Module: DispatchPolicy::InflightTracker

Extended by:
ActiveSupport::Concern
Defined in:
lib/dispatch_policy/inflight_tracker.rb

Overview

Around-perform that records each job execution in dispatch_policy_inflight_jobs while it runs, so the concurrency gate can count active jobs per partition.

While the job runs we spawn a heartbeat thread that bumps ‘heartbeat_at` every `config.inflight_heartbeat_interval` seconds. Without this, jobs longer than `inflight_stale_after` (default 5 min) get their inflight row prematurely swept and the concurrency gate over-admits.

Defined Under Namespace

Classes: Heartbeat, ThreadSafeFlag

Constant Summary collapse

HEARTBEAT_KEY =

—– heartbeat thread —–

:__dispatch_policy_heartbeat_token__

Class Method Summary collapse

Class Method Details

.lookup_admitted_at(active_job_id) ⇒ Object

Reads the admitted_at column from the inflight row that the Tick pre-inserted. Used as the start-of-queue-wait reference for the adaptive_concurrency feedback signal (queue_lag = perform_start

  • admitted_at). nil if the row vanished or the lookup fails —

the observation is then skipped.



75
76
77
78
79
80
81
82
83
84
85
86
87
 
# File 'lib/dispatch_policy/inflight_tracker.rb', line 75

def self.lookup_admitted_at(active_job_id)
  result = ActiveRecord::Base.connection.exec_query(
    "SELECT admitted_at FROM dispatch_policy_inflight_jobs WHERE active_job_id = $1 LIMIT 1",
    "lookup_admitted_at",
    [active_job_id]
  )
  row = result.first
  return nil unless row
  ts = row["admitted_at"]
  ts.is_a?(Time) ? ts : Time.parse(ts.to_s)
rescue StandardError
  nil
end

.record_adaptive_observations(policy:, gates:, partition_key:, admitted_at:, perform_start:, succeeded:) ⇒ Object 



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
 
# File 'lib/dispatch_policy/inflight_tracker.rb', line 89

def self.record_adaptive_observations(policy:, gates:, partition_key:, admitted_at:, perform_start:, succeeded:)
  return if gates.empty?

  queue_lag_ms = if admitted_at
    ((perform_start - admitted_at) * 1000).to_i
  else
    # No admitted_at means we can't measure queue wait. Treat as 0
    # so the observation still increments sample_count and the
    # cap can grow if everything else is healthy.
    0
  end

  gates.each do |gate|
    gate.record_observation(
      policy_name:   policy.name,
      partition_key: partition_key,
      queue_lag_ms:  queue_lag_ms,
      succeeded:     succeeded
    )
  rescue StandardError => e
    DispatchPolicy.config.logger&.warn(
      "[dispatch_policy] adaptive observation failed for #{policy.name}/#{partition_key}: #{e.class}: #{e.message}"
    )
  end
end

.start_heartbeat(active_job_id) ⇒ Object 



121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
 
# File 'lib/dispatch_policy/inflight_tracker.rb', line 121

def self.start_heartbeat(active_job_id)
  interval = DispatchPolicy.config.inflight_heartbeat_interval.to_f
  return nil if interval <= 0

  stop_flag = Concurrent::AtomicBoolean.new(false) if defined?(Concurrent::AtomicBoolean)
  stop_flag ||= ThreadSafeFlag.new

  thread = Thread.new do
    Thread.current.name = "dispatch_policy.heartbeat:#{active_job_id}"

    until stop_flag.true?
      # Sleep in small slices so stop is responsive without polling tight.
      slept = 0.0
      slice = [interval, 1.0].min
      while slept < interval && !stop_flag.true?
        sleep(slice)
        slept += slice
      end
      break if stop_flag.true?

      begin
        ActiveRecord::Base.connection_pool.with_connection do
          Repository.heartbeat_inflight!(active_job_id: active_job_id)
        end
      rescue StandardError => e
        DispatchPolicy.config.logger&.warn("[dispatch_policy] heartbeat #{active_job_id} failed: #{e.class}: #{e.message}")
      end
    end
  end

  Heartbeat.new(thread, stop_flag)
end

.stop_heartbeat(heartbeat) ⇒ Object 



154
155
156
157
158
159
160
161
162
163
164
 
# File 'lib/dispatch_policy/inflight_tracker.rb', line 154

def self.stop_heartbeat(heartbeat)
  return if heartbeat.nil?

  heartbeat.stop_flag.make_true
  # Wake the thread out of any in-progress sleep so we don't wait the full slice.
  heartbeat.thread.wakeup if heartbeat.thread.alive?
  heartbeat.thread.join(1.0)
rescue StandardError
  # Worst case: the thread is killed by GC; the inflight row gets a stale
  # heartbeat_at and the sweeper will reclaim it after inflight_stale_after.
end

.track(job) ⇒ Object 



24
25
26
27
28
29
30
31
32
33
34
35
36
37
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
 
# File 'lib/dispatch_policy/inflight_tracker.rb', line 24

def self.track(job)
  policy_name = job.class.respond_to?(:dispatch_policy_name) && job.class.dispatch_policy_name
  return yield unless policy_name

  policy = DispatchPolicy.registry.fetch(policy_name)
  return yield unless policy

  ctx           = policy.build_context(job.arguments, queue_name: job.queue_name&.to_s)
  partition_key = policy.partition_key_for(ctx)

  Repository.insert_inflight!([{
    policy_name:    policy.name,
    partition_key:  partition_key,
    active_job_id:  job.job_id
  }])

  adaptive_gates = policy.gates.select { |g| g.name == :adaptive_concurrency }
  admitted_at    = adaptive_gates.any? ? lookup_admitted_at(job.job_id) : nil
  perform_start  = Time.current

  heartbeat = start_heartbeat(job.job_id)

  succeeded = false
  begin
    yield
    succeeded = true
  ensure
    stop_heartbeat(heartbeat)

    record_adaptive_observations(
      policy:        policy,
      gates:         adaptive_gates,
      partition_key: partition_key,
      admitted_at:   admitted_at,
      perform_start: perform_start,
      succeeded:     succeeded
    )

    begin
      Repository.delete_inflight!(active_job_id: job.job_id)
    rescue StandardError => e
      DispatchPolicy.config.logger&.warn("[dispatch_policy] failed to delete inflight row #{job.job_id}: #{e.class}: #{e.message}")
    end
  end
end