From 7f9208af58e5143a02541be9e6b9ce5fa504156c Mon Sep 17 00:00:00 2001
From: Christian Pointner <equinox@spreadspace.org>
Date: Sat, 16 Oct 2021 12:35:32 +0200
Subject: prometues: add some more smartmon alerts

---
 .../prometheus/server/defaults/main/rules_node.yml | 47 +++++++++++++++++++++-
 1 file changed, 46 insertions(+), 1 deletion(-)

(limited to 'roles/monitoring')

diff --git a/roles/monitoring/prometheus/server/defaults/main/rules_node.yml b/roles/monitoring/prometheus/server/defaults/main/rules_node.yml
index ffe616b7..af47e7f7 100644
--- a/roles/monitoring/prometheus/server/defaults/main/rules_node.yml
+++ b/roles/monitoring/prometheus/server/defaults/main/rules_node.yml
@@ -309,10 +309,55 @@ prometheus_server_rules_node:
       description: "S.M.A.R.T. reports errors for disk {{ '{{' }} $labels.device {{ '}}' }} on host {{ '{{' }} $labels.instance {{ '}}' }}.\n  VALUE = {{ '{{' }} $value {{ '}}' }}\n  LABELS = {{ '{{' }} $labels {{ '}}' }}"
 
   - alert: SmartmonAttributeBelowThreshold
-    expr: smartmon_attr_value < smartmon_attr_threshold
+    expr: smartmon_attr_value <= smartmon_attr_threshold
     for: 0m
     labels:
       severity: critical
     annotations:
       summary: Host disk S.M.A.R.T. attribute is below it's threshold (instance {{ '{{' }} $labels.instance {{ '}}' }})
       description: "S.M.A.R.T. attribute {{ '{{' }} $labels.name {{ '}}' }} for disk {{ '{{' }} $labels.device {{ '}}' }} on host {{ '{{' }} $labels.instance {{ '}}' }} has fallen below it's threshold.\n  VALUE = {{ '{{' }} $value {{ '}}' }}\n  LABELS = {{ '{{' }} $labels {{ '}}' }}"
+
+  - alert: SmartmonReallocatedSectorsCount
+    expr: smartmon_attr_raw_value{name="reallocated_sector_ct"} > 0
+    for: 0m
+    labels:
+      severity: warning
+    annotations:
+      summary: Host disk S.M.A.R.T. reports reallocated sectors (instance {{ '{{' }} $labels.instance {{ '}}' }})
+      description: "S.M.A.R.T. reports reallocated sectors for disk {{ '{{' }} $labels.device {{ '}}' }} on host {{ '{{' }} $labels.instance {{ '}}' }}, the drive might be failing.\n  VALUE = {{ '{{' }} $value {{ '}}' }}\n  LABELS = {{ '{{' }} $labels {{ '}}' }}"
+
+  - alert: SmartmonReallocationEventCount
+    expr: smartmon_attr_raw_value{name="reallocated_event_count"} > 0
+    for: 0m
+    labels:
+      severity: warning
+    annotations:
+      summary: Host disk S.M.A.R.T. reports reallocation events (instance {{ '{{' }} $labels.instance {{ '}}' }})
+      description: "S.M.A.R.T. reports reallocation events for disk {{ '{{' }} $labels.device {{ '}}' }} on host {{ '{{' }} $labels.instance {{ '}}' }}, the drive might be failing.\n  VALUE = {{ '{{' }} $value {{ '}}' }}\n  LABELS = {{ '{{' }} $labels {{ '}}' }}"
+
+  - alert: SmartmonCurrentPendingSectors
+    expr: smartmon_attr_raw_value{name="current_pending_sector"} > 0
+    for: 0m
+    labels:
+      severity: warning
+    annotations:
+      summary: Host disk S.M.A.R.T. reports current pending sectors (instance {{ '{{' }} $labels.instance {{ '}}' }})
+      description: "S.M.A.R.T. reports current pending sectors for disk {{ '{{' }} $labels.device {{ '}}' }} on host {{ '{{' }} $labels.instance {{ '}}' }}, the drive might be failing.\n  VALUE = {{ '{{' }} $value {{ '}}' }}\n  LABELS = {{ '{{' }} $labels {{ '}}' }}"
+
+  - alert: SmartmonOfflineUncorrectable
+    expr: smartmon_attr_raw_value{name="offline_uncorrectable"} > 0
+    for: 0m
+    labels:
+      severity: warning
+    annotations:
+      summary: Host disk S.M.A.R.T. reports offline uncorrectable errors (instance {{ '{{' }} $labels.instance {{ '}}' }})
+      description: "S.M.A.R.T. reports offline uncorrectable errors for disk {{ '{{' }} $labels.device {{ '}}' }} on host {{ '{{' }} $labels.instance {{ '}}' }}, the drive might be failing.\n  VALUE = {{ '{{' }} $value {{ '}}' }}\n  LABELS = {{ '{{' }} $labels {{ '}}' }}"
+
+  - alert: SmartmonSpinRetryCount
+    expr: smartmon_attr_raw_value{name="spin_retry_count"} > 0
+    for: 0m
+    labels:
+      severity: warning
+    annotations:
+      summary: Host disk S.M.A.R.T. reports spin-up retries (instance {{ '{{' }} $labels.instance {{ '}}' }})
+      description: "S.M.A.R.T. reports spin-up retries for disk {{ '{{' }} $labels.device {{ '}}' }} on host {{ '{{' }} $labels.instance {{ '}}' }}, the drive might be failing.\n  VALUE = {{ '{{' }} $value {{ '}}' }}\n  LABELS = {{ '{{' }} $labels {{ '}}' }}"
-- 
cgit v1.2.3