X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/326c0527e3a2616f9f808d14d533f7c761c73616..c810e6117c2e4e131aa859942566e2734e815148:/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 d29b7272d..fd25f2523 100644
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -8,6 +8,20 @@ groups:
         for: 5m
         labels:
           alertgroup: "prometheus"
+  - name: apache
+    rules:
+      - alert: apache down
+        expr: apache_up == 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+      - alert: apache workers busy
+        expr: sum(apache_workers{state="busy"}) by (instance) / sum(apache_scoreboard) by (instance) > 0.8
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          busy_workers: "{{ $value | humanizePercentage }}"
   - name: database
     rules:
       - alert: postgres replication delay
@@ -68,6 +82,11 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           voltage_volts: "{{ with printf \"ipmi_voltage_volts{instance='%s',id='%s'}\" $labels.instance $labels.id | query }}{{ . | first | value | humanize }}V{{end}}"
+      - alert: ipmi power alarm
+        expr: ipmi_power_state > 0 or ipmi_sensor_state{type=~"Power .*"} > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
   - name: mdadm
     rules:
       - alert: mdadm array inactive
@@ -113,6 +132,102 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           new_oom_kills: "{{ $value }}"
+  - name: network
+    rules:
+      - alert: interface transmit rate
+        expr: rate(node_network_transmit_bytes_total[1m]) / node_network_speed_bytes > 0.98
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          bandwidth_used: "{{ $value | humanizePercentage }}"
+      - alert: interface receive rate
+        expr: rate(node_network_receive_bytes_total[1m]) / node_network_speed_bytes > 0.98
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          bandwidth_used: "{{ $value | humanizePercentage }}"
+      - alert: interface transmit errors
+        expr: rate(node_network_transmit_errs_total[1m]) / rate(node_network_transmit_packets_total[1m]) > 0.01
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          error_rate: "{{ $value | humanizePercentage }}"
+      - alert: interface receive errors
+        expr: rate(node_network_receive_errs_total[1m]) / rate(node_network_receive_packets_total[1m]) > 0.01
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          error_rate: "{{ $value | humanizePercentage }}"
+      - alert: conntrack entries
+        expr: node_nf_conntrack_entries / node_nf_conntrack_entries_limit > 0.8
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          entries_used: "{{ $value | humanizePercentage }}"
+  - name: postgresql
+    rules:
+      - alert: postgresql down
+        expr: pg_up == 0
+        for: 1m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+      - alert: postgresql replication delay
+        expr: pg_replication_lag_seconds > 5
+        for: 1m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          delay: "{{ $value | humanizeDuration }}"
+      - alert: postgresql connection limit
+        expr: sum (pg_stat_activity_count) by (instance, server) / sum (pg_settings_max_connections) by (instance, server) > 0.8
+        for: 1m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          connections_used: "{{ $value | humanizePercentage }}"
+      - alert: postgresql deadlocks
+        expr: increase(pg_stat_database_deadlocks[1m]) > 5
+        for: 0m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          new_deadlocks: "{{ $value }}"
+      - alert: postgresql slow queries
+        expr: pg_slow_queries > 0
+        for: 5m
+        labels:
+          alertgroup: "{{ $labels.instance }}"
+        annotations:
+          queries: "{{ $value }}"
+  - name: ssl
+    rules:
+      - alert: ssl certificate probe failed
+        expr: ssl_probe_success == 0
+        for: 60m
+        labels:
+          alertgroup: ssl
+      - alert: ssl certificate expiry
+        expr: ssl_verified_cert_not_after{chain_no="0"} - time() < 86400 * 14
+        for: 0m
+        labels:
+          alertgroup: ssl
+        annotations:
+          expires_in: "{{ $value | humanizeDuration }}"
+      - alert: ssl certificate revoked
+        expr: ssl_ocsp_response_status == 1
+        for: 0m
+        labels:
+          alertgroup: ssl
+      - alert: ocsp status unknown
+        expr: ssl_ocsp_response_status == 1
+        for: 0m
+        labels:
+          alertgroup: ssl
   - name: tile
     rules:
       - alert: renderd replication delay