CNPG pods (they live on knoe-dev-cnpg-0), so
# every cnpg-grafana dashboard rendered "No data". This stack runs Prometheus
# locally on the DB cluster, auto-discovers CNPG's PodMonitor, and is queried
# directly by the app-cluster Grafana via an internal-LB datasource.
#
# Topology:
#   knoe-dev-cnpg-0  -[this stack]-> Prometheus + node-exporter + ksm
#                                      |
#                                      v
#                                 ILB (prometheus-cnpg-ilb @ 10.180.x.x:9090)
#                                      ^
#                                      |  (datasource: cnpg-prometheus)
#   knoe-dev-0       --------------- Grafana (single canonical Grafana)
#
# Apply:
#   helm --kube-context=$DB_CLUSTER_KUBECONTEXT upgrade --install kps \
#     prometheus-community/kube-prometheus-stack --version 84.3.0 \
#     --namespace monitoring --create-namespace \
#     -f monitoring/kps-cnpg-values.yaml --wait --timeout 5m
#
# Storage: every PVC pinned to `standard-hdd` (pd-standard). The cluster also
# has `standard-rwo` (pd-balanced, partial-SSD) and `premium-rwo` (pd-ssd) —
# both would consume the SSD_TOTAL_GB regional quota that's shared with CNPG.
# Honor the standing rule: monitoring is HDD-only.
---

# One Grafana stays canonical on the app cluster (svc.knoe.dev/grafana).
# This stack is a metrics-only sidecar.
grafana:
  enabled: false

# Alerting routes through the app-cluster Alertmanager (follow-up: federation).
# For now, no AM here, no rules either (avoid scrape-rule noise into a void).
alertmanager:
  enabled: false
defaultRules:
  create: false

# DB-cluster object metrics (pods/PVCs/Jobs etc.) for kube-state dashboards.
kube-state-metrics:
  enabled: true

# Per-DB-node host metrics (CPU/mem/disk/network).
prometheus-node-exporter:
  enabled: true

prometheus:
  prometheusSpec:
    # Short retention — this Prometheus is mostly a query-target. App-cluster
    # Grafana queries it on demand; we don't archive long-form.
    retention: 7d

    # CRITICAL: select ALL PodMonitors / ServiceMonitors / PrometheusRules,
    # not just helm-release-labelled ones. The CNPG operator-emitted
    # PodMonitor at knoe-db-0/knoe-db doesn't carry our `release=kps` label,
    # but we want to scrape it. Selector with empty `{}` matches everything.
    podMonitorSelectorNilUsesHelmValues: false
    serviceMonitorSelectorNilUsesHelmValues: false
    ruleSelectorNilUsesHelmValues: false
    podMonitorSelector: {}
    serviceMonitorSelector: {}

    # Tight resource shape — small cluster, low metric cardinality (3 PG
    # pods + ~50 metrics each + 15s scrape ≈ 30k samples/day ≈ 250 MB/day).
    resources:
      requests:
        cpu: 100m
        memory: 512Mi
      limits:
        cpu: 500m
        memory: 1Gi

    storageSpec:
      volumeClaimTemplate:
        spec:
          accessModes:
            - ReadWriteOnce
          resources:
            requests:
              storage: 20Gi
          storageClassName: standard-hdd   # explicit HDD (do not let default fall through)

prometheusOperator:
  resources:
    requests:
      cpu: 50m
      memory: 128Mi
    limits:
      cpu: 200m
      memory: 256Mi