prometheus-bird-role → dedicated role + various alerts

filter_plugins/prometheus.py

@@ -1,10 +1,12 @@
 from ansible.parsing.yaml.objects import AnsibleUnicode
 
+
 class FilterModule:
     def filters(self):
         return {
             "prometheus__convert_jobs": convert_jobs,
             "interp": interp,
+            "interp_float": interp_float,
         }
 
 
@@ -12,6 +14,10 @@ def interp(string):
     return AnsibleUnicode(f"{{{{ {string} }}}}")
 
 
+def interp_float(string):
+    return AnsibleUnicode(f'{{{{ printf "%.0f" {string} }}}}')
+
+
 def convert_jobs(config):
 
     for name, job in config.items():

playbooks/prometheus.yml

@@ -6,10 +6,22 @@
   roles:
     - prometheus_node
 
+- hosts:
+    - edge-1.rtr.infra.auro.re
+    - edge-2.rtr.infra.auro.re
+    - isp-1.rtr.infra.auro.re
+    - isp-2.rtr.infra.auro.re
+    - infra-1.rtr.infra.auro.re
+    - infra-2.rtr.infra.auro.re
+  roles:
+    - prometheus_bird
+
 - hosts:
     - prometheus-1.monit.infra.auro.re
     - prometheus-2.monit.infra.auro.re
   vars:
+    prometheus__alertmanager_targets:
+      - docker-ovh.adm.auro.re:9093
     prometheus__tsdb_retention_time: 90d
     prometheus__scraping:
       node:
@@ -18,7 +30,25 @@
                      | flatten }}"
         address:
           port: 9100
+      prometheus:
+        targets:
+          - prometheus-1.monit.infra.auro.re
+          - prometheus-2.monit.infra.auro.re
+        address:
+          port: 9090
+      bird:
+        targets:
+          - edge-1.rtr.infra.auro.re
+          - edge-2.rtr.infra.auro.re
+        address:
+          port: 9324
     prometheus__alert_rules:
+      prometheus:
+        - alert: PrometheusTsdbCompactionFailed
+          expr: "increase(prometheus_tsdb_compactions_failed_total[1m]) > 0"
+          for: 0m
+          labels:
+            severity: critical
       node:
         - alert: MachineDown
           expr: "up == 0"
@@ -36,7 +66,7 @@
           labels:
             severity: warning
           annotations:
-            summary: "Mémoire libre à {{ '$value' | interp }}%"
+            summary: "Mémoire libre à {{ '$value' | interp_float }}%"
         - alert: HostSwapIsFillingUp
           expr: "( 1 - ( node_memory_SwapFree_bytes
                          / node_memory_SwapTotal_bytes ) )
@@ -45,14 +75,14 @@
           labels:
             severity: critical
           annotations:
-            summary: "Swap {{ '$value' | interp }}%"
+            summary: "Swap {{ '$value' | interp_float }}%"
         - alert: HostPhysicalComponentTooHot
           expr: "node_hwmon_temp_celsius > 79"
           for: 3m
           labels:
             severity: critical
           annotations:
-            summary: "{{ '$value' | interp }}°C :
+            summary: "{{ '$value' | interp_float }}°C :
                       {{ '$labels.chip' | interp }},
                       {{ '$labels.sensor' | interp }}"
         - alert: HostNodeOvertemperatureAlarm
@@ -63,6 +93,20 @@
           annotations:
             summary: "{{ '$labels.chip' | interp }},
                       {{ '$labels.sensor' | interp }}"
+        - alert: HostRaidArrayGotInactive
+          expr: 'node_md_state{state="inactive"} > 0'
+          for: 0m
+          labels:
+            severity: critical
+          annotations:
+            summary: "{{ '$labels.device' | interp }}"
+        - alert: HostRaidDiskFailure
+          expr: 'node_md_disks{state="failed"} > 0'
+          for: 0m
+          labels:
+            severity: critical
+          annotations:
+            severity: "{{ '$labels.md_device' | interp }}"
         - alert: HostOomKillDetected
           expr: "increase(node_vmstat_oom_kill[1m]) > 0"
           for: 0m
@@ -77,15 +121,51 @@
             severity: warning
           annotations:
             summary: "{{ '$value' | interp }} erreurs corrigées"
+        - alert: HostEdacUncorrectableErrorsDetected
+          expr: "increase(node_edac_uncorrectable_errors_total[1m]) > 0"
+          for: 0m
+          labels:
+            severity: warning
+          annotations:
+            summary: "{{ '$value' | interp }} erreurs corrigées"
         - alert: OutOfDiskSpace
-          expr: "node_filesystem_free_bytes
+          expr: "( node_filesystem_free_bytes
-                 / node_filesystem_size_bytes * 100 < 10"
+                   / node_filesystem_size_bytes * 100 < 10 )
+                 and on (instance, device, mountpoint)
+                  node_filesystem_readonly == 0"
+          for: 5m
+          labels:
+            severity: critical
+          annotations:
+            summary: "{{ '$labels.mountpoint' | interp }} :
+                      {{ '$value' | interp_float }}% libre"
+        - alert: HostConntrackLimit
+          expr: "( node_nf_conntrack_entries
+                   / node_nf_conntrack_entries_limit ) * 100 > 80"
           for: 5m
           labels:
             severity: warning
           annotations:
-            summary: "{{ '$labels.mountpoint' | interp }} :
+            summary: "{{ '$value' | interp_float }}% complet"
-                      {{ '$value' | interp }}% libre"
+        - alert: HostClockSkew
+          expr: "(node_timex_offset_seconds > 0.05
+                 and deriv(node_timex_offset_seconds[5m]) >= 0)
+                 or (node_timex_offset_seconds < -0.05
+                 and deriv(node_timex_offset_seconds[5m]) <= 0)"
+          for: 2m
+          labels:
+            severity: warning
+        - alert: HostClockNotSynchronising
+          expr: "min_over_time(node_timex_sync_status[1m]) == 0
+                 and node_timex_maxerror_seconds >= 16"
+          for: 2m
+          labels:
+            severity: warning
+        - alert: HostRequiresReboot
+          expr: "node_reboot_required > 0"
+          for: 5m
+          labels:
+            severity: warning
         - alert: OutOfInodes
           expr: "node_filesystem_files_free
                  / node_filesystem_files * 100 < 10"
@@ -94,7 +174,7 @@
             severity: warning
           annotations:
             summary: "{{ '$labels.mountpoint' | interp }} :
-                      {{ '$value' | interp }}% libre"
+                      {{ '$value' | interp_float }}% libre"
         - alert: CpuUsage
           expr: '( 100 - avg by (instance)
                    ( irate(node_cpu_seconds_total{mode="idle"}[5m]) ) 
@@ -103,7 +183,7 @@
           labels:
             severity: warning
           annotations:
-            summary: "{{ '$value' | interp }}%"
+            summary: "{{ '$value' | interp_float }}%"
         - alert: SystemdServiceFailed
           expr: 'node_systemd_unit_state{state="failed"} == 1'
           for: 10m
@@ -117,7 +197,7 @@
           labels:
             severity: warning
           annotations:
-            summary: "{{ '$value' | interp }}"
+            summary: "{{ '$value' | interp_float }}"
         - alert: UnhealthyDisk
           expr: "smartmon_device_smart_healthy < 1"
           for: 10m
@@ -125,6 +205,24 @@
             severity: critical
           annotations:
             summary: "{{ '$labels.disk' | interp }}"
+        - alert: HostCpuStealNoisyNeighbor
+          expr: 'avg by(instance)
+                  (rate(node_cpu_seconds_total{mode="steal"}[5m]))
+                  * 100 > 10'
+          for: 5m
+          labels:
+            severity: warning
+          annotations:
+            summary: "{{ '$labels.disk' | interp }}"
+      bird:
+        - alert: BirdProtocolDown
+          expr: "bird_protocol_up == 0"
+          for: 0m
+          labels:
+            severity: critical
+          annotations:
+            summary: "{{ '$labels.name' | interp }} :
+                      {{ '$labels.state' | interp }}"
   roles:
     - prometheus
 
@@ -332,38 +430,7 @@
 #    - prometheus
 #    - update_motd
 #
-#- hosts: prometheus-federate.adm.auro.re
-#  vars:
-#    prometheus_alertmanager: docker-ovh.adm.auro.re:9093
-#    snmp_unifi_password: "{{ vault_snmp_unifi_password }}"
-#    snmp_pdu_user: "{{ vault_snmp_pdu_user }}"
-#    snmp_pdu_password: "{{ vault_snmp_pdu_password }}"
-#    snmp_ilo_user: aurore
-#    snmp_ilo_auth: "{{ vault_snmp_ilo_auth }}"
-#    snmp_ilo_priv: "{{ vault_snmp_ilo_priv }}"
-#
-#    prometheus_servers_targets:
-#      - prometheus-edc.adm.auro.re
-#      - prometheus-gs.adm.auro.re
-#      - prometheus-fleming.adm.auro.re
-#      - prometheus-pacaterie.adm.auro.re
-#      - prometheus-rives.adm.auro.re
-#      - prometheus-aurore.adm.auro.re
-#      - prometheus-ovh.adm.auro.re
-#
-#    update_motd:
-#      prometheus_federate: >-
-#        Prometheus (en configuration fédération) est déployé (/etc/prometheus).
-#  roles:
-#    - prometheus_federate
-#    - update_motd
-#
 ## Postgres Exporters
 #- hosts: bdd,radius
 #  roles:
 #    - prometheus_postgres
-#
-## Monitor all hosts
-#- hosts: all,!edc_unifi,!fleming_unifi,!pacaterie_unifi,!gs_unifi,!rives_unifi,!aurore_testing_vm,!ovh_container
-#  roles:
-#    - prometheus_node

roles/bird/defaults/main.yml

@@ -11,5 +11,4 @@ bird__radv_dns_servers: []
 bird__radv_max_interval: 5
 bird__static_unreachable: []
 bird__bgp_sessions: []
-bird__prometheus_listen_address: 0.0.0.0:9324
 ...

roles/bird/handlers/main.yml

@@ -3,9 +3,4 @@
   systemd:
     name: bird.service
     state: reloaded
-
-- name: Restart prometheus-bird-exporter
-  systemd:
-    name: prometheus-bird-exporter.service
-    state: restarted
 ...

roles/bird/tasks/main.yml

@@ -1,9 +1,7 @@
 ---
 - name: Install bird
   apt:
-    name:
+    name: bird2
-      - bird2
-      - prometheus-bird-exporter
 
 - name: Configure bird
   template:
@@ -15,26 +13,9 @@
   notify:
     - Reload bird
 
-- name: Configure prometheus-bird-exporter
-  template:
-    src: prometheus-bird-exporter.j2
-    dest: /etc/default/prometheus-bird-exporter
-    owner: root
-    group: root
-    mode: u=rw,g=r,o=
-  notify:
-    - Restart prometheus-bird-exporter
-
 - name: Enable and start bird
   systemd:
     name: bird.service
     state: started
     enabled: true
-
-- name: Enable and start prometheus-bird-exporter
-  systemd:
-    name: prometheus-bird-exporter.service
-    state: started
-    enabled: true
-
 ...

roles/prometheus_bird/defaults/main.yml

@@ -0,0 +1,3 @@
+---
+bird__prometheus_listen_address: ':9324'
+...

roles/prometheus_bird/handlers/main.yml

@@ -0,0 +1,6 @@
+---
+- name: Restart prometheus-bird-exporter
+  systemd:
+    name: prometheus-bird-exporter.service
+    state: restarted
+...

roles/prometheus_bird/tasks/main.yml

@@ -0,0 +1,21 @@
+---
+- name: Install prometheus-bird-exporter
+  apt:
+      name: prometheus-bird-exporter
+
+- name: Configure prometheus-bird-exporter
+  template:
+    src: prometheus-bird-exporter.j2
+    dest: /etc/default/prometheus-bird-exporter
+    owner: root
+    group: root
+    mode: u=rw,g=r,o=
+  notify:
+    - Restart prometheus-bird-exporter
+
+- name: Enable and start prometheus-bird-exporter
+  systemd:
+    name: prometheus-bird-exporter.service
+    state: started
+    enabled: true
+...