From 9288b25cac9ef24252b314adcbe72efdb4ba7ee2 Mon Sep 17 00:00:00 2001
From: iminfinity <suwarna@obmondo.com>
Date: Thu, 14 Nov 2024 16:53:26 +0530
Subject: [PATCH] add smartmon_sata rule to handle sata cable failures

---
 .../rules/smartmon_sata.yaml                  | 18 ++++++++++
 .../prometheusRule-smartmon-sata.yaml         |  9 +++++
 .../tests/smartmon_sata.yaml                  | 35 +++++++++++++++++++
 .../prometheus-linuxaid/values.yaml           |  1 +
 4 files changed, 63 insertions(+)
 create mode 100644 argocd-helm-charts/prometheus-linuxaid/rules/smartmon_sata.yaml
 create mode 100644 argocd-helm-charts/prometheus-linuxaid/templates/prometheusRule-smartmon-sata.yaml
 create mode 100644 argocd-helm-charts/prometheus-linuxaid/tests/smartmon_sata.yaml

diff --git a/argocd-helm-charts/prometheus-linuxaid/rules/smartmon_sata.yaml b/argocd-helm-charts/prometheus-linuxaid/rules/smartmon_sata.yaml
new file mode 100644
index 00000000..bd83f863
--- /dev/null
+++ b/argocd-helm-charts/prometheus-linuxaid/rules/smartmon_sata.yaml
@@ -0,0 +1,18 @@
+groups:
+  - name: monitor::system::service::smartmon
+    rules:
+    - alert: monitor::system::service::smartmon::smartmon_udma_crc_error_count_raw_value
+      expr: |
+        smartmon_udma_crc_error_count_raw_value != 0 and on(certname) obmondo_monitoring
+      for: 3h
+      labels:
+        severity: critical
+        alert_id: monitor::system::service::smartmon::smartmon_udma_crc_error_count_raw_value
+      annotations:
+        summary: 'Disk **{{ $labels.disk }}** on **{{ $labels.certname }}** has disk sata failure'
+        description: |
+          Disk **{{ $labels.disk }}** on **{{ $labels.certname }}** has disk sata failure,
+          instance="**{{ $labels.instance }}**",
+          type="**{{ $labels.type }}**",
+
+          UDMA_CRC_Error_Count - The number of errors related to data transfer over the interface. A value of **{{ $value }}** is concerning and indicates potential issues with the data cable or connections.
diff --git a/argocd-helm-charts/prometheus-linuxaid/templates/prometheusRule-smartmon-sata.yaml b/argocd-helm-charts/prometheus-linuxaid/templates/prometheusRule-smartmon-sata.yaml
new file mode 100644
index 00000000..4a1e6acf
--- /dev/null
+++ b/argocd-helm-charts/prometheus-linuxaid/templates/prometheusRule-smartmon-sata.yaml
@@ -0,0 +1,9 @@
+{{- if .Values.prometheusRule.smartmon_sata }}
+apiVersion: monitoring.coreos.com/v1
+kind: PrometheusRule
+metadata:
+  name: smartmon-sata-rules
+  namespace: monitoring-{{ .Values.customerid }}
+spec:
+  {{- $.Files.Get "rules/smartmon_sata.yaml" | nindent 4 }}
+{{- end }}
diff --git a/argocd-helm-charts/prometheus-linuxaid/tests/smartmon_sata.yaml b/argocd-helm-charts/prometheus-linuxaid/tests/smartmon_sata.yaml
new file mode 100644
index 00000000..78279177
--- /dev/null
+++ b/argocd-helm-charts/prometheus-linuxaid/tests/smartmon_sata.yaml
@@ -0,0 +1,35 @@
+---
+evaluation_interval: 1m
+
+rule_files:
+  - ../rules/smartmon_sata.yaml
+
+tests:
+  - interval: 1m
+    input_series:
+      - series: obmondo_monitoring{certname="jacen.enableit", alert_id="monitor::system::service::smartmon::smartmon_udma_crc_error_count_raw_value"}
+        values: 1x200
+      - series: smartmon_udma_crc_error_count_raw_value{certname="jacen.enableit", disk="/dev/sdb", type="sat", instance="htzhel1-ax41a.enableit:63385"}
+        values: 1x200
+      - series: smartmon_udma_crc_error_count_raw_value{certname="jacen.enableit", disk="/dev/sba", type="sat", instance="htzhel1-ax41a.enableit:63385"}
+        values: 0x200
+
+    alert_rule_test:
+      - alertname: 'monitor::system::service::smartmon::smartmon_udma_crc_error_count_raw_value'
+        eval_time: 3h
+        exp_alerts:
+          - exp_labels:
+              severity: 'critical'
+              certname: 'jacen.enableit'
+              disk: '/dev/sdb'
+              type: sat
+              instance: "htzhel1-ax41a.enableit:63385"
+              alert_id: 'monitor::system::service::smartmon::smartmon_udma_crc_error_count_raw_value'
+            exp_annotations:
+              summary: 'Disk **/dev/sdb** on **jacen.enableit** has disk sata failure'
+              description: |
+                Disk **/dev/sdb** on **jacen.enableit** has disk sata failure,
+                instance="**htzhel1-ax41a.enableit:63385**",
+                type="**sat**",
+
+                UDMA_CRC_Error_Count - The number of errors related to data transfer over the interface. A value of **1** is concerning and indicates potential issues with the data cable or connections.
diff --git a/argocd-helm-charts/prometheus-linuxaid/values.yaml b/argocd-helm-charts/prometheus-linuxaid/values.yaml
index 8aca4848..1dd55b4e 100644
--- a/argocd-helm-charts/prometheus-linuxaid/values.yaml
+++ b/argocd-helm-charts/prometheus-linuxaid/values.yaml
@@ -54,6 +54,7 @@ prometheusRule:
   smart: true
   ssacli: true
   zfsExporter: true
+  smartmon_sata: true
 alertmanager:
   version: v0.27.0
   config: