]> git.openstreetmap.org Git - chef.git/blobdiff - cookbooks/prometheus/templates/default/alert_rules.yml.erb
Fix errors in alert rules
[chef.git] / cookbooks / prometheus / templates / default / alert_rules.yml.erb
index 30a14a471262afc886129a99075e15500452d144..6a9ad6120fd7b1fc73e0afd64f4b82460bf8eaf5 100644 (file)
@@ -115,7 +115,12 @@ groups:
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
       - alert: fastly healthcheck failing
-        expr: fastly_healthcheck_status == 0
+        expr: count(fastly_healthcheck_status == 0) > 0
+        for: 15m
+        labels:
+          alertgroup: fastly
+      - alert: fastly healthcheck failing
+        expr: count(fastly_healthcheck_status == 0) > 4
         for: 5m
         labels:
           alertgroup: fastly
@@ -336,6 +341,31 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           entries_used: "{{ $value | humanizePercentage }}"
+  - name: nominatim
+    rules:
+      - alert: nominatim replication delay
+        expr: nominatim_replication_delay > 10800
+        for: 1h
+        labels:
+          alertgroup: nominatim
+        annotations:
+          delay: "{{ $value | humanizeDuration }}"
+  - name: overpass
+    rules:
+      - alert: overpass osm database age
+        expr: overpass_database_age_seconds{database="osm"} > 300
+        for: 5m
+        labels:
+          alertgroup: overpass
+        annotations:
+          age: "{{ $value | humanizeDuration }}"
+      - alert: overpass area database age
+        expr: overpass_database_age_seconds{database="area"} > 86400
+        for: 1h
+        labels:
+          alertgroup: overpass
+        annotations:
+          age: "{{ $value | humanizeDuration }}"
   - name: planet
     rules:
       - alert: planet dump overdue
@@ -402,7 +432,7 @@ groups:
         annotations:
           connections_used: "{{ $value | humanizePercentage }}"
       - alert: postgresql deadlocks
-        expr: increase(pg_stat_database_deadlocks[1m]) > 5
+        expr: increase(pg_stat_database_deadlocks{datname!="nominatim"}[1m]) > 5
         for: 0m
         labels:
           alertgroup: "{{ $labels.instance }}"
@@ -439,6 +469,22 @@ groups:
         for: 5m
         labels:
           alertgroup: "{{ $labels.instance }}"
+  - name: rasdaemon
+    rules:
+      - alert: memory controller errors
+        expr: increase(rasdaemon_mc_events_total[1m]) > 0
+        for: 0m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          new_errors: "{{ $value }}"
+      - alert: pcie aer errors
+        expr: increase(rasdaemon_aer_events_total[1m]) > 0
+        for: 0m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          new_ercrors: "{{ $value }}"
   - name: smart
     rules:
       - alert: smart failure
@@ -447,7 +493,7 @@ groups:
         labels:
           alertgroup: "{{ $labels.instance }}"
       - alert: smart ssd wearout approaching
-        expr: smart_percentage_used >= 90
+        expr: smart_percentage_used >= 80
         for: 60m
         labels:
           alertgroup: "{{ $labels.instance }}"