def _handle_running_task(self, task_id):
keyt = 'task:%s' % task_id
_, service = self._get_service(keyt=keyt)
self._logger.debug('- checking activity of task: %s', task_id)
data = json.loads(self._redis.hget(keyt, 'job'))
try:
status = service.status(task_id, data)
if status == 'dead':
self._logger.info(
'%s: task no longer running on %s, request termination',
task_id, service.name)
task.terminate(self._redis, task_id, phase='exited')
else:
task.work_queue(self._redis,
task_id,
service.name,
delay=service.is_notifying_activity and 600
or 120)
except Exception as e:
self._logger.info('cannot get status for [%s] - %s', task_id,
str(e))
self._redis.hincrby(keyt, 'status_fail', 1)
if int(self._redis.hget(keyt, 'status_fail')) > 4:
return task.terminate(self._redis,
task_id,
phase='lost_connection')
self._redis.hdel(keyt, 'status_fail')
return None
def terminate(task_id):
with redis.acquire_lock(task_id):
current_status = task.info(redis, task_id, "status")
if current_status is None:
flask.abort(flask.make_response(flask.jsonify(message="task %s unknown" % task_id), 404))
elif current_status == "stopped":
return flask.jsonify(message="%s already stopped" % task_id)
phase = flask.request.args.get('phase')
task.terminate(redis, task_id, phase=phase)
return flask.jsonify(message="terminating %s" % task_id)
def terminate(task_id):
with redis.acquire_lock(task_id):
current_status = task.info(redis, taskfile_dir, task_id, "status")
if current_status is None:
abort(flask.make_response(flask.jsonify(message="task %s unknown" % task_id), 404))
elif current_status == "stopped":
return flask.jsonify(message="%s already stopped" % task_id)
phase = flask.request.args.get('phase')
res = post_function('GET/task/terminate', task_id, phase)
if res:
task.terminate(redis, task_id, phase="publish_error")
return flask.jsonify(message="problem while posting model: %s" % res)
task.terminate(redis, task_id, phase=phase)
return flask.jsonify(message="terminating %s" % task_id)
def try_create(next_task_id):
next_keyt = 'task:%s' % next_task_id
parent = self._redis.hget(next_keyt, 'parent')
task_entity = task.get_owner_entity(self._redis, next_task_id)
if task_entity not in resource_mgr.entities_usage:
self._logger.error(
"\t[Task %s] entity %s - without usage limit !",
next_task_id, task_entity)
return None
# check parent dependency
if parent:
keyp = 'task:%s' % parent
if self._redis.exists(keyp):
# if the parent task is in the database, check for dependencies
parent_status = self._redis.hget(keyp, 'status')
if parent_status != 'stopped':
if parent_status == 'running':
# parent is still running so update queued time to be as close
# as possible to terminate time of parent task
self._redis.hset(next_keyt, "queued_time",
time.time())
return None
if self._redis.hget(keyp, 'message') != 'completed':
task.terminate(self._redis,
next_task_id,
phase='dependency_error')
return None
task_capacity = Capacity(self._redis.hget(next_keyt, 'ngpus'),
self._redis.hget(next_keyt, 'ncpus'))
candidate_task = CandidateTask(
next_task_id, task_entity, self._redis, task_capacity,
resource_mgr.entities_usage[task_entity], self._logger)
# check now the task has a chance to be processed by any machine
for _, machine in six.iteritems(resource_mgr._machines):
can_be_processed = machine._is_authorized(candidate_task._entity, candidate_task._capacity) \
and candidate_task._capacity.inf_or_eq(machine._init_capacity)
if can_be_processed:
return candidate_task
return None
def run(self):
self._logger.info('Starting worker')
# Subscribe to beat expiration.
pubsub = self._redis.pubsub()
pubsub.psubscribe('__keyspace@0__:beat:*')
pubsub.psubscribe('__keyspace@0__:queue:*')
while True:
message = pubsub.get_message()
if message:
channel = message['channel']
data = message['data']
if data == 'expired':
if channel.startswith('__keyspace@0__:beat:'):
task_id = channel[20:]
self._logger.info('%s: task expired', task_id)
with self._redis.acquire_lock(task_id):
task.terminate(self._redis,
task_id,
phase='expired')
elif channel.startswith('__keyspace@0__:queue:'):
task_id = channel[21:]
task.queue(self._redis, task_id)
else:
task_id = task.unqueue(self._redis)
if task_id is not None:
try:
self._advance_task(task_id)
except RuntimeWarning:
self._logger.warning(
'%s: failed to acquire a lock, retrying', task_id)
task.queue(self._redis, task_id)
except Exception as e:
self._logger.error('%s: %s', task_id, str(e))
with self._redis.acquire_lock(task_id):
task.terminate(self._redis,
task_id,
phase="launch_error")
time.sleep(0.1)
def run(self):
signal.signal(signal.SIGTERM, graceful_exit)
signal.signal(signal.SIGINT, graceful_exit)
pubsub = self._redis.pubsub()
pubsub.psubscribe('__keyspace@0__:beat:*')
pubsub.psubscribe('__keyspace@0__:queue:*')
while True:
message = pubsub.get_message()
if message:
channel = message['channel']
data = message['data']
if data == 'expired':
# task expired, not beat was received
if channel.startswith('__keyspace@0__:beat:'):
task_id = channel[20:]
service = self._redis.hget('task:' + task_id, 'service')
if service in self._services:
self._logger.info('%s: task expired', task_id)
auth_token = self._redis.hget('task:%s' % task_id, 'token')
callback_url = self._services[service]._config.get('callback_url')
if auth_token:
callback_url = callback_url.replace("://", "://" + auth_token + ":x@")
r = requests.get(os.path.join(callback_url, "task/terminate", task_id),
params={'phase': 'expired'})
if r.status_code != 200:
self._logger.warning('incorrect result from \'task/terminate\' service: %s' % r.text)
with self._redis.acquire_lock(task_id):
task.terminate(self._redis, task_id, phase='expired')
# expired in the queue - comes back in the work queue
elif channel.startswith('__keyspace@0__:queue:'):
task_id = channel[21:]
service = self._redis.hget('task:' + task_id, 'service')
if service in self._services:
self._logger.info('%s: move to work queue', task_id)
task.work_queue(self._redis, task_id, service)
time.sleep(self._work_cycle)
def run(self):
signal.signal(signal.SIGTERM, graceful_exit)
signal.signal(signal.SIGINT, graceful_exit)
self._logger.info('Starting...')
counter = 0
while True:
# process one element from work queue
task_id = task.work_unqueue(self._redis, self._service)
if task_id is not None:
try:
self._advance_task(task_id)
except RuntimeWarning:
self._logger.warning(
'%s: failed to acquire a lock, retrying', task_id)
task.work_queue(self._redis, task_id, self._service)
except Exception as e:
self._logger.error('%s: %s', task_id, str(e))
with self._redis.acquire_lock(task_id):
task.set_log(self._redis, self._taskfile_dir, task_id,
str(e))
task.terminate(self._redis,
task_id,
phase="launch_error")
self._logger.info(traceback.format_exc())
# every 0.01s * refresh_counter - check if we can find some free resource
if counter > self._refresh_counter:
# if there are some queued tasks, look for free resources
if self._redis.exists('queued:%s' % self._service):
self._logger.debug('checking processes on : %s',
self._service)
self._select_best_task_to_process(
self._services[self._service])
counter = 0
counter += 1
time.sleep(self._work_cycle)
def run(self):
signal.signal(signal.SIGTERM, graceful_exit)
signal.signal(signal.SIGINT, graceful_exit)
pubsub = self._redis.pubsub()
pubsub.psubscribe('__keyspace@0__:beat:*')
pubsub.psubscribe('__keyspace@0__:queue:*')
while True:
message = pubsub.get_message()
if message:
channel = message['channel']
data = message['data']
if data == 'expired':
# task expired, not beat was received
if channel.startswith('__keyspace@0__:beat:'):
task_id = channel[20:]
service = self._redis.hget('task:' + task_id,
'service')
if service in self._services:
self._logger.info('%s: task expired', task_id)
with self._redis.acquire_lock(task_id):
task.terminate(self._redis,
task_id,
phase='expired')
# expired in the queue - comes back in the work queue
elif channel.startswith('__keyspace@0__:queue:'):
task_id = channel[21:]
service = self._redis.hget('task:' + task_id,
'service')
if service in self._services:
self._logger.info('%s: move to work queue',
task_id)
task.work_queue(self._redis, task_id, service)
time.sleep(self._work_cycle)
def _advance_task(self, task_id):
"""Tries to advance the task to the next status. If it can, re-queue it immediately
to process the next stage. Otherwise, re-queue it after some delay to try again.
"""
keyt = 'task:%s' % task_id
with self._redis.acquire_lock(keyt, acquire_timeout=1, expire_time=600):
status = self._redis.hget(keyt, 'status')
if status == 'stopped':
return
service_name = self._redis.hget(keyt, 'service')
if service_name not in self._services:
raise ValueError('unknown service %s' % service_name)
service = self._services[service_name]
self._logger.info('%s: trying to advance from status %s', task_id, status)
if status == 'queued':
resource = self._redis.hget(keyt, 'resource')
parent = self._redis.hget(keyt, 'parent')
if parent:
keyp = 'task:%s' % parent
# if the parent task is in the database, check for dependencies
if self._redis.exists(keyp):
status = self._redis.hget(keyp, 'status')
if status == 'stopped':
if self._redis.hget(keyp, 'message') != 'completed':
task.terminate(self._redis, task_id, phase='dependency_error')
return
else:
self._logger.warning('%s: depending on other task, waiting', task_id)
task.service_queue(self._redis, task_id, service.name)
return
ngpus = int(self._redis.hget(keyt, 'ngpus'))
resource, available_gpus = self._allocate_resource(task_id, resource, service, ngpus)
if resource is not None:
self._logger.info('%s: resource %s reserved (%d/%d)',
task_id, resource, available_gpus, ngpus)
self._redis.hset(keyt, 'alloc_resource', resource)
if ngpus == available_gpus:
task.set_status(self._redis, keyt, 'allocated')
else:
task.set_status(self._redis, keyt, 'allocating')
task.work_queue(self._redis, task_id, service_name)
else:
self._logger.warning('%s: no resources available, waiting', task_id)
task.service_queue(self._redis, task_id, service.name)
elif status == 'allocating':
resource = self._redis.hget(keyt, 'alloc_resource')
keyr = 'resource:%s:%s' % (service.name, resource)
ngpus = int(self._redis.hget(keyt, 'ngpus'))
already_allocated_gpus = 0
for k, v in six.iteritems(self._redis.hgetall(keyr)):
if v == task_id:
already_allocated_gpus += 1
capacity = service.list_resources()[resource]
available_gpus, remaining_gpus = self._reserve_resource(service, resource,
capacity, task_id,
ngpus - already_allocated_gpus,
0, -1, True)
self._logger.warning('task: %s - resource: %s (capacity %d)- already %d - available %d', task_id, resource, capacity, already_allocated_gpus, available_gpus)
if available_gpus == ngpus - already_allocated_gpus:
task.set_status(self._redis, keyt, 'allocated')
key_reserved = 'reserved:%s:%s' % (service.name, resource)
self._redis.delete(key_reserved)
task.work_queue(self._redis, task_id, service.name)
else:
task.work_queue(self._redis, task_id, service.name,
delay=service.is_notifying_activity and 120 or 30)
elif status == 'allocated':
content = json.loads(self._redis.hget(keyt, 'content'))
resource = self._redis.hget(keyt, 'alloc_resource')
self._logger.info('%s: launching on %s', task_id, service.name)
try:
keyr = 'resource:%s:%s' % (service.name, resource)
lgpu = []
for k, v in six.iteritems(self._redis.hgetall(keyr)):
if v == task_id:
lgpu.append(k)
self._redis.hset(keyt, 'alloc_lgpu', ",".join(lgpu))
data = service.launch(
task_id,
content['options'],
lgpu,
resource,
content['docker']['registry'],
content['docker']['image'],
content['docker']['tag'],
content['docker']['command'],
task.file_list(self._redis, task_id),
content['wait_after_launch'])
except EnvironmentError as e:
# the resource is not available and will be set busy
self._block_resource(resource, service, str(e))
# set the task as queued again
self._redis.hdel(keyt, 'alloc_resource')
self._release_resource(service, resource, task_id)
task.set_status(self._redis, keyt, 'queued')
task.service_queue(self._redis, task_id, service.name)
self._logger.info('could not launch [%s] %s on %s: blocking resource', str(e), task_id, resource)
return
except Exception as e:
# all other errors make the task fail
task.append_log(self._redis, task_id, str(e))
task.terminate(self._redis, task_id, phase='launch_error')
return
self._logger.info('%s: task started on %s', task_id, service.name)
self._redis.hset(keyt, 'job', json.dumps(data))
task.set_status(self._redis, keyt, 'running')
# For services that do not notify their activity, we should
# poll the task status more regularly.
task.work_queue(self._redis, task_id, service.name,
delay=service.is_notifying_activity and 120 or 30)
elif status == 'running':
self._logger.debug('- checking activity of task: %s', task_id)
data = json.loads(self._redis.hget(keyt, 'job'))
status = service.status(task_id, data)
if status == 'dead':
self._logger.info('%s: task no longer running on %s, request termination',
task_id, service.name)
task.terminate(self._redis, task_id, phase='exited')
else:
task.work_queue(self._redis, task_id, service.name,
delay=service.is_notifying_activity and 120 or 30)
elif status == 'terminating':
data = self._redis.hget(keyt, 'job')
if data is not None:
container_id = self._redis.hget(keyt, 'container_id')
data = json.loads(data)
data['container_id'] = container_id
self._logger.info('%s: terminating task (%s)', task_id, json.dumps(data))
try:
service.terminate(data)
self._logger.info('%s: terminated', task_id)
except Exception:
self._logger.warning('%s: failed to terminate', task_id)
resource = self._redis.hget(keyt, 'alloc_resource')
self._release_resource(service, resource, task_id)
task.set_status(self._redis, keyt, 'stopped')
task.disable(self._redis, task_id)
def run(self):
self._logger.info('Starting worker')
# Subscribe to beat expiration.
pubsub = self._redis.pubsub()
pubsub.psubscribe('__keyspace@0__:beat:*')
pubsub.psubscribe('__keyspace@0__:queue:*')
counter = 0
while True:
message = pubsub.get_message()
if message:
channel = message['channel']
data = message['data']
if data == 'expired':
self._logger.warning('received expired event on channel %s', channel)
if channel.startswith('__keyspace@0__:beat:'):
task_id = channel[20:]
service = self._redis.hget('task:'+task_id, 'service')
if service in self._services:
self._logger.info('%s: task expired', task_id)
with self._redis.acquire_lock(task_id):
task.terminate(self._redis, task_id, phase='expired')
elif channel.startswith('__keyspace@0__:queue:'):
task_id = channel[21:]
service = self._redis.hget('task:'+task_id, 'service')
if service in self._services:
task.work_queue(self._redis, task_id, service)
else:
for service in self._services:
task_id = task.work_unqueue(self._redis, service)
if task_id is not None:
try:
self._advance_task(task_id)
except RuntimeWarning:
self._logger.warning(
'%s: failed to acquire a lock, retrying', task_id)
task.work_queue(self._redis, task_id, service)
except Exception as e:
self._logger.error('%s: %s', task_id, str(e))
with self._redis.acquire_lock(task_id):
task.set_log(self._redis, task_id, str(e))
task.terminate(self._redis, task_id, phase="launch_error")
else:
if counter > self._refresh_counter:
resources = self._services[service].list_resources()
for resource in resources:
keyr = 'resource:%s:%s' % (service, resource)
key_busy = 'busy:%s:%s' % (service, resource)
key_reserved = 'reserved:%s:%s' % (service, resource)
if not self._redis.exists(key_busy) and self._redis.hlen(keyr) < resources[resource]:
if self._redis.exists(key_reserved) and self._redis.ttl('queue:'+self._redis.get(key_reserved))>10:
self._redis.expire('queue:'+self._redis.get(key_reserved), 5)
break
if self._redis.exists('queued:%s' % service):
resources = self._services[service].list_resources()
self._logger.debug('checking processes on : %s', service)
availableResource = False
for resource in resources:
keyr = 'resource:%s:%s' % (service, resource)
key_busy = 'busy:%s:%s' % (service, resource)
key_reserved = 'reserved:%s:%s' % (service, resource)
if not self._redis.exists(key_busy) and self._redis.hlen(keyr) < resources[resource]:
if not self._redis.exists(key_reserved):
availableResource = True
break
if availableResource:
self._logger.debug('resources available on %s - trying dequeuing', service)
self._service_unqueue(self._services[service])
if counter > self._refresh_counter:
counter = 0
counter += 1
time.sleep(0.01)
def _handle_allocated_task(self, task_id):
keyt = 'task:%s' % task_id
_, service = self._get_service(keyt=keyt)
content = json.loads(self._redis.hget(keyt, 'content'))
resource = self._redis.hget(keyt, 'alloc_resource')
self._logger.info('%s: launching on %s', task_id, service.name)
try:
entity_config = self._get_current_config(task_id)
keygr = 'gpu_resource:%s:%s' % (service.name, resource)
lgpu = []
for k, v in six.iteritems(self._redis.hgetall(keygr)):
if v == task_id:
lgpu.append(k)
self._redis.hset(keyt, 'alloc_lgpu', ",".join(lgpu))
keycr = 'cpu_resource:%s:%s' % (service.name, resource)
lcpu = []
for k, v in six.iteritems(self._redis.hgetall(keycr)):
if v == task_id:
lcpu.append(k)
self._redis.hset(keyt, 'alloc_lcpu', ",".join(lcpu))
data = service.launch(
task_id, content['options'], (lgpu, lcpu), resource,
entity_config["storages"], entity_config["docker"],
content['docker']['registry'], content['docker']['image'],
content['docker']['tag'], content['docker']['command'],
task.file_list(self._taskfile_dir,
task_id), content['wait_after_launch'],
self._redis.hget(keyt, 'token'),
content.get('support_statistics'))
except EnvironmentError as e:
# the resource is not available and will be set busy
self._block_resource(resource, service, str(e))
self._redis.hdel(keyt, 'alloc_resource')
# set the task as queued again
self._release_resource(
service, resource, task_id,
Capacity(self._redis.hget(keyt, 'ngpus'),
self._redis.hget(keyt, 'ncpus')))
status = self._redis.hget(keyt, 'status')
if status == 'terminating':
return None
task.set_status(self._redis, keyt, 'queued')
task.service_queue(self._redis, task_id, service.name)
self._logger.info(
'could not launch [%s] %s on %s: blocking resource', str(e),
task_id, resource)
self._logger.info(traceback.format_exc())
return None
except Exception as e:
# all other errors make the task fail
self._logger.info('fail task [%s] - %s', task_id, str(e))
self._logger.info(traceback.format_exc())
task.append_log(self._taskfile_dir, task_id, str(e))
auth_token = self._redis.hget(keyt, 'token')
callback_url = service._config.get('callback_url')
if auth_token:
callback_url = callback_url.replace("://",
"://" + auth_token + ":x@")
r = requests.get(os.path.join(callback_url, "task/terminate",
task_id),
params={'phase': 'launch_error'})
if r.status_code != 200:
raise RuntimeError(
'incorrect result from \'task/terminate\' service: %s' %
r.text) from e
task.terminate(self._redis, task_id, phase='launch_error')
self._logger.info(traceback.format_exc())
return None
self._logger.info('%s: task started on %s', task_id, service.name)
self._redis.hset(keyt, 'job', json.dumps(data))
status = self._redis.hget(keyt, 'status')
if status == 'terminating':
return None
task.set_status(self._redis, keyt, 'running')
# For services that do not notify their activity, we should
# poll the task status more regularly.
task.work_queue(self._redis,
task_id,
service.name,
delay=service.is_notifying_activity and 120 or 30)
return None
def _service_unqueue(self, service):
"""find the best next task to push to the work queue
"""
with self._redis.acquire_lock('service:' + service.name):
queue = 'queued:%s' % service.name
count = self._redis.llen(queue)
idx = 0
preallocated_task_count = {}
preallocated_task_resource = {}
avail_resource = {}
resources = service.list_resources()
reserved = {}
# list free cpu/gpus on each node
for resource in resources:
current_xpu_usage = Capacity()
capacity = resources[resource]
keygr = 'gpu_resource:%s:%s' % (self._service, resource)
keycr = 'cpu_resource:%s:%s' % (self._service, resource)
key_reserved = 'reserved:%s:%s' % (service.name, resource)
gpu_tasks = self._redis.hgetall(keygr)
cpu_tasks = self._redis.hgetall(keycr)
task_reserved = self._redis.get(key_reserved)
# can not launch multiple tasks on service with no multi-tasking (ec2)
if not service.resource_multitask and \
not task_reserved and \
(gpu_tasks or cpu_tasks):
continue
for k, v in six.iteritems(gpu_tasks):
if v in preallocated_task_count:
preallocated_task_count[v].incr_ngpus(1)
else:
preallocated_task_count[v] = Capacity(ngpus=1)
preallocated_task_resource[v] = resource
current_xpu_usage.incr_ngpus(1)
for k, v in six.iteritems(cpu_tasks):
if v in preallocated_task_count:
preallocated_task_count[v].incr_ncpus(1)
else:
preallocated_task_count[v] = Capacity(ncpus=1)
preallocated_task_resource[v] = resource
current_xpu_usage.incr_ncpus(1)
available_xpus = capacity - current_xpu_usage
avail_resource[resource] = available_xpus
reserved[resource] = task_reserved
self._logger.debug("\tresource %s - reserved: %s - free %s",
resource, task_reserved or "False",
available_xpus)
if len(avail_resource) == 0:
return
# Go through the tasks, find if there are tasks that can be launched and
# queue the best one
best_task_id = None
best_task_priority = -10000
best_task_queued_time = 0
while count > 0:
count -= 1
next_task_id = self._redis.lindex(queue, count)
if next_task_id is not None:
next_keyt = 'task:%s' % next_task_id
# self._logger.debug("\tcheck task: %s", next_task_id)
parent = self._redis.hget(next_keyt, 'parent')
# check parent dependency
if parent:
keyp = 'task:%s' % parent
if self._redis.exists(keyp):
# if the parent task is in the database, check for dependencies
parent_status = self._redis.hget(keyp, 'status')
if parent_status != 'stopped':
if parent_status == 'running':
# parent is still running so update queued time to be as close
# as possible to terminate time of parent task
self._redis.hset(next_keyt, "queued_time",
time.time())
continue
else:
if self._redis.hget(keyp,
'message') != 'completed':
task.terminate(self._redis,
next_task_id,
phase='dependency_error')
continue
nxpus = Capacity(self._redis.hget(next_keyt, 'ngpus'),
self._redis.hget(next_keyt, 'ncpus'))
foundResource = False
if next_task_id in preallocated_task_count:
# if task is pre-allocated, can only continue on the same node
r = preallocated_task_resource[next_task_id]
nxpus -= preallocated_task_count[next_task_id]
avail_r = avail_resource[r]
foundResource = (nxpus.ngpus == 0 and avail_r.ncpus !=
0) or (nxpus.ngpus != 0
and avail_r.ngpus != 0)
else:
# can the task be launched on any node
for r, v in six.iteritems(avail_resource):
# cannot launch a new task on a reserved node
if reserved[r]:
continue
if ((nxpus.ngpus > 0
and resources[r].ngpus >= nxpus.ngpus
and v.ngpus > 0)
or (nxpus.ngpus == 0 and v.ncpus >= 0)):
foundResource = True
break
if not foundResource:
continue
priority = int(self._redis.hget(next_keyt, 'priority'))
queued_time = float(
self._redis.hget(next_keyt, 'queued_time'))
if priority > best_task_priority or (
priority == best_task_priority
and best_task_queued_time > queued_time):
best_task_priority = priority
best_task_id = next_task_id
best_task_queued_time = queued_time
if best_task_id:
self._logger.info('selected %s to be launched on %s',
best_task_id, service.name)
task.work_queue(self._redis, best_task_id, service.name)
self._redis.lrem(queue, 0, best_task_id)
def run(self):
self._logger.info('Starting worker')
# Subscribe to beat expiration.
pubsub = self._redis.pubsub()
pubsub.psubscribe('__keyspace@0__:beat:*')
pubsub.psubscribe('__keyspace@0__:queue:*')
counter = 0
counter_beat = 1000
while True:
counter_beat += 1
# every 1000 * 0.01s (10s) - check & reset beat of the worker
if counter_beat > 1000:
counter_beat = 0
if self._redis.exists(self._worker_id):
self._redis.hset(self._worker_id, "beat_time", time.time())
self._redis.expire(self._worker_id, 1200)
else:
self._logger.info('stopped by key expiration/removal')
sys.exit(0)
# every 100 * 0.01s (1s) - check worker administration command
if counter_beat % 100 == 0:
workeradmin.process(self._logger, self._redis, self._service)
if (self._default_config_timestamp
and self._redis.hget('default', 'timestamp') !=
self._default_config_timestamp):
self._logger.info(
'stopped by default configuration change')
sys.exit(0)
# process one message from the queue
message = pubsub.get_message()
if message:
channel = message['channel']
data = message['data']
if data == 'expired':
# task expired, not beat was received
if channel.startswith('__keyspace@0__:beat:'):
task_id = channel[20:]
service = self._redis.hget('task:' + task_id,
'service')
if service in self._services:
self._logger.info('%s: task expired', task_id)
with self._redis.acquire_lock(task_id):
task.terminate(self._redis,
task_id,
phase='expired')
# expired in the queue - comes back in the work queue
elif channel.startswith('__keyspace@0__:queue:'):
task_id = channel[21:]
service = self._redis.hget('task:' + task_id,
'service')
if service in self._services:
self._logger.info('%s: move to work queue',
task_id)
task.work_queue(self._redis, task_id, service)
# process one element from work queue
task_id = task.work_unqueue(self._redis, self._service)
if task_id is not None:
try:
self._advance_task(task_id)
except RuntimeWarning:
self._logger.warning(
'%s: failed to acquire a lock, retrying', task_id)
task.work_queue(self._redis, task_id, self._service)
except Exception as e:
self._logger.error('%s: %s', task_id, str(e))
with self._redis.acquire_lock(task_id):
task.set_log(self._redis, self._taskfile_dir, task_id,
str(e))
task.terminate(self._redis,
task_id,
phase="launch_error")
# every 0.01s * refresh_counter - check if we can find some free resource
if counter > self._refresh_counter:
# if there are some queued tasks, look for free resources
if self._redis.exists('queued:%s' % self._service):
self._logger.debug('checking processes on : %s',
self._service)
self._service_unqueue(self._services[self._service])
counter = 0
counter += 1
time.sleep(0.01)
def _service_unqueue(self, service):
"""find the best next task to push to the work queue
"""
with self._redis.acquire_lock('service:' + service.name):
queue = 'queued:%s' % service.name
count = self._redis.llen(queue)
idx = 0
preallocated_task_count = {}
preallocated_task_resource = {}
avail_resource = {}
resources = service.list_resources()
reserved = {}
# list free cpu/gpus on each node
for resource in resources:
keyr = 'gpu_resource:%s:%s' % (self._service, resource)
keyc = 'ncpus:%s:%s' % (self._service, resource)
available_cpus = int(self._redis.get(keyc))
current_gpu_usage = 0
gpu_capacity = resources[resource]
for k, v in six.iteritems(self._redis.hgetall(keyr)):
if v in preallocated_task_count:
preallocated_task_count[v] += 1
else:
preallocated_task_count[v] = 1
preallocated_task_resource[v] = resource
current_gpu_usage += 1
available_gpus = gpu_capacity - current_gpu_usage
avail_resource[resource] = (available_cpus, available_gpus)
key_reserved = 'reserved:%s:%s' % (service.name, resource)
reserved[resource] = self._redis.get(key_reserved)
self._logger.debug(
"\tresource %s - reserved: %s - free gpus: %d, cpus: %d",
resource, reserved[resource] or "False", available_gpus,
available_cpus)
# Go through the task, find if there are tasks that can be launched and
# queue the best one
best_task_id = None
best_task_priority = -10000
best_task_queued_time = 0
while count > 0:
count -= 1
next_task_id = self._redis.lindex(queue, count)
if next_task_id is not None:
next_keyt = 'task:%s' % next_task_id
# self._logger.debug("\tcheck task: %s", next_task_id)
parent = self._redis.hget(next_keyt, 'parent')
# check parent dependency
if parent:
keyp = 'task:%s' % parent
if self._redis.exists(keyp):
# if the parent task is in the database, check for dependencies
parent_status = self._redis.hget(keyp, 'status')
if parent_status != 'stopped':
if parent_status == 'running':
# parent is still running so update queued time to be as close
# as possible to terminate time of parent task
self._redis.hset(next_keyt, "queued_time",
time.time())
continue
else:
if self._redis.hget(keyp,
'message') != 'completed':
task.terminate(self._redis,
next_task_id,
phase='dependency_error')
continue
ngpus = int(self._redis.hget(next_keyt, 'ngpus'))
ncpus = int(self._redis.hget(next_keyt, 'ncpus'))
foundResource = False
if next_task_id in preallocated_task_count:
# if task is pre-allocated, can only continue on the same node
r = preallocated_task_resource[next_task_id]
ngpus -= preallocated_task_count[next_task_id]
avail_r = avail_resource[r]
foundResource = (ngpus == 0 or avail_r[1] != 0) and (
ngpus != 0 or ncpus <= avail_r[1])
else:
# can the task be launched on any node
for r, v in six.iteritems(avail_resource):
# cannot launch a new task on a reserved node
if reserved[r]:
continue
if ((ngpus > 0 and resources[r] >= ngpus
and v[1] > 0)
or (ngpus == 0 and v[0] >= ncpus)):
foundResource = True
break
if not foundResource:
continue
priority = int(self._redis.hget(next_keyt, 'priority'))
queued_time = float(
self._redis.hget(next_keyt, 'queued_time'))
if priority > best_task_priority or (
priority == best_task_priority
and best_task_queued_time > queued_time):
best_task_priority = priority
best_task_id = next_task_id
best_task_queued_time = queued_time
if best_task_id:
self._logger.info('selected %s to be launched on %s',
best_task_id, service.name)
task.work_queue(self._redis, best_task_id, service.name)
self._redis.lrem(queue, 0, best_task_id)
def _advance_task(self, task_id):
"""Tries to advance the task to the next status. If it can, re-queue it immediately
to process the next stage. Otherwise, re-queue it after some delay to try again.
"""
keyt = 'task:%s' % task_id
with self._redis.acquire_lock(keyt, acquire_timeout=1,
expire_time=600):
status = self._redis.hget(keyt, 'status')
if status == 'stopped':
return
service_name = self._redis.hget(keyt, 'service')
if service_name not in self._services:
raise ValueError('unknown service %s' % service_name)
service = self._services[service_name]
self._logger.info('%s: trying to advance from status %s', task_id,
status)
if status == 'queued':
resource = self._redis.hget(keyt, 'resource')
resource = self._allocate_resource(task_id, resource, service)
if resource is not None:
self._logger.info('%s: resource %s reserved', task_id,
resource)
self._redis.hset(keyt, 'resource', resource)
task.set_status(self._redis, keyt, 'allocated')
task.queue(self._redis, task_id)
else:
self._logger.warning('%s: no resources available, waiting',
task_id)
self._wait_for_resource(service, task_id)
elif status == 'allocated':
content = json.loads(self._redis.hget(keyt, 'content'))
resource = self._redis.hget(keyt, 'resource')
self._logger.info('%s: launching on %s', task_id, service.name)
data = service.launch(task_id, content['options'], resource,
content['docker']['registry'],
content['docker']['image'],
content['docker']['tag'],
content['docker']['command'],
task.file_list(self._redis, task_id),
content['wait_after_launch'])
self._logger.info('%s: task started on %s', task_id,
service.name)
self._redis.hset(keyt, 'job', json.dumps(data))
task.set_status(self._redis, keyt, 'running')
# For services that do not notify their activity, we should
# poll the task status more regularly.
task.queue(self._redis,
task_id,
delay=service.is_notifying_activity and 120 or 30)
elif status == 'running':
data = json.loads(self._redis.hget(keyt, 'job'))
status = service.status(data)
if status == 'dead':
self._logger.info(
'%s: task no longer running on %s, request termination',
task_id, service.name)
task.terminate(self._redis, task_id, phase='exited')
else:
task.queue(self._redis,
task_id,
delay=service.is_notifying_activity and 120
or 30)
elif status == 'terminating':
data = self._redis.hget(keyt, 'job')
if data is not None:
data = json.loads(data)
self._logger.info('%s: terminating task', task_id)
try:
service.terminate(data)
self._logger.info('%s: terminated', task_id)
except Exception:
self._logger.warning('%s: failed to terminate',
task_id)
resource = self._redis.hget(keyt, 'resource')
self._release_resource(service, resource, task_id)
task.set_status(self._redis, keyt, 'stopped')
task.disable(self._redis, task_id)
