# DO NOT EDIT - This file is being maintained by Chef groups: - name: alertmanager rules: - alert: prometheus target missing expr: up == 0 for: 5m labels: alertgroup: "prometheus" - name: hwmon rules: - alert: hwmon fan alarm expr: node_hwmon_fan_alarm == 1 for: 5m labels: alertgroup: "{{ $labels.instance }}" annotations: fan_rpm: "{{ with printf \"node_hwmon_fan_rpm{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}rpm{{end}}" fan_min_rpm: "{{ with printf \"node_hwmon_fan_min_rpm{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}rpm{{end}}" - alert: hwmon temperature alarm expr: node_hwmon_temp_alarm == 1 for: 5m labels: alertgroup: "{{ $labels.instance }}" annotations: temp_celsius: "{{ with printf \"node_hwmon_temp_celsius{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}C{{end}}" temp_max_celsius: "{{ with printf \"node_hwmon_temp_max_celsius{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}C{{end}}" temp_crit_celsius: "{{ with printf \"node_hwmon_temp_crit_celsius{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}C{{end}}" - alert: hwmon voltage alarm expr: node_hwmon_in_alarm == 1 for: 5m labels: alertgroup: "{{ $labels.instance }}" annotations: in_volts: "{{ with printf \"node_hwmon_in_volts{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}V{{end}}" in_min_volts: "{{ with printf \"node_hwmon_in_min_volts{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}V{{end}}" in_max_volts: "{{ with printf \"node_hwmon_in_max_volts{instance='%s',chip='%s',sensor='%s'}\" $labels.instance $labels.chip $labels.sensor | query }}{{ . | first | value | humanize }}V{{end}}" - name: ipmi rules: - alert: ipmi fan alarm expr: ipmi_fan_speed_state > 0 for: 5m labels: alertgroup: "{{ $labels.instance }}" annotations: fan_speed_rpm: "{{ with printf \"ipmi_fan_speed_rpm{instance='%s',id='%s'}\" $labels.instance $labels.id | query }}{{ . | first | value | humanize }}rpm{{end}}" - alert: ipmi temperature alarm expr: ipmi_temperature_state > 0 for: 5m labels: alertgroup: "{{ $labels.instance }}" annotations: temperature_celsius: "{{ with printf \"ipmi_temperature_celsius{instance='%s',id='%s'}\" $labels.instance $labels.id | query }}{{ . | first | value | humanize }}C{{end}}" - alert: ipmi voltage alarm expr: ipmi_voltage_state > 0 for: 5m labels: 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}}" - name: mdadm rules: - alert: mdadm array inactive expr: node_md_state{state="inactive"} > 0 for: 0m labels: alertgroup: "{{ $labels.instance }}" annotations: required: "{{ with printf \"node_md_disks_required{instance='%s',device='%s'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" active: "{{ with printf \"node_md_disks{instance='%s',device='%s',state='active'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" failed: "{{ with printf \"node_md_disks{instance='%s',device='%s',state='failed'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" spare: "{{ with printf \"node_md_disks{instance='%s',device='%s',state='spare'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" - alert: mdadm disk failed expr: node_md_disks{state="failed"} > 0 for: 0m labels: alertgroup: "{{ $labels.instance }}" annotations: required: "{{ with printf \"node_md_disks_required{instance='%s',device='%s'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" active: "{{ with printf \"node_md_disks{instance='%s',device='%s',state='active'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" failed: "{{ with printf \"node_md_disks{instance='%s',device='%s',state='failed'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" spare: "{{ with printf \"node_md_disks{instance='%s',device='%s',state='spare'}\" $labels.instance $labels.device | query }}{{ . | first | value | humanize }} disks{{end}}" - name: memory rules: - alert: low memory expr: node_memory_MemAvailable_bytes / node_memory_MemTotal_bytes < 0.1 for: 5m labels: alertgroup: "{{ $labels.instance }}" annotations: memory_free: "{{ $value | humanizePercentage }}" - alert: memory pressure expr: rate(node_vmstat_pgmajfault[1m]) > 1000 for: 5m labels: alertgroup: "{{ $labels.instance }}" annotations: major_page_faults: "{{ $value }} faults/s" - alert: oom kill detected expr: increase(node_vmstat_oom_kill[1m]) > 0 for: 0m labels: alertgroup: "{{ $labels.instance }}" annotations: new_oom_kills: "{{ $value }}"