Attach extra labels to jobs

[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 521d2c92ba936a2209760eb6c63f222acc5446fa..396de8de43001ed12dbce7aeca8a26f1fc55b0df 100644 (file)
--- a/cookbooks/prometheus/templates/default/alert_rules.yml.erb
+++ b/cookbooks/prometheus/templates/default/alert_rules.yml.erb
@@ -3,6 +3,13 @@
 groups:
   - name: amsterdam
     rules:
 groups:
   - name: amsterdam
     rules:

+      - alert: uplink
+        expr: ifOperStatus{site="amsterdam",ifName=~"ge-[01]/2/2"} != 1
+        for: 6m
+        labels:
+          alertgroup: "amsterdam"
+        annotations:
+          status: "{{ $value }}"

       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="amsterdam",rPDU2PhaseStatusIndex="1"} / 10 > 28
         for: 6m
       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="amsterdam",rPDU2PhaseStatusIndex="1"} / 10 > 28
         for: 6m


@@ -11,7 +18,7 @@ groups:
         annotations:
           current: "{{ $value | humanize }}A"
       - alert: site power
         annotations:
           current: "{{ $value | humanize }}A"
       - alert: site power

-        expr: sum(avg_over_time(rPDU2PhaseStatusApparentPower{site="amsterdam",rPDU2PhaseStatusIndex="1"}[1h]) / 100) > 3
+        expr: sum(avg_over_time(rPDU2PhaseStatusApparentPower{site="amsterdam",rPDU2PhaseStatusIndex="1"}[1h]) / 100) > 3.5

         for: 6m
         labels:
           alertgroup: "amsterdam"
         for: 6m
         labels:
           alertgroup: "amsterdam"


@@ -45,13 +52,6 @@ groups:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           busy_workers: "{{ $value | humanizePercentage }}"
           alertgroup: "{{ $labels.instance }}"
         annotations:
           busy_workers: "{{ $value | humanizePercentage }}"

-      - alert: apache low request rate
-        expr: rate(apache_accesses_total[5m]) / rate(apache_accesses_total[1h] offset 1w) < 0.25 and rate(apache_accesses_total[1h] offset 1w) > 2
-        for: 15m
-        labels:
-          alertgroup: "{{ $labels.instance }}"
-        annotations:
-          request_rate: "{{ $value | humanizePercentage }}"

   - name: chef
     rules:
       - alert: chef client not running
   - name: chef
     rules:
       - alert: chef client not running


@@ -116,6 +116,13 @@ groups:
           failure_rate: "{{ $value }} jobs/s"
   - name: dublin
     rules:
           failure_rate: "{{ $value }} jobs/s"
   - name: dublin
     rules:

+      - alert: uplink
+        expr: ifOperStatus{site="dublin",ifName=~"ge-[01]/2/2"} != 1
+        for: 6m
+        labels:
+          alertgroup: "dublin"
+        annotations:
+          status: "{{ $value }}"

       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10 > 28
         for: 6m
       - alert: pdu current draw
         expr: rPDU2PhaseStatusCurrent{site="dublin",rPDU2PhaseStatusIndex="1"} / 10 > 28
         for: 6m


@@ -256,10 +263,12 @@ groups:
   - name: juniper
     rules:
       - alert: juniper cpu alarm
   - name: juniper
     rules:
       - alert: juniper cpu alarm

-        expr: jnxOperatingCPU{jnxOperatingContentsIndex="7"} > 30
+        expr: jnxOperating5MinLoadAvg{jnxOperatingContentsIndex="9"} / 200 > 0.5

         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"
         for: 5m
         labels:
           alertgroup: "{{ $labels.site }}"

+        annotations:
+          load_average: "{{ $value | humanizePercentage }}"

       - alert: juniper fan alarm
         expr: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"} > 0
         for: 5m
       - alert: juniper fan alarm
         expr: jnxOperatingState{jnxOperatingContentsIndex="4",jnxOperatingState!~"running.*"} > 0
         for: 5m


@@ -278,7 +287,7 @@ groups:
         labels:
           alertgroup: "{{ $labels.instance }}"
       - alert: exim queue length
         labels:
           alertgroup: "{{ $labels.instance }}"
       - alert: exim queue length

-        expr: exim_queue > exim_queue_limit
+        expr: exim_queue > ignoring(job) exim_queue_limit

         for: 60m
         labels:
           alertgroup: mail
         for: 60m
         labels:
           alertgroup: mail


@@ -363,14 +372,14 @@ groups:
   - name: network
     rules:
       - alert: interface transmit rate
   - name: network
     rules:
       - alert: interface transmit rate

-        expr: rate(node_network_transmit_bytes_total[1m]) / node_network_speed_bytes > 0.98
+        expr: rate(node_network_transmit_bytes_total[1m]) / node_network_speed_bytes > 0.99

         for: 5m
         labels:
           alertgroup: "{{ $labels.instance }}"
         annotations:
           bandwidth_used: "{{ $value | humanizePercentage }}"
       - alert: interface receive rate
         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
+        expr: rate(node_network_receive_bytes_total[1m]) / node_network_speed_bytes > 0.99

         for: 5m
         labels:
           alertgroup: "{{ $labels.instance }}"
         for: 5m
         labels:
           alertgroup: "{{ $labels.instance }}"


@@ -727,8 +736,8 @@ groups:
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
       - alert: job processing rate
         annotations:
           error_rate: "{{ $value | humanizePercentage }}"
       - alert: job processing rate

-        expr: rate(pg_stat_user_tables_n_tup_del{datname="openstreetmap",relname="delayed_jobs"}[5m]) / rate(pg_stat_user_tables_n_tup_ins{datname="openstreetmap",relname="delayed_jobs"}[5m]) < 0.9 and ignoring(job, name, datname, relname, schemaname, server) chef_role{name="db-master"} == 1
-        for: 15m
+        expr: rate(pg_stat_user_tables_n_tup_del{datname="openstreetmap",relname="delayed_jobs"}[1h]) / rate(pg_stat_user_tables_n_tup_ins{datname="openstreetmap",relname="delayed_jobs"}[1h]) < 0.9 and ignoring(job, name, datname, relname, schemaname, server) chef_role{name="db-master"} == 1
+        for: 1h

         labels:
           alertgroup: web
         annotations:
         labels:
           alertgroup: web
         annotations: