#!/usr/bin/env ruby

# Forwards information on a Riak node to Riemann.

require File.expand_path('../../lib/riemann/tools', __FILE__)

require 'net/http'
require 'yajl/json_gem'

class Riemann::Tools::Riak
  include Riemann::Tools

  opt :riak_host, "Riak host", :default => Socket.gethostname
  opt :data_dir, "Riak data directory", :default => '/var/lib/riak'
  opt :stats_port, "Riak HTTP port for stats", :default => 8098
  opt :stats_path, "Riak HTTP stats path", :default => '/stats'
  opt :node_name, "Riak erlang node name", :default => "riak@#{Socket.gethostname}"
  
  opt :get_50_warning, "FSM 50% get time warning threshold (ms)", :default => 1000
  opt :put_50_warning, "FSM 50% put time warning threshold (ms)", :default => 1000
  opt :get_95_warning, "FSM 95% get time warning threshold (ms)", :default => 2000
  opt :put_95_warning, "FSM 95% put time warning threshold (ms)", :default => 2000
  opt :get_99_warning, "FSM 99% get time warning threshold (ms)", :default => 10000
  opt :put_99_warning, "FSM 99% put time warning threshold (ms)", :default => 10000

  def initialize
    if `which escript` =~ /^\s*$/
      puts "No escript; disabling ring/key checks."
      @no_escript = true
    end
  end

  def check_ring
    return if @no_escript

    str = `#{File.expand_path(File.dirname(__FILE__))}/riemann-riak-ring #{opts[:node_name]}`.chomp
    if str =~ /^TRUE/
      report(
        :host => opts[:riak_host],
        :service => 'riak ring',
        :state => 'ok',
        :description => str
      )
    else
      report(
        :host => opts[:riak_host],
        :service => 'riak ring',
        :state => 'warning',
        :description => str
      )
    end
  end

  def check_keys
    return if @no_escript
    
    keys = `#{File.expand_path(File.dirname(__FILE__))}/riemann-riak-keys #{opts[:node_name]}`.chomp
    if keys =~ /^\d+$/
      report(
        :host => opts[:riak_host],
        :service => 'riak keys',
        :state => 'ok',
        :metric => keys.to_i,
        :description => keys
      )
    else
      report(
        :host => opts[:riak_host],
        :servie => 'riak keys',
        :state => 'unknown',
        :description => keys
      )
    end
  end

  def check_disk
    gb = `du -s #{opts[:data_dir]}`.split(/\s+/).first.to_i / (1024.0**2)
    report(
      :host => opts[:riak_host],
      :service => 'riak disk',
      :state => 'ok',
      :metric => gb,
      :description => "#{gb} GB in /var/lib/riak/"
    )
  end

  # Returns the riak stat for the given fsm type and percentile.
  def fsm_stat(type, percentile)
    "node_#{type}_fsm_time_#{percentile == 50 ? 'median' : percentile}"
  end

  # Returns the alerts state for the given fsm.
  def fsm_state(type, percentile, val)
    limit = opts["#{type}_#{percentile}_warning".to_sym]
    case val
    when 0 .. limit
      'ok'
    when limit .. limit * 2
      'warning'
    else
      'critical'
    end
  end

  def check_stats
    begin
      res = Net::HTTP.start(opts[:riak_host], opts[:stats_port]) do |http|
        http.get opts[:stats_path]
      end
    rescue => e
      report(
        :host => opts[:riak_host],
        :service => 'riak',
        :state => 'critical',
        :description => "error fetching #{opts[:riak_host]}:#{opts[:stats_port]} #{e.class}, #{e.message}"
      )
      return
    end

    if res.code.to_i == 200
      stats = JSON.parse(res.body)
    else
      report(
        :host => opts[:riak_host],
        :service => 'riak',
        :state => 'critical',
        :description => "stats returned HTTP #{res.code}:\n\n#{res.body}"
      )
      return
    end

    report(
      :host => opts[:riak_host],
      :service => 'riak',
      :state => 'ok'
    )

    # Gets/puts/rr
    [
      'vnode_gets', 
      'vnode_puts', 
      'node_gets', 
      'node_puts', 
      'read_repairs' 
    ].each do |s|
      report(
        :host => opts[:riak_host],
        :service => "riak #{s}",
        :state => 'ok',
        :metric => stats[s]/60.0,
        :description => "#{stats[s]/60.0}/sec"
      )
    end

    # FSMs
    ['get', 'put'].each do |type|
      [50, 95, 99].each do |percentile|
        val = stats[fsm_stat(type, percentile)] || 0
        val = 0 if val == 'undefined'
        val /= 1000.0 # Convert us to ms
        state = fsm_state(type, percentile, val)
        report(
          :host => opts[:riak_host],
          :service => "riak #{type} #{percentile}",
          :state => state,
          :metric => val,
          :description => "#{val} ms"
        )
      end
    end
  end

  def tick
    # This can utterly destroy a cluster, so we disable
    # check_keys
    check_stats
    check_ring
    check_disk
  end
end

Riemann::Tools::Riak.run