##// END OF EJS Templates
Merging from trunk.
Merging from trunk.

File last commit:

r1234:52b55407
r1268:bb43f0d6 merge
Show More
task.py
799 lines | 27.6 KiB | text/x-python | PythonLexer
Brian E Granger
This is a manual merge of certain things in the ipython1-dev branch, revision 46, into the main ...
r1234 # encoding: utf-8
# -*- test-case-name: IPython.kernel.tests.test_task -*-
"""Task farming representation of the ControllerService."""
__docformat__ = "restructuredtext en"
#-------------------------------------------------------------------------------
# Copyright (C) 2008 The IPython Development Team
#
# Distributed under the terms of the BSD License. The full license is in
# the file COPYING, distributed as part of this software.
#-------------------------------------------------------------------------------
#-------------------------------------------------------------------------------
# Imports
#-------------------------------------------------------------------------------
import copy, time
from types import FunctionType as function
import zope.interface as zi, string
from twisted.internet import defer, reactor
from twisted.python import components, log, failure
# from IPython.genutils import time
from IPython.kernel import engineservice as es, error
from IPython.kernel import controllerservice as cs
from IPython.kernel.twistedutil import gatherBoth, DeferredList
from IPython.kernel.pickleutil import can,uncan, CannedFunction
def canTask(task):
t = copy.copy(task)
t.depend = can(t.depend)
if t.recovery_task:
t.recovery_task = canTask(t.recovery_task)
return t
def uncanTask(task):
t = copy.copy(task)
t.depend = uncan(t.depend)
if t.recovery_task and t.recovery_task is not task:
t.recovery_task = uncanTask(t.recovery_task)
return t
time_format = '%Y/%m/%d %H:%M:%S'
class Task(object):
"""Our representation of a task for the `TaskController` interface.
The user should create instances of this class to represent a task that
needs to be done.
:Parameters:
expression : str
A str that is valid python code that is the task.
pull : str or list of str
The names of objects to be pulled as results. If not specified,
will return {'result', None}
push : dict
A dict of objects to be pushed into the engines namespace before
execution of the expression.
clear_before : boolean
Should the engine's namespace be cleared before the task is run.
Default=False.
clear_after : boolean
Should the engine's namespace be cleared after the task is run.
Default=False.
retries : int
The number of times to resumbit the task if it fails. Default=0.
recovery_task : Task
This is the Task to be run when the task has exhausted its retries
Default=None.
depend : bool function(properties)
This is the dependency function for the Task, which determines
whether a task can be run on a Worker. `depend` is called with
one argument, the worker's properties dict, and should return
True if the worker meets the dependencies or False if it does
not.
Default=None - run on any worker
options : dict
Any other keyword options for more elaborate uses of tasks
Examples
--------
>>> t = Task('dostuff(args)')
>>> t = Task('a=5', pull='a')
>>> t = Task('a=5\nb=4', pull=['a','b'])
>>> t = Task('os.kill(os.getpid(),9)', retries=100) # this is a bad idea
# A dependency case:
>>> def hasMPI(props):
... return props.get('mpi') is not None
>>> t = Task('mpi.send(blah,blah)', depend = hasMPI)
"""
def __init__(self, expression, pull=None, push=None,
clear_before=False, clear_after=False, retries=0,
recovery_task=None, depend=None, **options):
self.expression = expression
if isinstance(pull, str):
self.pull = [pull]
else:
self.pull = pull
self.push = push
self.clear_before = clear_before
self.clear_after = clear_after
self.retries=retries
self.recovery_task = recovery_task
self.depend = depend
self.options = options
self.taskid = None
class ResultNS:
"""The result namespace object for use in TaskResult objects as tr.ns.
It builds an object from a dictionary, such that it has attributes
according to the key,value pairs of the dictionary.
This works by calling setattr on ALL key,value pairs in the dict. If a user
chooses to overwrite the `__repr__` or `__getattr__` attributes, they can.
This can be a bad idea, as it may corrupt standard behavior of the
ns object.
Example
--------
>>> ns = ResultNS({'a':17,'foo':range(3)})
>>> print ns
NS{'a':17,'foo':range(3)}
>>> ns.a
17
>>> ns['foo']
[0,1,2]
"""
def __init__(self, dikt):
for k,v in dikt.iteritems():
setattr(self,k,v)
def __repr__(self):
l = dir(self)
d = {}
for k in l:
# do not print private objects
if k[:2] != '__' and k[-2:] != '__':
d[k] = getattr(self, k)
return "NS"+repr(d)
def __getitem__(self, key):
return getattr(self, key)
class TaskResult(object):
"""
An object for returning task results.
This object encapsulates the results of a task. On task
success it will have a keys attribute that will have a list
of the variables that have been pulled back. These variables
are accessible as attributes of this class as well. On
success the failure attribute will be None.
In task failure, keys will be empty, but failure will contain
the failure object that encapsulates the remote exception.
One can also simply call the raiseException() method of
this class to re-raise any remote exception in the local
session.
The TaskResult has a .ns member, which is a property for access
to the results. If the Task had pull=['a', 'b'], then the
Task Result will have attributes tr.ns.a, tr.ns.b for those values.
Accessing tr.ns will raise the remote failure if the task failed.
The engineid attribute should have the engineid of the engine
that ran the task. But, because engines can come and go in
the ipython task system, the engineid may not continue to be
valid or accurate.
The taskid attribute simply gives the taskid that the task
is tracked under.
"""
taskid = None
def _getNS(self):
if isinstance(self.failure, failure.Failure):
return self.failure.raiseException()
else:
return self._ns
def _setNS(self, v):
raise Exception("I am protected!")
ns = property(_getNS, _setNS)
def __init__(self, results, engineid):
self.engineid = engineid
if isinstance(results, failure.Failure):
self.failure = results
self.results = {}
else:
self.results = results
self.failure = None
self._ns = ResultNS(self.results)
self.keys = self.results.keys()
def __repr__(self):
if self.failure is not None:
contents = self.failure
else:
contents = self.results
return "TaskResult[ID:%r]:%r"%(self.taskid, contents)
def __getitem__(self, key):
if self.failure is not None:
self.raiseException()
return self.results[key]
def raiseException(self):
"""Re-raise any remote exceptions in the local python session."""
if self.failure is not None:
self.failure.raiseException()
class IWorker(zi.Interface):
"""The Basic Worker Interface.
A worked is a representation of an Engine that is ready to run tasks.
"""
zi.Attribute("workerid", "the id of the worker")
def run(task):
"""Run task in worker's namespace.
:Parameters:
task : a `Task` object
:Returns: `Deferred` to a `TaskResult` object.
"""
class WorkerFromQueuedEngine(object):
"""Adapt an `IQueuedEngine` to an `IWorker` object"""
zi.implements(IWorker)
def __init__(self, qe):
self.queuedEngine = qe
self.workerid = None
def _get_properties(self):
return self.queuedEngine.properties
properties = property(_get_properties, lambda self, _:None)
def run(self, task):
"""Run task in worker's namespace.
:Parameters:
task : a `Task` object
:Returns: `Deferred` to a `TaskResult` object.
"""
if task.clear_before:
d = self.queuedEngine.reset()
else:
d = defer.succeed(None)
if task.push is not None:
d.addCallback(lambda r: self.queuedEngine.push(task.push))
d.addCallback(lambda r: self.queuedEngine.execute(task.expression))
if task.pull is not None:
d.addCallback(lambda r: self.queuedEngine.pull(task.pull))
else:
d.addCallback(lambda r: None)
def reseter(result):
self.queuedEngine.reset()
return result
if task.clear_after:
d.addBoth(reseter)
return d.addBoth(self._zipResults, task.pull, time.time(), time.localtime())
def _zipResults(self, result, names, start, start_struct):
"""Callback for construting the TaskResult object."""
if isinstance(result, failure.Failure):
tr = TaskResult(result, self.queuedEngine.id)
else:
if names is None:
resultDict = {}
elif len(names) == 1:
resultDict = {names[0]:result}
else:
resultDict = dict(zip(names, result))
tr = TaskResult(resultDict, self.queuedEngine.id)
# the time info
tr.submitted = time.strftime(time_format, start_struct)
tr.completed = time.strftime(time_format)
tr.duration = time.time()-start
return tr
components.registerAdapter(WorkerFromQueuedEngine, es.IEngineQueued, IWorker)
class IScheduler(zi.Interface):
"""The interface for a Scheduler.
"""
zi.Attribute("nworkers", "the number of unassigned workers")
zi.Attribute("ntasks", "the number of unscheduled tasks")
zi.Attribute("workerids", "a list of the worker ids")
zi.Attribute("taskids", "a list of the task ids")
def add_task(task, **flags):
"""Add a task to the queue of the Scheduler.
:Parameters:
task : a `Task` object
The task to be queued.
flags : dict
General keywords for more sophisticated scheduling
"""
def pop_task(id=None):
"""Pops a Task object.
This gets the next task to be run. If no `id` is requested, the highest priority
task is returned.
:Parameters:
id
The id of the task to be popped. The default (None) is to return
the highest priority task.
:Returns: a `Task` object
:Exceptions:
IndexError : raised if no taskid in queue
"""
def add_worker(worker, **flags):
"""Add a worker to the worker queue.
:Parameters:
worker : an IWorker implementing object
flags : General keywords for more sophisticated scheduling
"""
def pop_worker(id=None):
"""Pops an IWorker object that is ready to do work.
This gets the next IWorker that is ready to do work.
:Parameters:
id : if specified, will pop worker with workerid=id, else pops
highest priority worker. Defaults to None.
:Returns:
an IWorker object
:Exceptions:
IndexError : raised if no workerid in queue
"""
def ready():
"""Returns True if there is something to do, False otherwise"""
def schedule():
"""Returns a tuple of the worker and task pair for the next
task to be run.
"""
class FIFOScheduler(object):
"""A basic First-In-First-Out (Queue) Scheduler.
This is the default Scheduler for the TaskController.
See the docstrings for IScheduler for interface details.
"""
zi.implements(IScheduler)
def __init__(self):
self.tasks = []
self.workers = []
def _ntasks(self):
return len(self.tasks)
def _nworkers(self):
return len(self.workers)
ntasks = property(_ntasks, lambda self, _:None)
nworkers = property(_nworkers, lambda self, _:None)
def _taskids(self):
return [t.taskid for t in self.tasks]
def _workerids(self):
return [w.workerid for w in self.workers]
taskids = property(_taskids, lambda self,_:None)
workerids = property(_workerids, lambda self,_:None)
def add_task(self, task, **flags):
self.tasks.append(task)
def pop_task(self, id=None):
if id is None:
return self.tasks.pop(0)
else:
for i in range(len(self.tasks)):
taskid = self.tasks[i].taskid
if id == taskid:
return self.tasks.pop(i)
raise IndexError("No task #%i"%id)
def add_worker(self, worker, **flags):
self.workers.append(worker)
def pop_worker(self, id=None):
if id is None:
return self.workers.pop(0)
else:
for i in range(len(self.workers)):
workerid = self.workers[i].workerid
if id == workerid:
return self.workers.pop(i)
raise IndexError("No worker #%i"%id)
def schedule(self):
for t in self.tasks:
for w in self.workers:
try:# do not allow exceptions to break this
cando = t.depend is None or t.depend(w.properties)
except:
cando = False
if cando:
return self.pop_worker(w.workerid), self.pop_task(t.taskid)
return None, None
class LIFOScheduler(FIFOScheduler):
"""A Last-In-First-Out (Stack) Scheduler. This scheduler should naively
reward fast engines by giving them more jobs. This risks starvation, but
only in cases with low load, where starvation does not really matter.
"""
def add_task(self, task, **flags):
# self.tasks.reverse()
self.tasks.insert(0, task)
# self.tasks.reverse()
def add_worker(self, worker, **flags):
# self.workers.reverse()
self.workers.insert(0, worker)
# self.workers.reverse()
class ITaskController(cs.IControllerBase):
"""The Task based interface to a `ControllerService` object
This adapts a `ControllerService` to the ITaskController interface.
"""
def run(task):
"""Run a task.
:Parameters:
task : an IPython `Task` object
:Returns: the integer ID of the task
"""
def get_task_result(taskid, block=False):
"""Get the result of a task by its ID.
:Parameters:
taskid : int
the id of the task whose result is requested
:Returns: `Deferred` to (taskid, actualResult) if the task is done, and None
if not.
:Exceptions:
actualResult will be an `IndexError` if no such task has been submitted
"""
def abort(taskid):
"""Remove task from queue if task is has not been submitted.
If the task has already been submitted, wait for it to finish and discard
results and prevent resubmission.
:Parameters:
taskid : the id of the task to be aborted
:Returns:
`Deferred` to abort attempt completion. Will be None on success.
:Exceptions:
deferred will fail with `IndexError` if no such task has been submitted
or the task has already completed.
"""
def barrier(taskids):
"""Block until the list of taskids are completed.
Returns None on success.
"""
def spin():
"""touch the scheduler, to resume scheduling without submitting
a task.
"""
def queue_status(self, verbose=False):
"""Get a dictionary with the current state of the task queue.
If verbose is True, then return lists of taskids, otherwise,
return the number of tasks with each status.
"""
class TaskController(cs.ControllerAdapterBase):
"""The Task based interface to a Controller object.
If you want to use a different scheduler, just subclass this and set
the `SchedulerClass` member to the *class* of your chosen scheduler.
"""
zi.implements(ITaskController)
SchedulerClass = FIFOScheduler
timeout = 30
def __init__(self, controller):
self.controller = controller
self.controller.on_register_engine_do(self.registerWorker, True)
self.controller.on_unregister_engine_do(self.unregisterWorker, True)
self.taskid = 0
self.failurePenalty = 1 # the time in seconds to penalize
# a worker for failing a task
self.pendingTasks = {} # dict of {workerid:(taskid, task)}
self.deferredResults = {} # dict of {taskid:deferred}
self.finishedResults = {} # dict of {taskid:actualResult}
self.workers = {} # dict of {workerid:worker}
self.abortPending = [] # dict of {taskid:abortDeferred}
self.idleLater = None # delayed call object for timeout
self.scheduler = self.SchedulerClass()
for id in self.controller.engines.keys():
self.workers[id] = IWorker(self.controller.engines[id])
self.workers[id].workerid = id
self.schedule.add_worker(self.workers[id])
def registerWorker(self, id):
"""Called by controller.register_engine."""
if self.workers.get(id):
raise "We already have one! This should not happen."
self.workers[id] = IWorker(self.controller.engines[id])
self.workers[id].workerid = id
if not self.pendingTasks.has_key(id):# if not working
self.scheduler.add_worker(self.workers[id])
self.distributeTasks()
def unregisterWorker(self, id):
"""Called by controller.unregister_engine"""
if self.workers.has_key(id):
try:
self.scheduler.pop_worker(id)
except IndexError:
pass
self.workers.pop(id)
def _pendingTaskIDs(self):
return [t.taskid for t in self.pendingTasks.values()]
#---------------------------------------------------------------------------
# Interface methods
#---------------------------------------------------------------------------
def run(self, task):
"""Run a task and return `Deferred` to its taskid."""
task.taskid = self.taskid
task.start = time.localtime()
self.taskid += 1
d = defer.Deferred()
self.scheduler.add_task(task)
# log.msg('Queuing task: %i' % task.taskid)
self.deferredResults[task.taskid] = []
self.distributeTasks()
return defer.succeed(task.taskid)
def get_task_result(self, taskid, block=False):
"""Returns a `Deferred` to a TaskResult tuple or None."""
# log.msg("Getting task result: %i" % taskid)
if self.finishedResults.has_key(taskid):
tr = self.finishedResults[taskid]
return defer.succeed(tr)
elif self.deferredResults.has_key(taskid):
if block:
d = defer.Deferred()
self.deferredResults[taskid].append(d)
return d
else:
return defer.succeed(None)
else:
return defer.fail(IndexError("task ID not registered: %r" % taskid))
def abort(self, taskid):
"""Remove a task from the queue if it has not been run already."""
if not isinstance(taskid, int):
return defer.fail(failure.Failure(TypeError("an integer task id expected: %r" % taskid)))
try:
self.scheduler.pop_task(taskid)
except IndexError, e:
if taskid in self.finishedResults.keys():
d = defer.fail(IndexError("Task Already Completed"))
elif taskid in self.abortPending:
d = defer.fail(IndexError("Task Already Aborted"))
elif taskid in self._pendingTaskIDs():# task is pending
self.abortPending.append(taskid)
d = defer.succeed(None)
else:
d = defer.fail(e)
else:
d = defer.execute(self._doAbort, taskid)
return d
def barrier(self, taskids):
dList = []
if isinstance(taskids, int):
taskids = [taskids]
for id in taskids:
d = self.get_task_result(id, block=True)
dList.append(d)
d = DeferredList(dList, consumeErrors=1)
d.addCallbacks(lambda r: None)
return d
def spin(self):
return defer.succeed(self.distributeTasks())
def queue_status(self, verbose=False):
pending = self._pendingTaskIDs()
failed = []
succeeded = []
for k,v in self.finishedResults.iteritems():
if not isinstance(v, failure.Failure):
if hasattr(v,'failure'):
if v.failure is None:
succeeded.append(k)
else:
failed.append(k)
scheduled = self.scheduler.taskids
if verbose:
result = dict(pending=pending, failed=failed,
succeeded=succeeded, scheduled=scheduled)
else:
result = dict(pending=len(pending),failed=len(failed),
succeeded=len(succeeded),scheduled=len(scheduled))
return defer.succeed(result)
#---------------------------------------------------------------------------
# Queue methods
#---------------------------------------------------------------------------
def _doAbort(self, taskid):
"""Helper function for aborting a pending task."""
# log.msg("Task aborted: %i" % taskid)
result = failure.Failure(error.TaskAborted())
self._finishTask(taskid, result)
if taskid in self.abortPending:
self.abortPending.remove(taskid)
def _finishTask(self, taskid, result):
dlist = self.deferredResults.pop(taskid)
result.taskid = taskid # The TaskResult should save the taskid
self.finishedResults[taskid] = result
for d in dlist:
d.callback(result)
def distributeTasks(self):
"""Distribute tasks while self.scheduler has things to do."""
# log.msg("distributing Tasks")
worker, task = self.scheduler.schedule()
if not worker and not task:
if self.idleLater and self.idleLater.called:# we are inside failIdle
self.idleLater = None
else:
self.checkIdle()
return False
# else something to do:
while worker and task:
# get worker and task
# add to pending
self.pendingTasks[worker.workerid] = task
# run/link callbacks
d = worker.run(task)
# log.msg("Running task %i on worker %i" %(task.taskid, worker.workerid))
d.addBoth(self.taskCompleted, task.taskid, worker.workerid)
worker, task = self.scheduler.schedule()
# check for idle timeout:
self.checkIdle()
return True
def checkIdle(self):
if self.idleLater and not self.idleLater.called:
self.idleLater.cancel()
if self.scheduler.ntasks and self.workers and \
self.scheduler.nworkers == len(self.workers):
self.idleLater = reactor.callLater(self.timeout, self.failIdle)
else:
self.idleLater = None
def failIdle(self):
if not self.distributeTasks():
while self.scheduler.ntasks:
t = self.scheduler.pop_task()
msg = "task %i failed to execute due to unmet dependencies"%t.taskid
msg += " for %i seconds"%self.timeout
# log.msg("Task aborted by timeout: %i" % t.taskid)
f = failure.Failure(error.TaskTimeout(msg))
self._finishTask(t.taskid, f)
self.idleLater = None
def taskCompleted(self, result, taskid, workerid):
"""This is the err/callback for a completed task."""
try:
task = self.pendingTasks.pop(workerid)
except:
# this should not happen
log.msg("Tried to pop bad pending task %i from worker %i"%(taskid, workerid))
log.msg("Result: %r"%result)
log.msg("Pending tasks: %s"%self.pendingTasks)
return
# Check if aborted while pending
aborted = False
if taskid in self.abortPending:
self._doAbort(taskid)
aborted = True
if not aborted:
if result.failure is not None and isinstance(result.failure, failure.Failure): # we failed
log.msg("Task %i failed on worker %i"% (taskid, workerid))
if task.retries > 0: # resubmit
task.retries -= 1
self.scheduler.add_task(task)
s = "Resubmitting task %i, %i retries remaining" %(taskid, task.retries)
log.msg(s)
self.distributeTasks()
elif isinstance(task.recovery_task, Task) and \
task.recovery_task.retries > -1:
# retries = -1 is to prevent infinite recovery_task loop
task.retries = -1
task.recovery_task.taskid = taskid
task = task.recovery_task
self.scheduler.add_task(task)
s = "Recovering task %i, %i retries remaining" %(taskid, task.retries)
log.msg(s)
self.distributeTasks()
else: # done trying
self._finishTask(taskid, result)
# wait a second before readmitting a worker that failed
# it may have died, and not yet been unregistered
reactor.callLater(self.failurePenalty, self.readmitWorker, workerid)
else: # we succeeded
# log.msg("Task completed: %i"% taskid)
self._finishTask(taskid, result)
self.readmitWorker(workerid)
else:# we aborted the task
if result.failure is not None and isinstance(result.failure, failure.Failure): # it failed, penalize worker
reactor.callLater(self.failurePenalty, self.readmitWorker, workerid)
else:
self.readmitWorker(workerid)
def readmitWorker(self, workerid):
"""Readmit a worker to the scheduler.
This is outside `taskCompleted` because of the `failurePenalty` being
implemented through `reactor.callLater`.
"""
if workerid in self.workers.keys() and workerid not in self.pendingTasks.keys():
self.scheduler.add_worker(self.workers[workerid])
self.distributeTasks()
components.registerAdapter(TaskController, cs.IControllerBase, ITaskController)