Sflow integration

common/roles/partition-prometheus-rules/README.md

@@ -14,3 +14,4 @@ Contains default variables for partition prometheus alert manager rules. This wa
 | partition_prometheus_rules_node_exporter     |           | node-exporter alert rules     |
 | partition_prometheus_rules_pixiecore         |           | pixiecore alert rules         |
 | partition_prometheus_rules_sonic_exporter    |           | sonic-exporter alert rules    |
+| partition_prometheus_rules_goflow2           |           | goflow2 sFlow collector rules |

common/roles/partition-prometheus-rules/defaults/main.yaml

@@ -343,3 +343,38 @@ partition_prometheus_rules_sonic_exporter: |
       annotations:
         description: "Sonic switch {{$labels.instance}} in partition {{$externalLabels.partition}} could not be scraped for more than 10 minutes."
   {% endraw %}
+
+partition_prometheus_rules_sonic_systemd: |
+  {% raw %}
+  - name: sonic-systemd.rules
+    rules:
+    - record: sonic:systemd_service_uptime_seconds
+      expr: time() - node_systemd_unit_start_time_seconds{name=~"(swss|bgp).service"}
+    - alert: SonicServiceNotActive
+      expr: node_systemd_unit_state{name=~"(swss|bgp).service", state="active"} == 0
+      for: 5m
+      labels:
+        severity: critical
+      annotations:
+        description: "SONiC service {{ $labels.name }} on {{ $labels.instance }} is not active."
+    - alert: SonicServiceRestarted
+      expr: changes(node_systemd_unit_start_time_seconds{name=~"(swss|bgp).service"}[8h]) > 0
+      labels:
+        severity: critical
+      annotations:
+        description: "SONiC service {{ $labels.name }} on {{ $labels.instance }} has restarted in the last 8 hours."
+  {% endraw %}
+
+partition_prometheus_rules_goflow2: |
+  {% raw %}
+  - name: goflow2-collector.rules
+    rules:
+    - alert: GoflowCollectorDown
+      expr: up{job="goflow2"} == 0
+      for: 2m
+      labels:
+        severity: critical
+      annotations:
+        summary: "goflow2 sFlow collector down ({{ $labels.instance }})"
+        description: "goflow2 instance {{ $labels.instance }} has been unreachable for more than 2 minutes. sFlow ingestion is halted."
+  {% endraw %}

control-plane/roles/monitoring/tasks/main.yaml

@@ -106,6 +106,8 @@
     - rethinkdb.yaml
     - alertmanager.yaml
     - sonic-exporter.yaml
+    - goflow2-north-south.yaml
+    - goflow2-east-west.yaml
   when: monitoring_grafana_enabled
 
 - name: Deploy Grafana dashboards for Gardener