Class: Gitlab::StackProf

Inherits:
Object
  • Object
show all
Defined in:
lib/gitlab/stack_prof.rb

Constant Summary collapse

DEFAULT_FILE_PREFIX =
Dir.tmpdir
DEFAULT_TIMEOUT_SEC =
30
DEFAULT_MODE =
:cpu
DEFAULT_INTERVAL_US =

Sample interval as a frequency in microseconds (~99hz); appropriate for CPU profiles

10_100
DEFAULT_INTERVAL_EVENTS =

Sample interval in event occurrences (n = every nth event); appropriate for allocation profiles

100

Class Method Summary collapse

Class Method Details

.installObject

this is a workaround for sidekiq, which defines its own SIGUSR2 handler. by defering to the sidekiq startup event, we get to set up our own handler late enough. see also: github.com/mperham/sidekiq/pull/4653


21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
# File 'lib/gitlab/stack_prof.rb', line 21

def self.install
  require 'stackprof'
  require 'tmpdir'

  if Gitlab::Runtime.sidekiq?
    Sidekiq.configure_server do |config|
      config.on :startup do
        on_worker_start
      end
    end
  else
    Gitlab::Cluster::LifecycleEvents.on_worker_start do
      on_worker_start
    end
  end
end

.interval(mode) ⇒ Object


135
136
137
# File 'lib/gitlab/stack_prof.rb', line 135

def self.interval(mode)
  mode == :object ? DEFAULT_INTERVAL_EVENTS : DEFAULT_INTERVAL_US
end

.log_event(event, labels = {}) ⇒ Object


127
128
129
130
131
132
133
# File 'lib/gitlab/stack_prof.rb', line 127

def self.log_event(event, labels = {})
  Gitlab::AppJsonLogger.info({
    event: 'stackprof',
    message: event,
    pid: Process.pid
  }.merge(labels.compact))
end

.on_worker_startObject


38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
# File 'lib/gitlab/stack_prof.rb', line 38

def self.on_worker_start
  log_event('listening for SIGUSR2 signal')

  # create a pipe in order to propagate signal out of the signal handler
  # see also: https://cr.yp.to/docs/selfpipe.html
  read, write = IO.pipe

  # create a separate thread that polls for signals on the pipe.
  #
  # this way we do not execute in signal handler context, which
  # lifts restrictions and also serializes the calls in a thread-safe
  # manner.
  #
  # it's very similar to a goroutine and channel design.
  #
  # another nice benefit of this method is that we can timeout the
  # IO.select call, allowing the profile to automatically stop after
  # a given interval (by default 30 seconds), avoiding unbounded memory
  # growth from a profile that was started and never stopped.
  t = Thread.new do
    timeout_s = ENV['STACKPROF_TIMEOUT_S']&.to_i || DEFAULT_TIMEOUT_SEC
    current_timeout_s = nil
    loop do
      read.getbyte if IO.select([read], nil, nil, current_timeout_s)

      if ::StackProf.running?
        stackprof_file_prefix = ENV['STACKPROF_FILE_PREFIX'] || DEFAULT_FILE_PREFIX
        stackprof_out_file = "#{stackprof_file_prefix}/stackprof.#{Process.pid}.#{SecureRandom.hex(6)}.profile"

        log_event(
          'stopping profile',
          profile_filename: stackprof_out_file,
          profile_timeout_s: timeout_s
        )

        ::StackProf.stop
        ::StackProf.results(stackprof_out_file)
        current_timeout_s = nil
      else
        mode = ENV['STACKPROF_MODE']&.to_sym || DEFAULT_MODE
        stackprof_interval = ENV['STACKPROF_INTERVAL']&.to_i
        stackprof_interval ||= interval(mode)

        log_event(
          'starting profile',
          profile_mode: mode,
          profile_interval: stackprof_interval,
          profile_timeout: timeout_s
        )

        ::StackProf.start(
          mode: mode,
          raw: Gitlab::Utils.to_boolean(ENV['STACKPROF_RAW'] || 'true'),
          interval: stackprof_interval
        )
        current_timeout_s = timeout_s
      end
    end
  rescue StandardError => e
    log_event("stackprof failed: #{e}")
  end
  t.abort_on_exception = true

  # in the case of puma, this will override the existing SIGUSR2 signal handler
  # that can be used to trigger a restart.
  #
  # puma cluster has two types of restarts:
  # * SIGUSR1: phased restart
  # * SIGUSR2: restart
  #
  # phased restart is not supported in our configuration, because we use
  # preload_app. this means we will always perform a normal restart.
  # additionally, phased restart is not supported when sending a SIGUSR2
  # directly to a puma worker (as opposed to the master process).
  #
  # the result is that the behaviour of SIGUSR1 and SIGUSR2 is identical in
  # our configuration, and we can always use a SIGUSR1 to perform a restart.
  #
  # thus, it is acceptable for us to re-appropriate the SIGUSR2 signal, and
  # override the puma behaviour.
  #
  # see also:
  # * https://github.com/puma/puma/blob/master/docs/signals.md#puma-signals
  # * https://github.com/mperham/sidekiq/wiki/Signals
  Signal.trap('SIGUSR2') do
    write.write('.')
  end
end