X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/b8d8c5b86947c933602e98057c106a2f97c040f7..41c52dd81080eb0a88e786966bed3fa6bac9cca5:/cookbooks/prometheus/templates/default/alert_rules.yml.erb?ds=inline

diff --git a/cookbooks/prometheus/templates/default/alert_rules.yml.erb b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
index f64b7fc08..dfd885bb8 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -25,14 +25,14 @@ groups:
         annotations:
           current: "{{ $value | humanize }}kVA"
       - alert: site temperature
-        expr: min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 < 18 or min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 > 26
+        expr: min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 < 15 or min(rPDU2SensorTempHumidityStatusTempC{site="amsterdam"}) / 10 > 32
         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
+        expr: max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 < 0.08 or max(rPDU2SensorTempHumidityStatusRelativeHumidity{site="amsterdam"}) / 100 > 0.8
         for: 6m
         labels:
           alertgroup: "amsterdam"
@@ -105,27 +105,20 @@ groups:
           pressure: "{{ $value | humanizePercentage }}"
   - name: database
     rules:
-      - alert: postgres replication delay
-        expr: pg_replication_lag_seconds > 30
-        for: 15m
-        labels:
-          alertgroup: database
-        annotations:
-          delay: "{{ $value | humanizeDuration }}"
       - alert: active rails queries
-        expr: sum(pg_stat_activity_count{instance="snap-01",datname="openstreetmap",usename="rails",state="active"}) > 50 and on (instance) chef_role{name="db-master"}
+        expr: sum(pg_stat_activity_count{datname="openstreetmap",usename="rails",state="active"}) by (instance) > 50 and on (instance) chef_role{name="db-master"}
         for: 5m
         labels:
           alertgroup: database
         annotations:
           queries: "{{ $value }}"
       - alert: active cgimap queries
-        expr: sum(pg_stat_activity_count{instance="snap-01",datname="openstreetmap",usename="cgimap",state="active"}) > 30 and on (instance) chef_role{name="db-master"}
+        expr: sum(pg_stat_activity_count{datname="openstreetmap",usename="cgimap",state="active"}) by (instance) > 30 and on (instance) chef_role{name="db-master"}
         for: 5m
         labels:
           alertgroup: database
         annotations:
-          delay: "{{ $value }}"
+          queries: "{{ $value }}"
   - name: discourse
     rules:
       - alert: discourse job failure rate
@@ -338,6 +331,15 @@ groups:
           alertgroup: "{{ $labels.site }}"
         annotations:
           power: "{{ $value }} dBm"
+  - name: load
+    rules:
+      - alert: load average
+        expr: sum(node_load5) by (instance) / count(node_cpu_frequency_max_hertz) by (instance) > 2
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          load: "{{ $value | humanizePercentage }}"
   - name: mail
     rules:
       - alert: exim down
@@ -428,6 +430,13 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           connections_used: "{{ $value | humanizePercentage }}"
+      - alert: mysql connection errors
+        expr: increase(mysql_global_status_connection_errors_total[1m]) > 0
+        for: 0m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          error_count: "{{ $value }}"
   - name: network
     rules:
       - alert: interface redundancy lost
@@ -488,6 +497,11 @@ groups:
           alertgroup: nominatim
         annotations:
           delay: "{{ $value | humanizeDuration }}"
+      - alert: nominatim connections
+        expr: sum(nginx_connections_writing and on (instance) chef_role{name="nominatim"}) > 2500
+        for: 15m
+        labels:
+          alertgroup: nominatim
   - name: overpass
     rules:
       - alert: overpass osm database age
@@ -612,6 +626,11 @@ groups:
         for: 10m
         labels:
           alertgroup: "prometheus"
+      - alert: node exporter text file scrape error
+        expr: node_textfile_scrape_error > 0
+        for: 10m
+        labels:
+          alertgroup: "prometheus"
   - name: raid
     rules:
       - alert: raid controller battery failed
@@ -788,7 +807,7 @@ groups:
   - name: web
     rules:
       - alert: web error rate
-        expr: sum(rate(api_call_count_total{status=~"50[0-8]|5[1-9][0-9]"}[5m])) by (instance) / sum(rate(api_call_count_total[5m])) by (instance) > 0.002
+        expr: sum(rate(api_call_count_total{status=~"50[0-8]|5[1-9][0-9]"}[5m])) by (instance) / sum(rate(api_call_count_total[5m])) by (instance) > 0.002 and sum(rate(api_call_count_total{status=~"50[0-8]|5[1-9][0-9]"}[5m])) by (instance) > 0.01
         for: 5m
         labels:
           alertgroup: web