X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/2c1479b9b228b234ea3514060ac2071e4caa1e4a..98c8d0026a2836abf350ac2a393930ff81cef41c:/cookbooks/prometheus/templates/default/alert_rules.yml.erb

diff --git a/cookbooks/prometheus/templates/default/alert_rules.yml.erb b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
index 3dd1af873..367e07650 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -5,28 +5,28 @@ groups:
     rules:
       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="amsterdam",rPDU2PhaseStatusIndex="1"} / 10 > 10
-        for: 5m
+        for: 6m
         labels:
           alertgroup: "amsterdam"
         annotations:
           current: "{{ $value | humanize }}A"
       - alert: site current draw
         expr: sum(rPDU2PhaseStatusCurrent{site="amsterdam",rPDU2PhaseStatusIndex="1"} / 10) > 13
-        for: 5m
+        for: 6m
         labels:
           alertgroup: "amsterdam"
         annotations:
           current: "{{ $value | humanize }}A"
       - alert: site temperature
-        expr: min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 < 18 or min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 > 25
-        for: 5m
+        expr: min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 < 18 or min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 > 26
+        for: 6m
         labels:
           alertgroup: "amsterdam"
         annotations:
           temperature: "{{ $value | humanize }}C"
       - alert: site humidity
         expr: max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 < 0.25 or max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 > 0.65
-        for: 5m
+        for: 6m
         labels:
           alertgroup: "amsterdam"
         annotations:
@@ -109,28 +109,28 @@ groups:
     rules:
       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10 > 13
-        for: 5m
+        for: 6m
         labels:
           alertgroup: "dublin"
         annotations:
           current: "{{ $value | humanize }}A"
       - alert: site current draw
         expr: sum(rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10) > 17
-        for: 5m
+        for: 6m
         labels:
           alertgroup: "dublin"
         annotations:
           current: "{{ $value | humanize }}A"
       - alert: site temperature
-        expr: min(rPDU2SensorTempHumidityStatusTempC{site="dublin"}) / 10 < 18 or min(rPDU2SensorTempHumidityStatusTempC{site="dublin"}) / 10 > 25
-        for: 5m
+        expr: min(rPDU2SensorTempHumidityStatusTempC{site="dublin"}) / 10 < 18 or min(rPDU2SensorTempHumidityStatusTempC{site="dublin"}) / 10 > 26
+        for: 6m
         labels:
           alertgroup: "dublin"
         annotations:
           temperature: "{{ $value | humanize }}C"
       - alert: site humidity
         expr: max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="dublin"}) / 100 < 0.25 or max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="dublin"}) / 100 > 0.65
-        for: 5m
+        for: 6m
         labels:
           alertgroup: "dublin"
         annotations:
@@ -157,7 +157,7 @@ groups:
   - name: filesystem
     rules:
       - alert: readonly filesystem
-        expr: node_filesystem_readonly == 1
+        expr: node_filesystem_readonly > min_over_time(node_filesystem_readonly[7d])
         for: 0m
         labels:
           alertgroup: "{{ $labels.instance }}"
@@ -263,6 +263,11 @@ groups:
           alertgroup: "{{ $labels.site }}"
   - name: mail
     rules:
+      - alert: exim down
+        expr: exim_up == 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
       - alert: exim queue length
         expr: exim_queue > exim_queue_limit
         for: 60m
@@ -332,6 +337,20 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           new_oom_kills: "{{ $value }}"
+  - name: mysql
+    rules:
+      - alert: mysql down
+        expr: mysql_up == 0
+        for: 1m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+      - alert: mysql connection limit
+        expr: mysql_global_status_max_used_connections / mysql_global_variables_max_connections > 0.8
+        for: 1m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          connections_used: "{{ $value | humanizePercentage }}"
   - name: network
     rules:
       - alert: interface transmit rate
@@ -536,7 +555,7 @@ groups:
         labels:
           alertgroup: "{{ $labels.instance }}"
         annotations:
-          new_ercrors: "{{ $value }}"
+          new_errors: "{{ $value }}"
   - name: smart
     rules:
       - alert: smart failure
@@ -551,6 +570,15 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           percentage_used: "{{ $value | humanizePercentage }}"
+  - name: snmp
+    rules:
+      - alert: snmp pdus missing
+        expr: max_over_time(snmp_scrape_pdus_returned[1d]) - snmp_scrape_pdus_returned > 0
+        for: 15m
+        labels:
+          alertgroup: snmp
+        annotations:
+          missing_pdus: "{{ $value }}"
   - name: ssl
     rules:
       - alert: ssl certificate probe failed
@@ -610,6 +638,13 @@ groups:
           alertgroup: tile
         annotations:
           miss_rate: "{{ $value | humanizePercentage }}"
+      - alert: tile render rate
+        expr: sum(rate(renderd_zoom_metatiles_total[5m])) by (instance) == 0
+        for: 15m
+        labels:
+          alertgroup: tile
+        annotations:
+          render_rate: "{{ $value }} tiles/s"
   - name: time
     rules:
       - alert: clock not synchronising