Merge remote-tracking branch 'github/pull/550'

[chef.git] / 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 92983b27a698899e3738aa5f61474e1d00ffaa09..16496c12d0887aec40f6ead0702b45ea8e962834 100644 (file)
--- 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 > 25.5
+        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 > 25.5
+        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:
@@ -145,19 +145,19 @@ groups:
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
       - alert: fastly healthcheck failing
-        expr: count(fastly_healthcheck_status == 0) > 0
+        expr: count(fastly_healthcheck_status == 0) by (service) > 0
         for: 15m
         labels:
           alertgroup: fastly
-      - alert: multipe fastly healthchecks failing
-        expr: count(fastly_healthcheck_status == 0) > 4
+      - alert: multiple fastly healthchecks failing
+        expr: count(fastly_healthcheck_status == 0) by (service) > 4
         for: 5m
         labels:
           alertgroup: fastly
   - 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 }}"
@@ -246,18 +246,28 @@ groups:
           alertgroup: "{{ $labels.instance }}"
   - name: juniper
     rules:
+      - alert: juniper cpu alarm
+        expr: jnxOperatingCPU{jnxOperatingContentsIndex="7"} > 30
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.site }}"
       - alert: juniper fan alarm
-        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!="running"}[6m]) > 0
+        expr: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"} > 0
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
       - alert: juniper power alarm
-        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!="running"}[6m]) > 0
+        expr: jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!~"running.*"} > 0
         for: 5m
         labels:
           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
@@ -396,6 +406,23 @@ groups:
           alertgroup: overpass
         annotations:
           age: "{{ $value | humanizeDuration }}"
+  - name: passenger
+    rules:
+      - alert: passenger down
+        expr: passenger_up == 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+      - alert: passenger queuing
+        expr: passenger_top_level_request_queue > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+      - alert: passenger application queuing
+        expr: passenger_app_request_queue > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
   - name: planet
     rules:
       - alert: planet dump overdue