MON-96 update system readme and add load monitor

system/generic/README.md

@@ -17,7 +17,11 @@ Purpose
 -------
 Creates a DataDog monitors with the following checks :
 
-* System CPU High
+* CPU usage
+* CPU load ratio
+* Free memory
+* Free disk inodes
+* Free disk space
 
 Inputs
 ------
@@ -29,6 +33,11 @@ Inputs
 | cpu_high_threshold_critical | CPU high critical threshold | string | `95` | no |
 | cpu_high_threshold_warning | CPU high warning threshold | string | `80` | no |
 | cpu_high_timeframe | CPU high timeframe | string | `last_5m` | no |
+| cpu_load_message | Custom message for CPU load ratio monitor | string | `` | no |
+| cpu_load_silenced | Groups to mute for CPU load ratio monitor | map | `<map>` | no |
+| cpu_load_threshold_critical | CPU load ratio critical threshold | string | `4` | no |
+| cpu_load_threshold_warning | CPU load ratio warning threshold | string | `3` | no |
+| cpu_load_timeframe | CPU load timeframe | string | `last_5m` | no |
 | environment | Architecture Environment | string | - | yes |
 | delay | Delay in seconds for the metric evaluation | string | `15` | no |
 | filter_tags_custom | Tags used for custom filtering when filter_tags_use_defaults is false | string | `*` | no |

system/generic/inputs.tf

@@ -53,6 +53,33 @@ variable "cpu_high_threshold_critical" {
   default     = 95
 }
 
+variable "cpu_load_silenced" {
+  description = "Groups to mute for CPU load ratio monitor"
+  type        = "map"
+  default     = {}
+}
+
+variable "cpu_load_message" {
+  description = "Custom message for CPU load ratio monitor"
+  type        = "string"
+  default     = ""
+}
+
+variable "cpu_load_timeframe" {
+  description = "CPU load ratio timeframe"
+  default     = "last_5m"
+}
+
+variable "cpu_load_threshold_warning" {
+  description = "CPU load ratio warning threshold"
+  default     = 3
+}
+
+variable "cpu_load_threshold_critical" {
+  description = "CPU load ratio critical threshold"
+  default     = 4
+}
+
 variable "free_disk_space_silenced" {
   description = "Groups to mute for Free diskspace monitor"
   type        = "map"

system/generic/monitors-system.tf

@@ -38,6 +38,39 @@ resource "datadog_monitor" "datadog_cpu_too_high" {
   silenced = "${var.cpu_high_silenced}"
 }
 
+resource "datadog_monitor" "datadog_load_too_high" {
+  name    = "[${var.environment}] CPU load 5 {{#is_alert}}{{comparator}} {{threshold}}% ({{value}}%){{/is_alert}}{{#is_warning}}{{comparator}} {{warn_threshold}}% ({{value}}%){{/is_warning}}"
+  message = "${coalesce(var.cpu_load_message, var.message)}"
+
+  query = <<EOF
+    min(${var.cpu_load_timeframe}): (
+      avg:system.load.5{${data.template_file.filter.rendered}} by {region,host} /
+      avg:system.core.count{${data.template_file.filter.rendered}} by {region,host}
+    ) > ${var.cpu_load_threshold_critical}
+  EOF
+
+  type = "metric alert"
+
+  thresholds {
+    warning  = "${var.cpu_load_threshold_warning}"
+    critical = "${var.cpu_load_threshold_critical}"
+  }
+
+  tags = ["env:${var.environment}", "type:system", "resource:load"]
+
+  notify_no_data      = true
+  evaluation_delay    = "${var.delay}"
+  new_host_delay      = "${var.delay}"
+  notify_audit        = false
+  timeout_h           = 0
+  include_tags        = true
+  locked              = false
+  require_full_window = true
+  no_data_timeframe   = 20
+
+  silenced = "${var.cpu_load_silenced}"
+}
+
 resource "datadog_monitor" "datadog_free_disk_space_too_low" {
   name    = "[${var.environment}] Free disk space {{#is_alert}}{{comparator}} {{threshold}}% ({{value}}%){{/is_alert}}{{#is_warning}}{{comparator}} {{warn_threshold}}% ({{value}}%){{/is_warning}}"
   message = "${coalesce(var.free_disk_space_message, var.message)}"