sensu-plugins-mesos 0.1.1 → 1.0.0

data/bin/check-mesos-failed-tasks.rb

@@ -0,0 +1,146 @@
+#! /usr/bin/env ruby
+#
+#   check-mesos-failed-tasks
+#
+# DESCRIPTION:
+#   This plugin checks that there are less or same number of failed tasks than provided on a Mesos cluster
+#
+# OUTPUT:
+#   plain text
+#
+# PLATFORMS:
+#   Linux
+#
+# DEPENDENCIES:
+#   gem: sensu-plugin
+#   gem: rest-client
+#   gem: json
+#
+# USAGE:
+#   #YELLOW
+#
+# NOTES:
+#
+# LICENSE:
+#   Copyright 2016, Oskar Flores (oskar.flores@gmail.com)
+#   Released under the same terms as Sensu (the MIT license); see LICENSE
+#   for details.
+#
+
+require 'sensu-plugin/check/cli'
+require 'rest-client'
+require 'json'
+require 'daybreak'
+
+class MesosFailedTasksCheck < Sensu::Plugin::Check::CLI
+  check_name 'CheckMesosFailedTasks'
+  @metrics_name = 'master/tasks_failed'.freeze
+
+  class << self
+    attr_reader :metrics_name
+  end
+
+  option :server,
+         description: 'Mesos server',
+         short: '-s SERVER',
+         long: '--server SERVER',
+         default: 'localhost'
+
+  option :port,
+         description: 'port (default 5050)',
+         short: '-p PORT',
+         long: '--port PORT',
+         default: 5050,
+         required: false
+
+  option :uri,
+         description: 'Endpoint URI',
+         short: '-u URI',
+         long: '--uri URI',
+         default: '/metrics/snapshot'
+
+  option :timeout,
+         description: 'timeout in seconds',
+         short: '-t TIMEOUT',
+         long: '--timeout TIMEOUT',
+         proc: proc(&:to_i),
+         default: 5
+
+  option :value,
+         description: 'value to check against',
+         short: '-v VALUE',
+         long: '--value VALUE',
+         proc: proc(&:to_i),
+         default: 0,
+         required: false
+
+  option :delta,
+         short: '-d',
+         long: '--delta',
+         description: 'Use this flag to compare the metric with the previously retrieved value',
+         boolean: true
+
+  def run
+    if config[:value].to_i < 0
+      unknown 'Number of failed tasks cannot be negative'
+    end
+
+    server = config[:server]
+    port = config[:port]
+    uri = config[:uri]
+    timeout = config[:timeout].to_i
+    value = config[:value].to_i
+
+    begin
+      server = get_leader_url server, port
+      r = RestClient::Resource.new("#{server}#{uri}", timeout).get
+      tasks_failed = check_tasks(r)
+      if config[:delta]
+        db = Daybreak::DB.new '/tmp/mesos-metrics.db', default: 0
+        prev_value = db["task_#{MesosFailedTasksCheck.metrics_name}"]
+        db.lock do
+          db["task_#{MesosFailedTasksCheck.metrics_name}"] = tasks_failed
+        end
+        tasks_failed -= prev_value
+        db.flush
+        db.compact
+        db.close
+      end
+
+      if tasks_failed >= value
+        critical "The number of FAILED tasks [#{tasks_failed}] is bigger than provided [#{value}]!"
+      end
+    rescue Errno::ECONNREFUSED, RestClient::ResourceNotFound, SocketError
+      unknown "Mesos #{server} is not responding"
+    rescue RestClient::RequestTimeout
+      unknown "Mesos #{server} connection timed out"
+    end
+    ok
+  end
+
+  # Redirects server call to discover the Leader
+  # @param server [String] Server address
+  # @param port [Number] api port
+  # @return [Url] Url representing the Leader
+
+  def get_leader_url(server, port)
+    RestClient::Resource.new("http://#{server}:#{port}/redirect").get.request.url
+  end
+
+  # Parses JSON data as returned from Mesos's metrics API
+  # @param data [String] Server response
+  # @return [Integer] Number of failed tasks in Mesos
+  def check_tasks(data)
+    begin
+      tasks_failed = JSON.parse(data)[MesosFailedTasksCheck.metrics_name]
+    rescue JSON::ParserError
+      raise "Could not parse JSON response: #{data}"
+    end
+
+    if tasks_failed.nil?
+      raise "No metrics for [#{MesosFailedTasksCheck.metrics_name}] in server response: #{data}"
+    end
+
+    tasks_failed.round.to_i
+  end
+end

data/bin/check-mesos-gpu-balance.rb

@@ -0,0 +1,154 @@
+#! /usr/bin/env ruby
+#
+#   check-mesos-gpu-balance
+#
+# DESCRIPTION:
+#   This plugin checks that there is less GPU imbalance than specified on a certain mesos cluster
+#
+# OUTPUT:
+#   plain text
+#
+# PLATFORMS:
+#   Linux
+#
+# DEPENDENCIES:
+#   gem: sensu-plugin
+#   gem: rest-client
+#   gem: json
+#
+# USAGE:
+#   #YELLOW
+#
+# NOTES:
+#
+# LICENSE:
+#   Released under the same terms as Sensu (the MIT license); see LICENSE
+#   for details.
+#
+
+require 'sensu-plugin/check/cli'
+require 'rest-client'
+require 'json'
+
+class MesosGpuBalanceCheck < Sensu::Plugin::Check::CLI
+  check_name 'MesosGpuBalanceCheck'
+  @metrics_name = 'slaves'.freeze
+  CHECK_TYPE = 'gpus'.freeze
+
+  class << self
+    attr_reader :metrics_name
+  end
+
+  option :server,
+         description: 'Mesos server',
+         short: '-s SERVER',
+         long: '--server SERVER',
+         default: 'localhost'
+
+  option :port,
+         description: 'port (default 5050)',
+         short: '-p PORT',
+         long: '--port PORT',
+         default: 5050,
+         required: false
+
+  option :uri,
+         description: 'Endpoint URI',
+         short: '-u URI',
+         long: '--uri URI',
+         default: '/master/slaves'
+
+  option :timeout,
+         description: 'timeout in seconds',
+         short: '-t TIMEOUT',
+         long: '--timeout TIMEOUT',
+         proc: proc(&:to_i),
+         default: 5
+
+  option :crit,
+         description: 'Critical value to check against',
+         short: '-c VALUE',
+         long: '--critical VALUE',
+         proc: proc(&:to_i),
+         default: 0,
+         required: false
+
+  option :warn,
+         description: 'Warning value to check against',
+         short: '-w VALUE',
+         long: '--warning VALUE',
+         proc: proc(&:to_i),
+         default: 0,
+         required: false
+
+  def run
+    if config[:crit] < 0 || config[:warn] < 0
+      unknown "Thresholds cannot be negative, crit: #{config[:crit]}, warn: #{config[:warn]}"
+    end
+
+    server = config[:server]
+    port = config[:port]
+    uri = config[:uri]
+    timeout = config[:timeout]
+    crit = config[:crit]
+    warn = config[:warn]
+
+    begin
+      server = get_leader_url server, port
+      r = RestClient::Resource.new("#{server}#{uri}", timeout).get
+      compare = get_check_diff(get_slaves(r))
+      if compare['diff'] >= crit
+        critical "There is a GPU usage diff of #{compare['diff']} bigger than #{crit} " + compare['msg']
+      end
+      if compare['diff'] >= warn
+        warning "There is a GPU usage diff of #{compare['diff']} bigger than #{warn} " + compare['msg']
+      end
+    rescue Errno::ECONNREFUSED, RestClient::ResourceNotFound, SocketError
+      unknown  "Mesos #{server} is not responding"
+    rescue RestClient::RequestTimeout
+      unknown  "Mesos #{server} connection timed out"
+    end
+    ok
+  end
+
+  # Redirects server call to discover the Leader
+  # @param server [String] Server address
+  # @param port [Number] api port
+  # @return [Url] Url representing the Leader
+
+  def get_leader_url(server, port)
+    RestClient::Resource.new("http://#{server}:#{port}/redirect").get.request.url
+  end
+
+  # Parses JSON data as returned from Mesos's metrics API
+  # @param data [String] Server response
+  # @return [Integer] Number of failed tasks in Mesos
+  def get_slaves(data)
+    begin
+      slaves = JSON.parse(data)[MesosGpuBalanceCheck.metrics_name]
+    rescue JSON::ParserError
+      raise "Could not parse JSON response: #{data}"
+    end
+
+    if slaves.nil?
+      raise "No metrics for [#{MesosGpuBalanceCheck.metrics_name}] in server response: #{data}"
+    end
+
+    slaves
+  end
+
+  def get_check_diff(slavelist)
+    begin
+      usages = {}
+      check_diff = {}
+      slavelist.each do |slaveinfo|
+        usages.store(slaveinfo['hostname'], slaveinfo['used_resources'][CHECK_TYPE] * 100 / slaveinfo['resources'][CHECK_TYPE])
+      end
+      sorted = usages.sort_by { |_hostname, total| total }
+      max = usages.length - 1
+      check_diff['diff'] = sorted[max][1] - sorted[0][1]
+      check_diff['msg'] = "Hostname #{sorted[0][0]} uses #{sorted[0][1]}% and Hostname #{sorted[max][0]} uses #{sorted[max][1]}%"
+    end
+    check_diff
+  end
+end

data/bin/check-mesos-leader-status.rb

@@ -0,0 +1,75 @@
+#! /usr/bin/env ruby
+#
+#   check-mesos-leader-status
+#
+# DESCRIPTION:
+#   This plugin checks that the health url of the leader master returns 200 OK
+#
+# OUTPUT:
+#   plain text
+#
+# PLATFORMS:
+#   Linux
+#
+# DEPENDENCIES:
+#   gem: sensu-plugin
+#   gem: rest-client
+#
+# USAGE:
+#   #YELLOW
+#
+# NOTES:
+#
+# LICENSE:
+#   Copyright 2016, Oskar Flores (oskar.flores@gmail.com)
+#   Released under the same terms as Sensu (the MIT license); see LICENSE
+#   for details.
+#
+
+require 'sensu-plugin/check/cli'
+require 'rest-client'
+
+class MesosLeaderNodeStatus < Sensu::Plugin::Check::CLI
+  option :server,
+         description: 'Mesos server',
+         short: '-s SERVER',
+         long: '--server SERVER',
+         default: 'localhost'
+
+  option :port,
+         description: 'port (default 5050)',
+         short: '-p PORT',
+         long: '--port PORT',
+         default: 5050,
+         required: false
+
+  option :uri,
+         description: 'Endpoint URI',
+         short: '-u URI',
+         long: '--uri URI',
+         default: '/redirect'
+
+  option :timeout,
+         description: 'timeout in seconds',
+         short: '-t TIMEOUT',
+         long: '--timeout TIMEOUT',
+         proc: proc(&:to_i),
+         default: 5
+
+  def run
+    server = config[:server]
+    port = config[:port]
+    uri = config[:uri]
+    begin
+      r = RestClient::Resource.new("http://#{server}:#{port}#{uri}", timeout: config[:timeout]).get
+      if r.code == 503
+        critical "Master on #{server} is not responding"
+      end
+    rescue Errno::ECONNREFUSED, RestClient::ResourceNotFound, SocketError
+      critical "Mesos on #{server} is not responding"
+    rescue RestClient::RequestTimeout
+      critical "Mesos on #{server} connection timed out"
+    end
+    ok
+  end
+end

data/bin/check-mesos-lost-tasks.rb

@@ -0,0 +1,139 @@
+#! /usr/bin/env ruby
+#
+#   check-mesos-lost-tasks
+#
+# DESCRIPTION:
+#   This plugin checks that there are less or same number of lost tasks than provided on a Mesos cluster
+#
+# OUTPUT:
+#   plain text
+#
+# PLATFORMS:
+#   Linux
+#
+# DEPENDENCIES:
+#   gem: sensu-plugin
+#   gem: rest-client
+#   gem: json
+#
+# USAGE:
+#   #YELLOW
+#
+# NOTES:
+#
+# LICENSE:
+#   Copyright 2016, Oskar Flores (oskar.flores@gmail.com)
+#   Released under the same terms as Sensu (the MIT license); see LICENSE
+#   for details.
+#
+
+require 'sensu-plugin/check/cli'
+require 'rest-client'
+require 'json'
+require 'daybreak'
+
+class MesosLostTasksCheck < Sensu::Plugin::Check::CLI
+  check_name 'CheckMesosLostTasks'
+  @metrics_name = 'master/tasks_lost'.freeze
+
+  class << self
+    attr_reader :metrics_name
+  end
+
+  option :server,
+         description: 'Mesos server',
+         short: '-s SERVER',
+         long: '--server SERVER',
+         default: 'localhost'
+
+  option :port,
+         description: 'port (default 5050)',
+         short: '-p PORT',
+         long: '--port PORT',
+         default: 5050,
+         required: false
+
+  option :timeout,
+         description: 'timeout in seconds',
+         short: '-t TIMEOUT',
+         long: '--timeout TIMEOUT',
+         proc: proc(&:to_i),
+         default: 5
+
+  option :uri,
+         description: 'Endpoint URI',
+         short: '-u URI',
+         long: '--uri URI',
+         default: '/metrics/snapshot'
+
+  option :value,
+         description: 'value to check against',
+         short: '-v VALUE',
+         long: '--value VALUE',
+         default: 0,
+         proc: proc(&:to_i),
+         required: false
+
+  option :delta,
+         short: '-d',
+         long: '--delta',
+         description: 'Use this flag to compare the metric with the previously retreived value',
+         boolean: true
+
+  def run
+    if config[:value] < 0
+      unknown 'Number of lost tasks cannot be negative, please set --value to a number greater or equal to 0'
+    end
+
+    server = config[:server]
+    port = config[:port]
+    uri = config[:uri]
+    timeout = config[:timeout]
+    value = config[:value]
+
+    begin
+      server = get_leader_url server, port
+      # remove comment for debugging purpose
+      # puts(server)
+
+      r = RestClient::Resource.new("#{server}#{uri}", timeout).get
+      tasks_lost = check_tasks(r)
+      if config[:delta]
+        db = Daybreak::DB.new '/tmp/mesos-metrics.db', default: 0
+        prev_value = db["task_#{MesosLostTasksCheck.metrics_name}"]
+        db.lock do
+          db["task_#{MesosLostTasksCheck.metrics_name}"] = tasks_lost
+        end
+        tasks_lost -= prev_value
+        db.flush
+        db.compact
+        db.close
+      end
+      if tasks_lost >= value
+        critical "The number of LOST tasks [#{tasks_lost}] is bigger than provided [#{value}]!"
+      end
+    end
+    ok
+  end
+
+  def get_leader_url(server, port)
+    RestClient::Resource.new("http://#{server}:#{port}/redirect").get.request.url
+  end
+
+  # Parses JSON data as returned from Mesos's metrics API
+  # @param data [String] Server response
+  # @return [Integer] Number of lost tasks in Mesos
+  def check_tasks(data)
+    begin
+      tasks_lost = JSON.parse(data)[MesosLostTasksCheck.metrics_name]
+    rescue JSON::ParserError
+      raise "Could not parse JSON response: #{data}"
+    end
+
+    if tasks_lost.nil?
+      raise "No metrics for [#{MesosLostTasksCheck.metrics_name}] in server response: #{data}"
+    end
+
+    tasks_lost.round.to_i
+  end
+end