def load_job(path, nodes, context=None):
if not os.path.exists(path):
raise ValueError('Job definition does not exist.')
job = import_job(path)
job_name = job.name.replace(' ', '_')
if job.debug:
job_name += '_debug'
holder = os.path.join(root_dir(), 'data', 'master', 'jobs', job_name)
if not os.path.exists(holder):
os.makedirs(holder)
lock_f = os.path.join(holder, 'lock')
if os.path.exists(lock_f):
raise JobMasterRunning('There has been a running job master')
open(lock_f, 'w').close()
rpc_server = create_rpc_server(job)
try:
loader = JobLoader(job, nodes, rpc_server, context=context)
loader.run()
# nofify master watcher finishing
master_watcher = '%s:%s' % (get_ip(), main_conf.master.port)
client_call(master_watcher, 'finish_job', job.real_name)
finally:
os.remove(lock_f)
rpc_server.shutdown()
def _check_workers(self):
while not self.stopped.is_set():
for worker, info in self.worker_tracker.workers.iteritems():
# if loose connection
if int(time.time()) - info.last_update \
> HEARTBEAT_CHECK_INTERVAL:
info.continous_register = 0
if info.status == RUNNING:
info.status = HANGUP
elif info.status == HANGUP:
info.status = STOPPED
self.black_list.append(worker)
for job in self.job_tracker.running_jobs:
self.job_tracker.remove_worker(job, worker)
# if continously connect for more than 10 min
elif info.continous_register >= CONTINOUS_HEARTBEAT:
if info.status != RUNNING:
info.status = RUNNING
if worker in self.black_list:
self.black_list.remove(worker)
for job in self.job_tracker.running_jobs:
if not client_call(worker, 'has_job'):
client_call(worker, 'prepare', job)
client_call(worker, 'run_job', job)
self.job_tracker.add_worker(job, worker)
self.stopped.wait(HEARTBEAT_CHECK_INTERVAL)
def stop(self):
for node in self.nodes:
try:
client_call(node, 'stop')
except socket.error:
pass
self.stopped = True
def add_node(self, node):
for node in self.nodes:
self.logger.debug('add node %s to nodes' % node)
client_call(node, 'add_node', node, ignore=True)
self.nodes.append(node)
self.logger.debug('run new node %s ' % node)
client_call(node, 'run', ignore=True)
def load_job(path, nodes, context=None):
if not os.path.exists(path):
raise ValueError('Job definition does not exist.')
job = import_job(path)
job_name = job.name.replace(' ', '_')
if job.debug:
job_name += '_debug'
holder = os.path.join(root_dir(), 'data', 'master', 'jobs', job_name)
if not os.path.exists(holder):
os.makedirs(holder)
lock_f = os.path.join(holder, 'lock')
if os.path.exists(lock_f):
raise JobMasterRunning('There has been a running job master')
open(lock_f, 'w').close()
rpc_server = create_rpc_server(job)
try:
loader = JobLoader(job, nodes, rpc_server, context=context)
loader.run()
# nofify master watcher finishing
master_watcher = '%s:%s' % (get_ip(), main_conf.master.port)
client_call(master_watcher, 'finish_job', job.real_name)
finally:
os.remove(lock_f)
rpc_server.shutdown()
def stop(self):
for node in self.nodes:
try:
client_call(node, 'stop')
except socket.error:
pass
self.finish()
def _check_workers(self):
while not self.stopped.is_set():
for worker, info in self.worker_tracker.workers.iteritems():
# if loose connection
if int(time.time()) - info.last_update \
> HEARTBEAT_CHECK_INTERVAL:
info.continous_register = 0
if info.status == RUNNING:
info.status = HANGUP
elif info.status == HANGUP:
info.status = STOPPED
self.black_list.append(worker)
for job in self.job_tracker.running_jobs:
self.job_tracker.remove_worker(job, worker)
# if continously connect for more than 10 min
elif info.continous_register >= CONTINOUS_HEARTBEAT:
if info.status != RUNNING:
info.status = RUNNING
if worker in self.black_list:
self.black_list.remove(worker)
for job in self.job_tracker.running_jobs:
if not client_call(worker, 'has_job'):
client_call(worker, 'prepare', job)
client_call(worker, 'run_job', job)
self.job_tracker.add_worker(job, worker)
self.stopped.wait(HEARTBEAT_CHECK_INTERVAL)
def _remote_or_local_batch_put(self, addr, objs):
if self._check_empty(objs):
return
if addr == self.addr_:
self.mq_node.batch_put(objs)
else:
client_call(addr, self.prefix+'batch_put', pickle.dumps(objs))
def clear_job(self, job_name):
job_name = job_name.replace(' ', '_')
path = os.path.join(self.job_dir, job_name)
shutil.rmtree(path)
for watcher in self.nodes_watchers:
client_call(watcher, 'clear_job')
def finish(self):
all_pages = self.pages()
self.release_lock(self.ready_lock)
self.release_lock(self.finish_lock)
LimitionJobLoader.finish(self)
JobLoader.finish(self)
self.stop_logger_server()
try:
for handler in self.logger.handlers:
handler.close()
except:
pass
if self.client is not None:
rpc_client = '%s:%s' % (
self.client.split(':')[0],
main_conf.client.port
)
client_call(rpc_client, 'stop', ignore=True)
self.logger.info('All nodes finishes visiting pages size: %s' % all_pages)
self.stopped = True
def stop(self):
# stop all jobs
for job_name in self.running_jobs.keys():
self.stop_job(job_name)
for watcher in self.nodes_watchers:
client_call(watcher, 'stop')
self.finish()
def add_worker(self, worker):
if worker in self.workers:
return
# rpc call the other workers to add this worker
for node in self.workers:
client_call(node, 'add_node', worker)
self.workers.append(worker)
def _remote_or_local_put(self, addr, objs, force=False, priority=0):
if self._check_empty(objs):
return
if addr == self.addr_:
self.mq_node.put(objs, force=force, priority=priority)
else:
client_call(addr, self.prefix+'put', pickle.dumps(objs),
force, priority)
def stop(self):
# stop all jobs
for job_name in self.running_jobs.keys():
self.stop_job(job_name)
for watcher in self.nodes_watchers:
client_call(watcher, 'stop', ignore=True)
self.finish()
def remove_worker(self, worker):
if worker not in self.workers:
return
# rpc call the other workers to remove this worker
self.workers.remove(worker)
for node in self.workers:
client_call(node, 'remove_node', worker)
def add_worker(self, worker):
if worker in self.workers:
return
# rpc call the other workers to add this worker
for node in self.workers:
client_call(node, 'add_node', worker)
self.workers.append(worker)
def remove_worker(self, worker):
if worker not in self.workers:
return
# rpc call the other workers to remove this worker
self.workers.remove(worker)
for node in self.workers:
client_call(node, 'remove_node', worker)
def _remote_or_local_put_backup(self, addr, backup_addr, objs,
force=False):
if self._check_empty(objs):
return
if addr == self.addr_:
self.mq_node.put_backup(backup_addr, objs, force=force)
else:
client_call(addr, self.prefix+'put_backup', backup_addr,
pickle.dumps(objs), force)
def put(self, objs):
addrs_objs, addrs_backup_objs = \
self.distributors.distribute(objs)
for addr, objs in addrs_objs.iteritems():
client_call(addr, self.prefix + 'batch_put', pickle.dumps(objs))
for addr, m in addrs_backup_objs.iteritems():
for b_addr, objs in m.iteritems():
client_call(addr, self.prefix + 'put_backup', b_addr,
pickle.dumps(objs))
def stop(self):
for watcher in self.nodes_watchers:
client_call(watcher, 'stop')
# stop all jobs
for job_info in self.running_jobs.values():
try:
client_call(job_info.job_master, 'stop')
except socket.error:
pass
self.stopped = True
def put(self, objs):
addrs_objs, addrs_backup_objs = \
self.distributors.distribute(objs)
for addr, objs in addrs_objs.iteritems():
client_call(addr, self.prefix+'batch_put', pickle.dumps(objs))
for addr, m in addrs_backup_objs.iteritems():
for b_addr, objs in m.iteritems():
client_call(addr, self.prefix+'put_backup', b_addr,
pickle.dumps(objs))
def action(self, name):
if name == 'stop all':
print 'Trying to stop master and all workers.'
try:
client_call(self.master, 'stop')
except socket.error:
print 'Cannot connect to cola master.'
else:
print 'Cola cluster has been shutdown.'
elif name == 'list jobs':
print 'Running jobs: '
for job in client_call(self.master, 'list_jobs'):
print job
elif name == 'list workers':
print 'Cola workers: '
for worker in client_call(self.master, 'list_workers'):
print worker
elif name == 'list job dirs':
print 'Runnable job dirs: '
for dir_ in client_call(self.master, 'list_job_dirs'):
print dir_
elif name.startswith('run remote job '):
print 'Remote job will run in background.'
job_dir = name[len('run remote job '):]
if job_dir not in client_call(self.master, 'list_job_dirs'):
print 'Remote job dir not exists!'
else:
client_call(self.master, 'start_job', job_dir, False)
elif name.startswith('run local job '):
print 'Job has been committed and will run in background.'
start = len('run local job ')
path = name[start:].strip().strip('"').strip("'")
if not os.path.exists(path):
print 'Job path not exists!'
else:
try:
job = import_job(path)
except (ImportError, AttributeError):
print 'Job path is illegal!'
return
dir_ = tempfile.mkdtemp()
try:
zip_filename = os.path.split(path)[1].replace(' ', '_') + '.zip'
zip_file = os.path.join(dir_, zip_filename)
ZipHandler.compress(zip_file, path, type_filters=("pyc", ))
FileTransportClient(self.master, zip_file).send_file()
client_call(self.master, 'start_job', zip_filename)
finally:
shutil.rmtree(dir_)
def sync(self):
with self.lock:
if isinstance(self.server, basestring):
client_call(self.server, self.prefix + 'inc_merge',
self.inc_counter.container)
client_call(self.server, self.prefix + 'acc_merge',
self.acc_counter.container)
else:
self.server.inc_merge(self.inc_counter.container)
self.server.acc_merge(self.acc_counter.container)
self.inc_counter.reset()
self.acc_counter.reset()
def sync(self):
with self.lock:
if isinstance(self.server, basestring):
client_call(self.server, self.prefix+'inc_merge',
self.inc_counter.container)
client_call(self.server, self.prefix+'acc_merge',
self.acc_counter.container)
else:
self.server.inc_merge(self.inc_counter.container)
self.server.acc_merge(self.acc_counter.container)
self.inc_counter.reset()
self.acc_counter.reset()
def run(self):
self.ready_lock.acquire()
if not self.stopped and len(self.not_registered) == 0:
self.mq_client.put(self.job.starts)
for node in self.nodes:
client_call(node, 'run')
self.finish_lock.acquire()
master_watcher = '%s:%s' % (get_ip(), main_conf.master.port)
client_call(master_watcher, 'finish_job', self.job.real_name, ignore=True)
def run(self):
self.ready_lock.acquire()
if not self.stopped and len(self.not_registered) == 0:
self.mq_client.put(self.job.starts)
for node in self.nodes:
client_call(node, 'run')
self.finish_lock.acquire()
master_watcher = '%s:%s' % (get_ip(), main_conf.master.port)
client_call(master_watcher, 'finish_job', self.job.real_name, ignore=True)
def run(self, args):
master_addr = args.master
ctx = Context(is_client=True, master_addr=master_addr)
if args.list is True:
jobs = ctx.list_jobs()
self.logger.info('list jobs at master: %s' % ctx.master_addr)
for job_id, info in jobs.iteritems():
self.logger.info(
'====> job id: %s, job description: %s, status: %s' % \
(job_id, info['name'], info['status']))
if len(jobs) == 0:
self.logger.info('no jobs exist')
elif args.kill is not None:
job_id = self._get_matched_job_name(ctx, args.kill)
if job_id is not None:
ctx.kill_job(job_id)
self.logger.info('killed job: %s' % job_id)
elif args.upload is not None:
path = os.path.abspath(args.upload)
if not os.path.exists(path):
self.logger.error('upload path does not exist')
return
job_id = None
try:
job_id = import_job_desc(path).uniq_name
except Exception, e:
self.logger.exception(e)
self.logger.error('uploading job description failed')
return
new_upload_dir = os.path.join(tempfile.gettempdir(), job_id)
if os.path.exists(new_upload_dir):
shutil.rmtree(new_upload_dir)
shutil.copytree(path, new_upload_dir)
temp_filename = os.path.join(tempfile.gettempdir(),
job_id + '.zip')
ZipHandler.compress(temp_filename,
new_upload_dir,
type_filters=('pyc', ))
try:
FileTransportClient(ctx.master_addr, temp_filename).send_file()
finally:
os.remove(temp_filename)
shutil.rmtree(new_upload_dir)
self.logger.info('upload job <id: %s> finished' % job_id)
if args.run == 'U':
client_call(ctx.master_addr, 'run_job', job_id, True)
self.logger.info('submit job <id: %s> to the cluster' % job_id)
def stop_job(self, job_real_name):
if job_real_name not in self.running_jobs:
return False
job_info = self.running_jobs[job_real_name]
try:
client_call(job_info.job_master, 'stop')
finally:
for watcher in self.nodes_watchers.keys():
client_call(watcher, 'kill', job_real_name)
self.kill(job_real_name)
return True
def stop_job(self, job_real_name):
if job_real_name not in self.running_jobs:
return False
job_info = self.running_jobs[job_real_name]
try:
client_call(job_info.job_master, 'stop', ignore=True)
finally:
for watcher in self.nodes_watchers.keys():
client_call(watcher, 'kill', job_real_name, ignore=True)
self.kill(job_real_name)
return True
def load_job(path, master=None):
if not os.path.exists(path):
raise ValueError('Job definition does not exist.')
job = import_job(path)
holder = os.path.join(root_dir(), 'data', 'worker', 'jobs', job.real_name)
mq_holder = os.path.join(holder, 'mq')
if not os.path.exists(mq_holder):
os.makedirs(mq_holder)
# Logger
logger = get_logger(os.path.join(holder, 'job.log'))
local_node = '%s:%s' % (get_ip(), job.context.job.port)
nodes = [local_node]
if master is not None:
nodes = client_call(master, 'get_nodes')
# Bloom filter hook
bloom_filter_file = os.path.join(holder, 'bloomfilter')
bloom_filter_hook = create_bloom_filter_hook(bloom_filter_file, job)
rpc_server = create_rpc_server(job)
loader = JobLoader(job, rpc_server, logger=logger, master=master)
loader.init_mq(nodes,
local_node,
mq_holder,
verify_exists_hook=bloom_filter_hook,
copies=2 if master else 1)
if master is None:
try:
loader.mq.put(job.starts)
loader.run()
finally:
rpc_server.shutdown()
else:
try:
client_call(master, 'ready', local_node)
def _start():
while not loader.stopped:
time.sleep(TIME_SLEEP)
loader.run()
thread = threading.Thread(target=_start)
thread.start()
thread.join()
finally:
rpc_server.shutdown()
def start_job(self, zip_filename, uncompress=True, client=None):
if uncompress:
zip_file = os.path.join(self.zip_dir, zip_filename)
# transfer zip file to workers
for watcher in self.nodes_watchers:
if watcher.split(':')[0] == self.ip_address:
continue
file_trans_client = FileTransportClient(watcher, zip_file)
file_trans_client.send_file()
job_dir = ZipHandler.uncompress(zip_file, self.job_dir)
else:
job_dir = os.path.join(self.job_dir,
zip_filename.rsplit('.', 1)[0])
job = import_job(job_dir)
worker_port = job.context.job.port
port = job.context.job.master_port
nodes = [watcher.split(':')[0] for watcher in self.nodes_watchers]
if len(nodes) > 0:
info = MasterJobInfo(port, nodes, worker_port)
self.running_jobs[job.real_name] = info
dirname = os.path.dirname(os.path.abspath(__file__))
f = os.path.join(dirname, 'loader.py')
workers = ['%s:%s' % (node, worker_port) for node in nodes]
cmds = [
'python', f, '-j', job_dir, '-i', self.ip_address, '-n',
' '.join(workers)
]
if self.data_path is not None:
cmds.extend(['-d', self.data_path])
if self.force:
cmds.append('-f')
if client is not None:
cmds.extend(['-c', client])
popen = subprocess.Popen(cmds)
info.popen = popen
# call workers to start job
for worker_watcher in self.nodes_watchers:
client_call(worker_watcher,
'start_job',
zip_filename,
uncompress,
ignore=True)
def load_job(path, master=None):
if not os.path.exists(path):
raise ValueError('Job definition does not exist.')
job = import_job(path)
holder = os.path.join(
root_dir(), 'data', 'worker', 'jobs', job.real_name)
mq_holder = os.path.join(holder, 'mq')
if not os.path.exists(mq_holder):
os.makedirs(mq_holder)
# Logger
logger = get_logger(os.path.join(holder, 'job.log'))
local_node = '%s:%s' % (get_ip(), job.context.job.port)
nodes = [local_node]
if master is not None:
nodes = client_call(master, 'get_nodes')
# Bloom filter hook
bloom_filter_file = os.path.join(holder, 'bloomfilter')
bloom_filter_hook = create_bloom_filter_hook(bloom_filter_file, job)
rpc_server = create_rpc_server(job)
loader = JobLoader(job, rpc_server, logger=logger, master=master)
loader.init_mq(nodes, local_node, mq_holder,
verify_exists_hook=bloom_filter_hook,
copies=2 if master else 1)
if master is None:
try:
loader.mq.put(job.starts)
loader.run()
finally:
rpc_server.shutdown()
else:
try:
client_call(master, 'ready', local_node)
def _start():
while not loader.stopped:
time.sleep(TIME_SLEEP)
loader.run()
thread = threading.Thread(target=_start)
thread.start()
thread.join()
finally:
rpc_server.shutdown()
def run(self, args):
master_addr = args.master
ctx = Context(is_client=True, master_addr=master_addr)
if args.list is True:
jobs = ctx.list_jobs()
self.logger.info('list jobs at master: %s' % ctx.master_addr)
for job_id, info in jobs.iteritems():
self.logger.info(
'====> job id: %s, job description: %s, status: %s' % \
(job_id, info['name'], info['status']))
if len(jobs) == 0:
self.logger.info('no jobs exist')
elif args.kill is not None:
job_id = self._get_matched_job_name(ctx, args.kill)
if job_id is not None:
ctx.kill_job(job_id)
self.logger.info('killed job: %s' % job_id)
elif args.upload is not None:
path = os.path.abspath(args.upload)
if not os.path.exists(path):
self.logger.error('upload path does not exist')
return
job_id = None
try:
job_id = import_job_desc(path).uniq_name
except Exception, e:
self.logger.exception(e)
self.logger.error('uploading job description failed')
return
new_upload_dir = os.path.join(tempfile.gettempdir(), job_id)
if os.path.exists(new_upload_dir):
shutil.rmtree(new_upload_dir)
shutil.copytree(path, new_upload_dir)
temp_filename = os.path.join(tempfile.gettempdir(), job_id+'.zip')
ZipHandler.compress(temp_filename, new_upload_dir, type_filters=('pyc', ))
try:
FileTransportClient(ctx.master_addr, temp_filename).send_file()
finally:
os.remove(temp_filename)
shutil.rmtree(new_upload_dir)
self.logger.info('upload job <id: %s> finished' % job_id)
if args.run == 'U':
client_call(ctx.master_addr, 'run_job', job_id, True)
self.logger.info('submit job <id: %s> to the cluster' % job_id)
def get(self, size=1, priority=0):
size = max(size, 1)
addrs = list(self.addrs)
shuffle(addrs)
results = []
for addr in addrs:
left = size - len(results)
if left <= 0:
break
objs = pickle.loads(client_call(addr, self.prefix+'get',
left, priority))
if objs is None:
continue
if not isinstance(objs, list):
objs = [objs, ]
results.extend(objs)
if size == 1:
if len(results) == 0:
return
return results[0]
return results
def get(self, size=1, priority=0):
size = max(size, 1)
addrs = list(self.addrs)
shuffle(addrs)
results = []
for addr in addrs:
left = size - len(results)
if left <= 0:
break
objs = pickle.loads(
client_call(addr, self.prefix + 'get', left, priority))
if objs is None:
continue
if not isinstance(objs, list):
objs = [
objs,
]
results.extend(objs)
if size == 1:
if len(results) == 0:
return
return results[0]
return results
def require(self, size=1):
if isinstance(self.server, basestring):
return client_call(self.server, self.prefix+'require', self.addr,
self.instance_id, size)
else:
return self.server.require(self.addr, self.instance_id,
size=size)
def pages(self):
all_pages = 0
for node in self.nodes:
pages = client_call(node, 'pages', ignore=True)
if pages is not None:
all_pages += int(pages)
return all_pages
def _call(i, worker):
try:
result = client_call(worker, self.remote_func, *args)
except Exception, e:
if self.logger:
self.logger.error(e)
result = False
def _client_call(*args):
try:
return client_call(*args)
except socket.error:
logger.error('Cannot connect to single running worker.')
except:
pass
def _call(i, worker):
try:
result = client_call(worker, self.remote_func, *args)
except Exception, e:
if self.logger:
self.logger.error(e)
result = False
def list_jobs(self):
jobs = {}
if self.is_master and self.master is not None:
runnable_jobs = self.master.list_runnable_jobs()
running_jobs = self.master.job_tracker.running_jobs
else:
runnable_jobs = client_call(self.master_addr, 'runnable_jobs')
running_jobs = client_call(self.master_addr, 'running_jobs')
for job_id, job_name in runnable_jobs.iteritems():
jobs[job_id] = {'name': job_name}
if job_id in running_jobs:
jobs[job_id]['status'] = 'running'
else:
jobs[job_id]['status'] = 'stopped'
return jobs
def list_jobs(self):
jobs = {}
if self.is_master and self.master is not None:
runnable_jobs = self.master.list_runnable_jobs()
running_jobs = self.master.job_tracker.running_jobs
else:
runnable_jobs = client_call(self.master_addr, 'runnable_jobs')
running_jobs = client_call(self.master_addr, 'running_jobs')
for job_id, job_name in runnable_jobs.iteritems():
jobs[job_id] = {'name': job_name}
if job_id in running_jobs:
jobs[job_id]['status'] = 'running'
else:
jobs[job_id]['status'] = 'stopped'
return jobs
def _client_call(*args):
try:
return client_call(*args)
except socket.error:
logger.error('Cannot connect to single running worker.')
except:
pass
def pages(self):
all_pages = 0
for node in self.nodes:
self.logger.debug('get pages from node %s' % node)
pages = client_call(node, 'pages', ignore=True)
if pages is not None:
all_pages += int(pages)
return all_pages
def _report():
while not self.stopped.is_set():
workers = client_call(self.master, 'register_heartbeat',
self.ctx.worker_addr)
self.ctx.addrs = [self.ctx.fix_addr(worker) for worker in workers]
self.ctx.ips = [self.ctx.fix_ip(worker) for worker in workers]
self.stopped.wait(HEARTBEAT_INTERVAL)
def run(self):
# wait until all the workers initialized
while not self.is_ready: pass
if self.limit_speed:
self._in_minute_clear()
self.mq_client.put(self.job.starts)
for node in self.nodes:
client_call(node, 'run')
def _run():
while not self.stopped:
time.sleep(TIME_SLEEP)
main_thread = threading.Thread(target=_run)
main_thread.start()
main_thread.join()
def run(self):
# wait until all the workers initialized
while not self.is_ready:
pass
if self.limit_speed:
self._in_minute_clear()
self.mq_client.put(self.job.starts)
for node in self.nodes:
client_call(node, 'run')
def _run():
while not self.stopped:
time.sleep(TIME_SLEEP)
main_thread = threading.Thread(target=_run)
main_thread.start()
main_thread.join()
def get_job_counter(self, job_id):
if self.is_master and self.master is not None:
return self.master.counter_server.output()
else:
from cola.functions.counter import FUNC_PREFIX
from cola.core.utils import get_rpc_prefix
func_name = '%s%s' % (get_rpc_prefix(job_id, FUNC_PREFIX), 'get_global')
return client_call(self.master_addr, func_name)
def get_job_counter(self, job_id):
if self.is_master and self.master is not None:
return self.master.counter_server.output()
else:
from cola.functions.counter import FUNC_PREFIX
from cola.core.utils import get_rpc_prefix
func_name = '%s%s' % (get_rpc_prefix(job_id, FUNC_PREFIX), 'get_global')
return client_call(self.master_addr, func_name)
def _require_budget(self):
if self.master is None or self.ctx.job.limits == 0:
return
if self.budget > 0:
self.budget -= 1
return
while self.budget == 0 and not self.stopped:
self.budget = client_call(self.master, 'require', BUDGET_REQUIRE)
def finish(self):
self.release_lock(self.ready_lock)
self.release_lock(self.finish_lock)
LimitionJobLoader.finish(self)
JobLoader.finish(self)
self.stop_logger_server()
try:
for handler in self.logger.handlers:
handler.close()
except:
pass
if self.client is not None:
rpc_client = '%s:%s' % (self.client.split(':')[0],
main_conf.client.port)
client_call(rpc_client, 'stop', ignore=True)
self.stopped = True
def load_job(job_path, data_path=None, master=None, force=False):
if not os.path.exists(job_path):
raise ValueError('Job definition does not exist.')
job = import_job(job_path)
if data_path is None:
data_path = os.path.join(root_dir(), 'data')
root = os.path.join(data_path, 'worker', 'jobs', job.real_name)
if not os.path.exists(root):
os.makedirs(root)
if master is None:
with StandaloneWorkerJobLoader(job, root, force=force) as job_loader:
job_loader.run()
else:
nodes = client_call(master, 'get_nodes')
local = '%s:%s' % (get_ip(), job.context.job.port)
client_call(master, 'ready', local)
with WorkerJobLoader(job, root, master, local=local, nodes=nodes, force=force) \
as job_loader:
client_call(master, 'ready', local)
job_loader.ready_for_run()
def start_job(self, zip_filename, uncompress=True):
if uncompress:
zip_file = os.path.join(self.zip_dir, zip_filename)
# transfer zip file to workers
for watcher in self.nodes_watchers:
if watcher.split(':')[0] == self.ip_address:
continue
file_trans_client = FileTransportClient(watcher, zip_file)
file_trans_client.send_file()
job_dir = ZipHandler.uncompress(zip_file, self.job_dir)
else:
job_dir = os.path.join(self.job_dir, zip_filename.rsplit('.', 1)[0])
job = import_job(job_dir)
worker_port = job.context.job.port
port = job.context.job.master_port
nodes = [watcher.split(':')[0] for watcher in self.nodes_watchers]
if len(nodes) > 0:
info = MasterJobInfo(port, nodes, worker_port)
self.running_jobs[job.real_name] = info
dirname = os.path.dirname(os.path.abspath(__file__))
f = os.path.join(dirname, 'loader.py')
workers = ['%s:%s'%(node, worker_port) for node in nodes]
subprocess.Popen('python "%(py)s" "%(job_dir)s" %(nodes)s' % {
'py': f,
'job_dir': job_dir,
'nodes': ' '.join(workers)
})
# call workers to start job
for worker_watcher in self.nodes_watchers:
client_call(worker_watcher, 'start_job', zip_filename, uncompress)
def _remote_or_local_get(self, addr, size=1, priority=0):
objs = None
if addr == self.addr_:
objs = self.mq_node.get(size=size, priority=priority)
else:
objs = pickle.loads(client_call(addr, self.prefix+'get',
size, priority))
addr_caches = self.caches.get(addr, [])
if size == 1 and objs is None and len(addr_caches) > 0:
return addr_caches.pop(0)
elif size > 1 and len(objs) == 0 and len(addr_caches) > 0:
return addr_caches[:size]
return objs