Mentor's intention is to simplify the process of writing python frameworks for Mesos. Mentor provides multiple components and interfaces to cover various levels of complexity needs.
pip install mentor or use daskos/mentor Docker image
Requirements:
Configuration:
MESOS_MASTER=zk://127.0.0.1:2181/mesosIt's almost identical to python's futures interface but runs processes on a Mesos cluster (concurrently).
from mentor.apis.futures import MesosPoolExecutor
from mentor.proxies.messages import Cpus, Mem
with MesosPoolExecutor(name='futures-pool') as executor:
def mul(a, b):
return a * b
future = executor.submit(mul, args=[3, 5])
assert future.result(timeout=5) == 3
it = executor.map(mul, range(10), range(10), timeout=5,
resources=[Cpus(0.1), Mem(128)])
assert list(it) == [i**2 for i in range(10)]It's similar to python's multiprocessing interface but runs processes on a Mesos cluster (concurrently).
from __future__ import print_function
from mentor.apis.multiprocessing import Pool
with Pool(name='mentor-pool') as pool:
def mul(a, b):
return a * b
res1 = pool.apply_async(lambda a, b: a + b, [1, 2])
res2 = pool.apply_async(mul, [2, 3])
pool.wait()
print(res1.get())
print(res2.get())Basic scheduler to submit various kind of workloads, eg.:
from __future__ import print_function
from mentor.scheduler import QueueScheduler, Running
from mentor.messages import PythonTask
from mentor.proxies.messages import Disk, Mem, Cpus
scheduler = QueueScheduler()
task = PythonTask(fn=sum, args=[range(10)], name='mentor-task',
resources=[Cpus(0.1), Mem(128), Disk(512)])
with Running(scheduler, name='mentor-scheduler'):
res = scheduler.submit(task) # return AsyncResult
print(res.get(timeout=30))You can make your own scheduler built on QueueScheduler or for more complex needs there's a Scheduler interface which you can use to create one from scratch. (However in this case you'll have to implement some of the functionalities already in QueueScheduler)
from __future__ import print_function
from mentor.scheduler import QueueScheduler, Running
from mentor.messages import PythonTask
from mentor.proxies.messages import Disk, Mem, Cpus
class CustomScheduler(QueueScheduler):
def on_update(self, driver, status):
"""You can hook on the events defined in the Scheduler interface.
They're just more conveniantly named methods for the basic
mesos.interface functions but this is how you can add some
custom logic to your framework in an easy manner.
"""
logging.info(
"Status update received for task {}".format(status.task_id))
super(CustomScheduler, self).on_update(driver, status)
scheduler = CustomScheduler()
task = PythonTask(fn=sum, args=[range(9)], name='mentor-task',
resources=[Cpus(0.1), Mem(128), Disk(512)])
with Running(scheduler, name='mentor-custom-scheduler'):
res = scheduler.submit(task)
print(res.get(timeout=60))Also this way you can easily implement your own resource offer handling logic by
overriding the on_offers(self, driver, offers) method in which we give you a
helping hand with comparable Offers and TaskInfos (basic arithmetic operators
are also overloaded).
from mentor.interface import Scheduler
from mentor.proxies.messages import Offer, TaskInfo
class CustomScheduler(Scheduler):
...
def on_offers(self, driver, offers):
...
task = self.get_next_task()
for offer in offers
if task < offer:
task.slave_id = offer.slave_id
driver.launch(offer, [task])
# decline unused offers or launch with empty task list
...Mentor implements multiple weighted heuristics to solve the Bin-Packing Problem:
see binpack.py.
The benefits of using bin-packing has been proven by Netflix/Fenzo in Heterogeneous Resource Scheduling Using Apache Mesos
The most basic task executes a simple command, Mesos will run CommandInfo's
value with /bin/sh -c. Also, if you want to run your task in a Docker
container you can provide some additional information for the task.
from mentor.proxies.messages import TaskInfo, CommandInfo
task = TaskInfo(name='command-task', command=CommandInfo(value='echo 100'))
task.container.type = 'DOCKER'
task.container.docker.image = 'daskos/mentor:latest'PythonTask is capable of running arbitrary python code on your cluster. It sends cloudpickled methods and arguments to the matched mesos-slave for execution. Note that python tasks run in daskos/mentor Docker container by default.
from mentor.messages import PythonTask
# You can pass a function or a lambda in place of sum for fn.
task = PythonTask(name='python-task', fn=sum, args=[range(5)])Customs tasks can be written by extending TaskInfo or any existing descendants. If you're walking down the former path you'll most likely have to deal with protobuf in your code; worry not, we have some magic wrappers for you to provide customizable messages.
from __future__ import print_function
from mentor.proxies.messages import TaskInfo
from mesos.interface import mesos_pb2
class CustomTask(TaskInfo):
# descriptive protobuf template the wrapper matched against
proto = mesos_pb2.TaskInfo(
labels=mesos_pb2.Labels(
labels=[mesos_pb2.Label(key='custom')]))
@property
def uppercase_task_name():
return self.name.upper()
def on_update(self, status):
logging.info('Custom task has received a status update')
def custom_method(self):
print("Arbitrary stuff")This Executor implementation simply runs the received python function with the provided arguments, then sends back the result in a reliable fashion.
class OneOffExecutor(Executor):
def on_launch(self, driver, task):
def run_task():
driver.update(task.status('TASK_RUNNING'))
logging.info('Sent TASK_RUNNING status update')
try:
logging.info('Executing task...')
result = task()
except Exception as e:
logging.exception('Task errored')
driver.update(task.status('TASK_FAILED', message=e.message))
logging.info('Sent TASK_RUNNING status update')
else:
driver.update(task.status('TASK_FINISHED', data=result))
logging.info('Sent TASK_FINISHED status update')
thread = threading.Thread(target=run_task)
thread.start()This is a pre-release!
are coming!