def handle(self, *arg, **options):
if options.get('status'):
print(Control('dispatcher').status())
return
if options.get('running'):
print(Control('dispatcher').running())
return
if options.get('reload'):
return Control('dispatcher').control({'control': 'reload'})
# It's important to close these because we're _about_ to fork, and we
# don't want the forked processes to inherit the open sockets
# for the DB and cache connections (that way lies race conditions)
django_connection.close()
django_cache.close()
# spawn a daemon thread to periodically enqueues scheduled tasks
# (like the node heartbeat)
periodic.run_continuously()
consumer = None
try:
queues = ['tower_broadcast_all', get_local_queuename()]
consumer = AWXConsumerPG('dispatcher', TaskWorker(), queues,
AutoscalePool(min_workers=4))
consumer.run()
except KeyboardInterrupt:
logger.debug('Terminating Task Dispatcher')
if consumer:
consumer.stop()
def handle(self, *arg, **options):
if options.get('status'):
print(Control('dispatcher').status())
return
if options.get('running'):
print(Control('dispatcher').running())
return
if options.get('reload'):
return Control('dispatcher').control({'control': 'reload'})
# It's important to close these because we're _about_ to fork, and we
# don't want the forked processes to inherit the open sockets
# for the DB and memcached connections (that way lies race conditions)
django_connection.close()
django_cache.close()
# spawn a daemon thread to periodically enqueues scheduled tasks
# (like the node heartbeat)
periodic.run_continuously()
reaper.reap()
consumer = None
# don't ship external logs inside the dispatcher's parent process
# this exists to work around a race condition + deadlock bug on fork
# in cpython itself:
# https://bugs.python.org/issue37429
AWXProxyHandler.disable()
with Connection(settings.BROKER_URL) as conn:
try:
bcast = 'tower_broadcast_all'
queues = [
Queue(q, Exchange(q), routing_key=q)
for q in (settings.AWX_CELERY_QUEUES_STATIC + [get_local_queuename()])
]
queues.append(
Queue(
construct_bcast_queue_name(bcast),
exchange=Exchange(bcast, type='fanout'),
routing_key=bcast,
reply=True
)
)
consumer = AWXConsumer(
'dispatcher',
conn,
TaskWorker(),
queues,
AutoscalePool(min_workers=4)
)
consumer.run()
except KeyboardInterrupt:
logger.debug('Terminating Task Dispatcher')
if consumer:
consumer.stop()
def handle(self, *arg, **options):
if options.get('status'):
print Control('dispatcher').status()
return
if options.get('running'):
print Control('dispatcher').running()
return
if options.get('reload'):
return Control('dispatcher').control({'control': 'reload'})
# It's important to close these because we're _about_ to fork, and we
# don't want the forked processes to inherit the open sockets
# for the DB and memcached connections (that way lies race conditions)
django_connection.close()
django_cache.close()
beat = Process(target=self.beat)
beat.daemon = True
beat.start()
reaper.reap()
consumer = None
with Connection(settings.BROKER_URL) as conn:
try:
bcast = 'tower_broadcast_all'
queues = [
Queue(q, Exchange(q), routing_key=q)
for q in (settings.AWX_CELERY_QUEUES_STATIC + [get_local_queuename()])
]
queues.append(
Queue(
construct_bcast_queue_name(bcast),
exchange=Exchange(bcast, type='fanout'),
routing_key=bcast,
reply=True
)
)
consumer = AWXConsumer(
'dispatcher',
conn,
TaskWorker(),
queues,
AutoscalePool(min_workers=4)
)
consumer.run()
except KeyboardInterrupt:
logger.debug('Terminating Task Dispatcher')
if consumer:
consumer.stop()
def handle(self, *arg, **options):
if options.get('status'):
print(Control('dispatcher').status())
return
if options.get('running'):
print(Control('dispatcher').running())
return
if options.get('reload'):
return Control('dispatcher').control({'control': 'reload'})
# It's important to close these because we're _about_ to fork, and we
# don't want the forked processes to inherit the open sockets
# for the DB and memcached connections (that way lies race conditions)
django_connection.close()
django_cache.close()
# spawn a daemon thread to periodically enqueues scheduled tasks
# (like the node heartbeat)
periodic.run_continuously()
reaper.reap()
consumer = None
# don't ship external logs inside the dispatcher's parent process
# this exists to work around a race condition + deadlock bug on fork
# in cpython itself:
# https://bugs.python.org/issue37429
AWXProxyHandler.disable()
try:
queues = ['tower_broadcast_all', get_local_queuename()]
consumer = AWXConsumerPG('dispatcher', TaskWorker(), queues,
AutoscalePool(min_workers=4))
consumer.run()
except KeyboardInterrupt:
logger.debug('Terminating Task Dispatcher')
if consumer:
consumer.stop()
def setup_method(self, test_method):
self.pool = AutoscalePool(min_workers=2, max_workers=10)
class TestAutoScaling:
def setup_method(self, test_method):
self.pool = AutoscalePool(min_workers=2, max_workers=10)
def teardown_method(self, test_method):
self.pool.stop(signal.SIGTERM)
def test_scale_up(self):
result_queue = multiprocessing.Queue()
self.pool.init_workers(SlowResultWriter().work_loop, result_queue)
# start with two workers, write an event to each worker and make it busy
assert len(self.pool) == 2
for i, w in enumerate(self.pool.workers):
w.put('Hello, Worker {}'.format(0))
assert len(self.pool) == 2
# wait for the subprocesses to start working on their tasks and be marked busy
time.sleep(1)
assert self.pool.should_grow
# write a third message, expect a new worker to spawn because all
# workers are busy
self.pool.write(0, 'Hello, Worker {}'.format(2))
assert len(self.pool) == 3
def test_scale_down(self):
self.pool.init_workers(ResultWriter().work_loop,
multiprocessing.Queue())
# start with two workers, and scale up to 10 workers
assert len(self.pool) == 2
for i in range(8):
self.pool.up()
assert len(self.pool) == 10
# cleanup should scale down to 8 workers
with mock.patch('awx.main.dispatch.reaper.reap') as reap:
self.pool.cleanup()
reap.assert_called()
assert len(self.pool) == 2
def test_max_scale_up(self):
self.pool.init_workers(ResultWriter().work_loop,
multiprocessing.Queue())
assert len(self.pool) == 2
for i in range(25):
self.pool.up()
assert self.pool.max_workers == 10
assert self.pool.full is True
assert len(self.pool) == 10
def test_equal_worker_distribution(self):
# if all workers are busy, spawn new workers *before* adding messages
# to an existing queue
self.pool.init_workers(SlowResultWriter().work_loop,
multiprocessing.Queue)
# start with two workers, write an event to each worker and make it busy
assert len(self.pool) == 2
for i in range(10):
self.pool.write(0, 'Hello, World!')
assert len(self.pool) == 10
for w in self.pool.workers:
assert w.busy
assert len(w.managed_tasks) == 1
# the queue is full at 10, the _next_ write should put the message into
# a worker's backlog
assert len(self.pool) == 10
for w in self.pool.workers:
assert w.messages_sent == 1
self.pool.write(0, 'Hello, World!')
assert len(self.pool) == 10
assert self.pool.workers[0].messages_sent == 2
def test_lost_worker_autoscale(self):
# if a worker exits, it should be replaced automatically up to min_workers
self.pool.init_workers(ResultWriter().work_loop,
multiprocessing.Queue())
# start with two workers, kill one of them
assert len(self.pool) == 2
assert not self.pool.should_grow
alive_pid = self.pool.workers[1].pid
self.pool.workers[0].process.terminate()
time.sleep(1) # wait a moment for sigterm
# clean up and the dead worker
with mock.patch('awx.main.dispatch.reaper.reap') as reap:
self.pool.cleanup()
reap.assert_called()
assert len(self.pool) == 1
assert self.pool.workers[0].pid == alive_pid
# the next queue write should replace the lost worker
self.pool.write(0, 'Hello, Worker')
assert len(self.pool) == 2
