Make fastly healthcheck alerts operate on a per-service basis

[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 df1329147e5b6f585fdaa555c8565e771a31c645..8c1aeee48bea93bfc9ec168c698b0ac45babdbb5 100644 (file)
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -105,6 +105,36 @@ groups:
           alertgroup: database
         annotations:
           delay: "{{ $value | humanizeDuration }}"
+  - name: dublin
+    rules:
+      - alert: pdu current draw
+        expr: rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10 > 13
+        for: 5m
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          current: "{{ $value | humanize }}A"
+      - alert: site current draw
+        expr: sum(rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10) > 17
+        for: 5m
+        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
+        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
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          humidity: "{{ $value | humanizePercentage }}"
   - name: fastly
     rules:
       - alert: fastly error rate
@@ -115,12 +145,12 @@ 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
@@ -217,12 +247,12 @@ groups:
   - name: juniper
     rules:
       - alert: juniper fan alarm
-        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!="running"}[5m]) > 0
+        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"}[6m]) > 0
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
       - alert: juniper power alarm
-        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!="running"}[5m]) > 0
+        expr: sum_over_time(jnxOperatingState{jnxOperatingContentsIndex="2",jnxOperatingState!~"running.*"}[6m]) > 0
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
@@ -353,8 +383,8 @@ groups:
   - name: overpass
     rules:
       - alert: overpass osm database age
-        expr: overpass_database_age_seconds{database="osm"} > 300
-        for: 5m
+        expr: overpass_database_age_seconds{database="osm"} > 3600
+        for: 1h
         labels:
           alertgroup: overpass
         annotations: