workhorse 1.5.2 → 2.0.0.rc0

data/test/workhorse/notifier_test.rb

@@ -0,0 +1,500 @@
+require 'test_helper'
+require 'tempfile'
+
+class Workhorse::NotifierTest < WorkhorseTest
+  def setup
+    super
+    @notifier = Workhorse.notifier
+    FileUtils.rm_f(wake_path)
+  end
+
+  def teardown
+    Workhorse.notifier = @notifier
+    FileUtils.rm_f(wake_path)
+  end
+
+  def test_default_notifier_is_none
+    assert_instance_of Workhorse::Notifiers::None, Workhorse.notifier
+    assert_nil Workhorse.notifier.token
+  end
+
+  def test_notifier_can_be_selected_by_symbol
+    Workhorse.notifier = :file
+    assert_instance_of Workhorse::Notifiers::FileSystem, Workhorse.notifier
+
+    Workhorse.notifier = :redis
+    assert_instance_of Workhorse::Notifiers::Redis, Workhorse.notifier
+
+    Workhorse.notifier = :none
+    assert_instance_of Workhorse::Notifiers::None, Workhorse.notifier
+  end
+
+  def test_notifier_can_be_set_to_nil
+    Workhorse.notifier = :file
+    Workhorse.notifier = nil
+
+    # Asserted on the stored value: the reader falls back to a None of its
+    # own, so it cannot tell a cleared notifier from an unset one.
+    assert_instance_of Workhorse::Notifiers::None, Workhorse.instance_variable_get(:@notifier)
+  end
+
+  def test_the_file_notifier_defaults_below_the_rails_root
+    assert_equal Rails.root.join('tmp', 'pids', 'workhorse.wake').to_s,
+                 Workhorse::Notifiers::FileSystem.new.path
+  end
+
+  def test_the_redis_notifier_takes_the_configured_client
+    redis = FakeRedis.new
+    Workhorse.notification_redis = redis
+    notifier = Workhorse::Notifiers::Redis.new
+
+    notifier.notify(queue: :mailer)
+
+    assert_equal [[Workhorse::Notifiers::Redis::DEFAULT_CHANNEL, 'mailer']], redis.published
+  ensure
+    Workhorse.notification_redis = nil
+  end
+
+  def test_unknown_notifier_is_rejected
+    assert_raises ArgumentError do
+      Workhorse.notifier = :carrier_pigeon
+    end
+  end
+
+  def test_file_notifier_token_changes_on_notify
+    notifier = file_notifier
+
+    assert_nil notifier.token
+
+    notifier.notify
+    first = notifier.token
+
+    refute_nil first
+    assert_equal first, notifier.token
+
+    sleep 0.01
+    notifier.notify
+
+    refute_equal first, notifier.token
+  end
+
+  def test_file_notifier_creates_its_directory
+    path = File.join(Dir.mktmpdir, 'deeply', 'nested', 'workhorse.wake')
+    notifier = Workhorse::Notifiers::FileSystem.new(path: path)
+
+    notifier.notify
+
+    assert File.exist?(path)
+  end
+
+  def test_file_notifier_swallows_errors
+    Tempfile.create('workhorse') do |file|
+      # A regular file cannot be a directory, on any platform.
+      notifier = Workhorse::Notifiers::FileSystem.new(path: File.join(file.path, 'workhorse.wake'))
+
+      assert_nothing_raised { notifier.notify }
+      assert_nil notifier.token
+    end
+  end
+
+  def test_enqueueing_notifies
+    Workhorse.notifier = file_notifier
+
+    assert_nil Workhorse.notifier.token
+
+    Workhorse.enqueue BasicJob.new(sleep_time: 0)
+
+    refute_nil Workhorse.notifier.token
+  end
+
+  def test_enqueueing_does_not_notify_for_jobs_that_are_not_due
+    Workhorse.notifier = file_notifier
+
+    Workhorse.enqueue BasicJob.new(sleep_time: 0), perform_at: Time.now + 60
+
+    assert_nil Workhorse.notifier.token
+  end
+
+  def test_notification_happens_after_commit
+    Workhorse.notifier = file_notifier
+
+    ActiveRecord::Base.transaction do
+      Workhorse.enqueue BasicJob.new(sleep_time: 0)
+
+      assert_nil Workhorse.notifier.token, 'must not notify before the commit'
+    end
+
+    refute_nil Workhorse.notifier.token, 'must notify after the commit'
+  end
+
+  def test_worker_starts_an_announced_job_ahead_of_its_polling_interval
+    Workhorse.notifier = file_notifier
+
+    log = capture_log do |logger|
+      with_worker(polling_interval: 60, pool_size: 1, auto_terminate: false, logger: logger) do
+        wait_for_first_poll(logger)
+
+        Workhorse.enqueue BasicJob.new(sleep_time: 0)
+
+        with_retries(30) do
+          # Re-announced on every attempt: a single announcement whose poll
+          # loses the global lock race would otherwise wait out the interval.
+          Workhorse.notifier.notify
+          assert_equal 1, Workhorse::DbJob.succeeded.count
+        end
+      end
+    end
+
+    assert_match(/Job was announced/, log)
+  end
+
+  def test_worker_without_a_notifier_waits_out_its_polling_interval
+    capture_log do |logger|
+      with_worker(polling_interval: 60, pool_size: 1, auto_terminate: false, logger: logger) do
+        wait_for_first_poll(logger)
+
+        Workhorse.enqueue BasicJob.new(sleep_time: 0)
+        sleep 1
+
+        assert_equal 1, Workhorse::DbJob.waiting.count
+      end
+    end
+  end
+
+  # Without this, marking every poll as brought forward - which would
+  # silently disable the max_global_lock_fails alarm - goes undetected.
+  def test_a_poll_that_waited_out_its_interval_is_not_marked_as_brought_forward
+    poller = Workhorse::Worker.new(polling_interval: 0.2, pool_size: 1).poller
+    poller.instance_variable_set(:@running, true)
+
+    poller.send(:sleep)
+
+    refute poller.instance_variable_get(:@poll_brought_forward),
+           'a poll that waited out its interval is the scheduled one'
+  end
+
+  def test_redis_notifier_publishes_and_counts
+    redis = FakeRedis.new
+    notifier = Workhorse::Notifiers::Redis.new(client: redis, channel: 'test:jobs')
+
+    notifier.notify(queue: :mailer)
+
+    assert_equal [['test:jobs', 'mailer']], redis.published
+
+    notifier.start
+    begin
+      redis.deliver('test:jobs', 'mailer')
+
+      with_retries(50, interval: 0.02) do
+        assert_equal 1, notifier.token
+      end
+    ensure
+      notifier.stop
+    end
+  end
+
+  # A reporter that is itself unreachable must not leave the process without
+  # a subscriber for the rest of its life.
+  def test_a_raising_exception_handler_does_not_kill_the_subscriber
+    redis = FakeRedis.new(fail_subscribe: 1)
+    notifier = Workhorse::Notifiers::Redis.new(client: redis, channel: 'test:jobs')
+
+    with_exception_handler ->(_e) { fail 'reporter is down' } do
+      notifier.start
+
+      begin
+        redis.deliver('test:jobs', 'mailer')
+
+        with_retries(100, interval: 0.05) { assert_equal 1, notifier.token }
+      ensure
+        notifier.stop
+      end
+    end
+  end
+
+  # The once-only guard has to re-arm, or only the first outage in a
+  # process's whole life is ever reported.
+  def test_a_second_outage_is_reported_again
+    redis = FakeRedis.new
+    notifier = Workhorse::Notifiers::Redis.new(client: redis, channel: 'test:jobs')
+    reported = []
+
+    with_exception_handler ->(e) { reported << e.message } do
+      notifier.start
+
+      begin
+        redis.drop!
+        with_retries(100, interval: 0.05) { assert_equal 1, reported.size }
+
+        redis.deliver('test:jobs', 'mailer')
+        with_retries(100, interval: 0.05) { assert_equal 1, notifier.token }
+
+        redis.drop!
+        with_retries(100, interval: 0.05) { assert_equal 2, reported.size }
+      ensure
+        notifier.stop
+      end
+    end
+  end
+
+  def test_redis_notifier_swallows_publish_errors
+    notifier = Workhorse::Notifiers::Redis.new(client: FakeRedis.new(fail_publish: true), channel: 'test:jobs')
+
+    assert_nothing_raised { notifier.notify(queue: :mailer) }
+  end
+
+  # A broken notifier must cost latency, not the worker.
+  def test_a_raising_notifier_does_not_take_the_worker_down
+    Workhorse.notifier = RaisingNotifier.new
+
+    with_worker(polling_interval: 0.2, pool_size: 1, auto_terminate: false) do |w|
+      Workhorse.enqueue BasicJob.new(sleep_time: 0)
+
+      with_retries(30) do
+        assert_equal 1, Workhorse::DbJob.succeeded.count
+      end
+
+      assert_equal :running, w.state
+    end
+  end
+
+  # Losing that race is expected when several workers are woken at once.
+  def test_lock_failures_of_brought_forward_polls_are_not_counted
+    Workhorse.notifier = file_notifier
+    w = Workhorse::Worker.new(polling_interval: 60, pool_size: 1)
+    poller = w.poller
+
+    poller.instance_variable_set(:@running, true)
+    poller.instance_variable_set(:@last_notification, Workhorse.notifier.token)
+
+    Workhorse.notifier.notify
+    poller.send(:sleep)
+
+    assert poller.instance_variable_get(:@poll_brought_forward),
+           'a poll following an announcement must be marked as brought forward'
+
+    with_global_lock_held do
+      poller.send(:poll)
+    end
+
+    assert_equal 0, poller.instance_variable_get(:@global_lock_fails)
+
+    poller.instance_variable_set(:@poll_brought_forward, false)
+
+    with_global_lock_held do
+      poller.send(:poll)
+    end
+
+    assert_equal 1, poller.instance_variable_get(:@global_lock_fails)
+  end
+
+  def test_redis_notifier_requires_a_client
+    Workhorse.notification_redis = nil
+    notifier = Workhorse::Notifiers::Redis.new
+
+    assert_raises RuntimeError do
+      notifier.client
+    end
+  end
+
+  # A worker with no free thread must leave the announcement pending, or the
+  # job waits out the whole polling interval once capacity frees up.
+  def test_an_announcement_arriving_while_busy_is_not_lost
+    Workhorse.notifier = file_notifier
+
+    with_worker(polling_interval: 60, pool_size: 1, auto_terminate: false) do
+      blocker = Workhorse.enqueue BasicJob.new(sleep_time: 1)
+      with_retries(60, interval: 0.05) { assert_equal 'started', blocker.reload.state }
+
+      job = Workhorse.enqueue BasicJob.new(sleep_time: 0)
+
+      with_retries(60, interval: 0.1) { assert_equal 'succeeded', job.reload.state }
+    end
+  end
+
+  # Several workers in one process share one subscriber thread; the first to
+  # stop must not take it away from the others.
+  def test_the_redis_subscriber_survives_until_the_last_worker_stops
+    redis = FakeRedis.new
+    notifier = Workhorse::Notifiers::Redis.new(client: redis, channel: 'test:jobs')
+
+    notifier.start
+    notifier.start
+    notifier.stop
+
+    begin
+      redis.deliver('test:jobs', 'mailer')
+
+      with_retries(50, interval: 0.02) { assert_equal 1, notifier.token }
+    ensure
+      notifier.stop
+    end
+  end
+
+  def test_the_file_notifier_follows_the_configured_path
+    Workhorse.notifier = :file
+    Workhorse.notification_path = wake_path
+
+    assert_equal wake_path, Workhorse.notifier.path
+  ensure
+    Workhorse.notification_path = nil
+  end
+
+  def test_the_redis_notifier_follows_the_configured_channel
+    Workhorse.notifier = :redis
+    Workhorse.notification_channel = 'custom:jobs'
+
+    assert_equal 'custom:jobs', Workhorse.notifier.channel
+  ensure
+    Workhorse.notification_channel = nil
+  end
+
+  # A subscribed connection cannot also publish, so the callable has to be
+  # asked again for the subscriber rather than the publisher being reused.
+  def test_the_redis_notifier_subscribes_on_a_client_of_its_own
+    built = []
+    notifier = Workhorse::Notifiers::Redis.new(
+      client: -> { FakeRedis.new.tap { |r| built << r } }, channel: 'test:jobs'
+    )
+
+    notifier.notify(queue: :mailer)
+    notifier.start
+
+    begin
+      with_retries(50, interval: 0.02) { assert_equal 2, built.size }
+
+      refute_equal built.first.object_id, built.last.object_id
+    ensure
+      notifier.stop
+    end
+  end
+
+  def test_the_redis_notifier_builds_a_client_from_a_callable
+    built = []
+    Workhorse.notification_redis = lambda do
+      built << :built
+      FakeRedis.new
+    end
+    notifier = Workhorse::Notifiers::Redis.new
+
+    notifier.notify(queue: :mailer)
+
+    assert_equal [:built], built
+  ensure
+    Workhorse.notification_redis = nil
+  end
+
+  private
+
+  def with_exception_handler(handler)
+    previous = Workhorse.on_exception
+    Workhorse.on_exception = handler
+    yield
+  ensure
+    Workhorse.on_exception = previous
+  end
+
+  def wake_path
+    return File.join(Dir.tmpdir, 'workhorse_test.wake')
+  end
+
+  # Holds workhorse's global lock on a connection of its own, so that the code
+  # under test sees it as taken by another worker.
+  def with_global_lock_held
+    connection = ActiveRecord::Base.connection_pool.checkout
+    connection.select_value("SELECT GET_LOCK(CONCAT(DATABASE(), '_workhorse'), 1)")
+    yield
+  ensure
+    if connection
+      connection.select_value("SELECT RELEASE_LOCK(CONCAT(DATABASE(), '_workhorse'))")
+      ActiveRecord::Base.connection_pool.checkin(connection)
+    end
+  end
+
+  def file_notifier
+    return Workhorse::Notifiers::FileSystem.new(path: wake_path)
+  end
+
+  # Waits until the worker has completed the poll it performs on startup, so
+  # that a job enqueued afterwards can only be found through a notification.
+  def wait_for_first_poll(logger)
+    with_retries(50, interval: 0.05) do
+      assert_match(/Polling DB for jobs/, logger.instance_variable_get(:@logdev).dev.string)
+    end
+  end
+
+  # Notifier whose token cannot be read, standing in for a broken custom one.
+  class RaisingNotifier < Workhorse::Notifiers::Base
+    def token
+      fail 'notifier is broken'
+    end
+  end
+
+  # Minimal stand-in for a Redis client, so that the notifier's logic can be
+  # tested without a server. `subscribe` blocks the way the real one does.
+  class FakeRedis
+    attr_reader :published
+
+    def initialize(fail_publish: false, fail_subscribe: 0)
+      @published = []
+      @fail_publish = fail_publish
+      @fail_subscribe = fail_subscribe
+      @queue = Queue.new
+    end
+
+    def publish(channel, message)
+      fail 'Connection refused' if @fail_publish
+
+      @published << [channel, message]
+    end
+
+    def deliver(channel, message)
+      @queue << [channel, message]
+    end
+
+    # Makes the current subscription fail, as a dropped connection would.
+    def drop!
+      @queue << :drop
+    end
+
+    def dup
+      return self
+    end
+
+    def subscribe(_channel)
+      if @fail_subscribe > 0
+        @fail_subscribe -= 1
+        fail 'Connection refused'
+      end
+
+      on = Callbacks.new
+      yield on
+      on.subscribed!
+
+      loop do
+        item = @queue.pop
+        fail 'Connection reset' if item == :drop
+
+        on.call(*item)
+      end
+    end
+
+    class Callbacks
+      def message(&block)
+        @message = block
+      end
+
+      def subscribe(&block)
+        @subscribe = block
+      end
+
+      def subscribed!
+        @subscribe&.call
+      end
+
+      def call(channel, message)
+        @message&.call(channel, message)
+      end
+    end
+  end
+end

data/test/workhorse/poller_test.rb

@@ -71,10 +71,8 @@ class Workhorse::PollerTest < WorkhorseTest
     assert_equal [], w.poller.send(:valid_queues)
   end
 
-  # Not every adapter returns a result set from `execute`: the Oracle enhanced
-  # adapter, for instance, returns `true` for queries, both before and after
-  # version 7.0.0. Querying the valid queues must therefore not rely on the
-  # return value of `execute`.
+  # Not every adapter returns a result set from `execute`, so querying the
+  # valid queues must not rely on its return value.
   def test_valid_queues_without_usable_execute
     w = Workhorse::Worker.new(polling_interval: 60)
 
@@ -139,6 +137,12 @@ class Workhorse::PollerTest < WorkhorseTest
     10.times do
       Process.fork do
         work 3, pool_size: 1, polling_interval: 0.1
+      ensure
+        # Exit without running the at_exit handlers of the test process: one
+        # of them is Minitest's, which joins threads this fork did not
+        # inherit and hangs the child, leaving waitall below waiting forever.
+        # In an ensure, as a child that raised must not run them either.
+        exit!(0)
       end
     end