linux/meta/heka.yml

{%- from "linux/map.jinja" import monitoring with context %}
metric_collector:
  trigger:
    linux_system_cpu_critical:
      description: 'The CPU usage is too high.'
      severity: critical
      rules:
      - metric: cpu_wait
        relational_operator: '>='
        threshold: 35
        window: 120
        periods: 0
        function: avg
      - metric: cpu_idle
        relational_operator: <=
        threshold: 5
        window: 120
        function: avg
    linux_system_cpu_warning:
      description: 'The CPU wait times are high.'
      severity: warning
      rules:
      - metric: cpu_wait
        relational_operator: '>='
        threshold: 15
        window: 120
        periods: 0
        function: avg
    linux_system_swap_usage_critical:
      description: 'There is no more swap free space'
      severity: critical
      rules:
      - metric: swap_free
        relational_operator: '=='
        threshold: 0
        window: 60
        periods: 0
        function: max
    linux_system_swap_activity_warning:
      description: 'The swap activity is high'
      severity: warning
      rules:
      - metric: swap_io_in
        relational_operator: '>='
        threshold: 1048576 # 1 Mb/s
        window: 120
        periods: 0
        function: avg
      - metric: swap_io_out
        relational_operator: '>='
        threshold: 1048576 # 1 Mb/s
        window: 120
        periods: 0
        function: avg
    linux_system_swap_usage_warning:
      description: 'The swap free space is low'
      severity: warning
      rules:
      - metric: swap_percent_used
        relational_operator: '>='
        threshold: 0.8
        window: 60
        periods: 0
        function: avg
    linux_system_root_fs_warning:
      description: "The root filesystem's free space is low"
      severity: warning
      rules:
      - metric: fs_space_percent_free
        field:
          fs: '/'
        relational_operator: '<'
        threshold: 10
        window: 60
        periods: 0
        function: min
    linux_system_root_fs_critical:
      description: "The root filesystem's free space is too low"
      severity: critical
      rules:
      - metric: fs_space_percent_free
        field:
          fs: '/'
        relational_operator: '<'
        threshold: 5
        window: 60
        periods: 0
        function: min
    linux_system_network_warning_dropped_rx:
      description: 'Some received packets have been dropped'
      severity: warning
      rules:
      - metric: if_dropped_rx
        relational_operator: '>'
        threshold: 100
        window: 60
        periods: 0
        function: avg
    linux_system_network_critical_dropped_rx:
      description: 'Too many received packets have been dropped'
      severity: critical
      rules:
      - metric: if_dropped_rx
        relational_operator: '>'
        threshold: 1000
        window: 60
        periods: 0
        function: avg
    linux_system_network_warning_dropped_tx:
      description: 'Some transmitted packets have been dropped'
      severity: warning
      rules:
      - metric: if_dropped_tx
        relational_operator: '>'
        threshold: 100
        window: 60
        periods: 0
        function: avg
    linux_system_network_critical_dropped_tx:
      description: 'Too many transmitted packets have been dropped'
      severity: critical
      rules:
      - metric: if_dropped_tx
        relational_operator: '>'
        threshold: 1000
        function: avg
        window: 60
    linux_system_hdd_errors_critical:
      description: 'Errors on hard drive(s) have been detected'
      severity: critical
      no_data_policy: okay
      rules:
      - metric: hdd_errors_rate
        group_by: [device]
        relational_operator: '>'
        threshold: 0
        window: 60
        periods: 0
        function: max
    {%- if monitoring.bond_status.interfaces is defined and monitoring.bond_status.interfaces is list %}
    linux_bond_status_critical:
      description: Bond members are down.
      rules:
      - function: last
        metric: bond_status_links_down
        periods: 0
        relational_operator: '>'
        threshold: 0
        window: 120
      severity: critical
    {%- endif %}
  alarm:
    linux_system_cpu:
      alerting: enabled
      triggers:
      - linux_system_cpu_warning
      - linux_system_cpu_critical
    linux_system_swap:
      alerting: enabled
      triggers:
      - linux_system_swap_usage_critical
      - linux_system_swap_activity_warning
      - linux_system_swap_usage_warning
    linux_system_root_fs:
      alerting: enabled
      triggers:
      - linux_system_root_fs_critical
      - linux_system_root_fs_warning
    linux_system_network_rx:
      alerting: enabled
      triggers:
      - linux_system_network_critical_dropped_rx
      - linux_system_network_warning_dropped_rx
    linux_system_network_tx:
      alerting: enabled
      triggers:
      - linux_system_network_critical_dropped_tx
      - linux_system_network_warning_dropped_tx
    linux_system_hdd_errors:
      alerting: enabled_with_notification
      triggers:
      - linux_system_hdd_errors_critical
    {%- if monitoring.bond_status.interfaces is defined and monitoring.bond_status.interfaces is list %}
    linux_bond_status:
      alerting: enabled
      triggers:
        - linux_bond_status_critical
    {%- endif %}