# Copyright 2011-2014 Biomedical Imaging Group Rotterdam, Departments of
# Medical Informatics and Radiology, Erasmus MC, Rotterdam, The Netherlands
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This module holds the ExecutionPluginManager as well as the base-class
for all ExecutionPlugins.
"""
from abc import abstractmethod
from collections import deque
import functools
import Queue
import sys
import threading
import traceback
import types
from enum import Enum
import fastr
from fastr import exceptions
from fastr.core.baseplugin import Plugin
from fastr.core.pluginmanager import PluginSubManager
from fastr.execution.job import Job, JobState, SourceJob, SinkJob
from fastr.utils import iohelpers
from fastr.utils.filesynchelper import FileSyncHelper, filesynchelper_enabled
[docs]class JobAction(Enum):
"""
Job actions that can be performed. This is used for checking
if held jobs should be queued, held longer or be cancelled.
"""
hold = 'hold'
queue = 'queue'
cancel = 'cancel'
[docs]class ExecutionPlugin(Plugin):
"""
This class is the base for all Plugins to execute jobs somewhere. There are
many methods already in place for taking care of stuff.
There are fall-backs for certain features, but if a system already
implements those it is usually preferred to skip the fall-back and
let the external system handle it. There are a few flags to enable
disable these features:
* ``cls.SUPPORTS_CANCEL`` indicates that the plugin can cancel queued jobs
* ``cls.SUPPORTS_HOLD_RELEASE`` indicates that the plugin can queue jobs in
a hold state and can release them again (if not, the base plugin will
create a hidden queue for held jobs). The plugin should respect the
``Job.status == JobState.hold`` when queueing jobs.
* ``cls.SUPPORTS_DEPENDENCY`` indicate that the plugin can manage job
dependencies, if not the base plugin job dependency system will be used
and jobs with only be submitted when all dependencies are met.
* ``cls.CANCELS_DEPENDENCIES`` indicates that if a job is cancelled it will
automatically cancel all jobs depending on that job. If not the plugin
traverse the dependency graph and kill each job manual.
.. note:: If a plugin supports dependencies it is assumed that when a
job gets cancelled, the depending job also get cancelled
automatically!
Most plugins should only need to redefine a few abstract methods:
* :py:meth:`__init__ <fastr.execution.executionpluginmanager.ExecutionPlugin.__init__>`
the constructor
* :py:meth:`cleanup <fastr.execution.executionpluginmanager.ExecutionPlugin.__init__>`
a clean up function that frees resources, closes connections, etc
* :py:meth:`_queue_job <fastr.execution.executionpluginmanager.ExecutionPlugin._queue_job>`
the method that queues the job for execution
Optionally an extra job finished callback could be added:
* :py:meth:`_job_finished <fastr.execution.executionpluginmanager.ExecutionPlugin._job_finished>`
extra callback for when a job finishes
If ``SUPPORTS_CANCEL`` is set to True, the plugin should also implement:
* :py:meth:`_cancel_job <fastr.execution.executionpluginmanager.ExecutionPlugin._cancel_job>`
cancels a previously queued job
If ``SUPPORTS_HOLD_RELEASE`` is set to True, the plugin should also implement:
* :py:meth:`_hold_job <fastr.execution.executionpluginmanager.ExecutionPlugin._hold_job>`
hold_job a job that is currently held
* :py:meth:`_release_job <fastr.execution.executionpluginmanager.ExecutionPlugin._release_job>`
releases a job that is currently held
If ``SUPPORTED_DEPENDENCY`` is set to True, the plugin should:
* Make sure to use the ``Job.hold_jobs`` as a list of its dependencies
Not all of the functions need to actually do anything for a plugin. There
are examples of plugins that do not really need a ``cleanup``, but for
safety you need to implement it. Just using a ``pass`` for the method could
be fine in such a case.
.. warning::
When overwriting other functions, extreme care must be taken not to break
the plugins working, as there is a lot of bookkeeping that can go wrong.
"""
#: Indicates if the plugin can cancel queued jobs
SUPPORTS_CANCEL = False
#: Indicates if the plugin can queue jobs in a hold state and can release
#: them again (if not, the base plugin will create a hidden queue for held
#: jobs)
SUPPORTS_HOLD_RELEASE = False
#: Indicate if the plugin can manage job dependencies, if not the base
#: plugin job dependency system will be used and jobs with only be
#: submitted when all dependencies are met.
SUPPORTS_DEPENDENCY = False
#: Indicates that when a job is cancelled the dependenceis
CANCELS_DEPENDENCIES = False
@abstractmethod
[docs] def __init__(self, finished_callback=None, cancelled_callback=None, status_callback=None):
"""
Setup the ExecutionPlugin
:param finished_callback: the callback to call after a job finished
:param cancelled_callback: the callback to call after a job cancelled
:return: newly created ExecutionPlugin
"""
super(ExecutionPlugin, self).__init__()
# Pylint seems to be unable to figure out the .dict() member
# pylint: disable=no-member
self.job_status = {}
self.job_dict = {}
self.job_archive = {}
self._finished_callback = finished_callback
self._cancelled_callback = cancelled_callback
self._status_callback = status_callback
# Dict indicating the depending jobs for a certain jobs (who is waiting on the key job id)
self.held_queue = {}
self.held_queue_lock = threading.Lock()
# A list for the source jobs that should be held
self.source_queue_lock = threading.Lock()
self.source_job_queue = deque()
self.queued_source_jobs = deque()
self.source_job_limit = fastr.config.source_job_limit
# Flag indicating the plugin is accepting new jobs queued
self.accepting = True
# Create a thread for processing callbacks
self.callback_queue = Queue.Queue()
self.processing_callbacks = True # Flag to keep thread alive
fastr.log.debug('Creating callback thread')
self.callback_processor = threading.Thread(name='CallbackProcessor-0', target=self.process_callbacks, args=())
self.callback_processor.daemon = True
fastr.log.debug('Starting callback thread')
self.callback_processor.start()
[docs] def __enter__(self):
return self
[docs] def __exit__(self, type_, value, tb):
self.cleanup()
[docs] def __del__(self):
"""
Cleanup if the variable was deleted on purpose
"""
fastr.log.debug('Calling cleanup')
self.cleanup()
@abstractmethod
[docs] def cleanup(self):
"""
Method to call to clean up the ExecutionPlugin. This can be to clear
temporary data, close connections, etc.
:param force: force cleanup (e.g. kill instead of join a process)
"""
# Stop accepting new jobs (close the queue)
self.accepting = False
# Stop processing of callbacks
self.processing_callbacks = False
# Cancel all queued jobs
while len(self.job_dict) > 0:
job_id, job = self.job_dict.popitem()
fastr.log.debug('Cleanup cancelling {}'.format(job_id))
self.cancel_job(job)
# End the callback processor
if self.callback_processor.isAlive():
fastr.log.debug('Terminating callback thread')
self.callback_processor.join()
[docs] def register_job(self, job):
self.job_dict[job.id] = job
self.job_status[job.id] = job.status
for hold_id in job.hold_jobs:
# Add job reference to held queue to receive signal when the
# required jobs are finished/failed. Do not subscribe for jobs that
# are already finished.
if hold_id in self.job_status and self.job_status[hold_id].done:
continue
with self.held_queue_lock:
if hold_id not in self.held_queue:
self.held_queue[hold_id] = []
# append to held_queue, because of the managed dict, we need to replace the value (not update)
self.held_queue[hold_id].append(job.id)
def _dispatch_job(self, job):
# Check the job requirements for dependencies and take appropriate
# action while keeping in mind the plugin capabilities.
action = self.check_job_requirements(job.id)
if action == JobAction.cancel:
self.cancel_job(job)
elif self.SUPPORTS_DEPENDENCY or action == JobAction.queue:
self.job_status[job.id] = job.status = JobState.queued
self._queue_job_limited(job)
else: # Thus action has to be JobAction.hold:
fastr.log.debug('Holding {} until dependencies are met'.format(job.id))
self.job_status[job.id] = job.status = JobState.hold
# If there is support for hold release the plugin should respect
# the job.status being JobState.hold and it is safe to queue in
# a held state already.
if self.SUPPORTS_HOLD_RELEASE:
self._queue_job_limited(job)
def _queue_job_limited(self, job):
# Check if it is a source job and limit if needed
if isinstance(job, (SourceJob, SinkJob)) and self.source_job_limit > 0:
with self.source_queue_lock:
if len(self.queued_source_jobs) < self.source_job_limit:
self.queued_source_jobs.append(job.id)
fastr.log.debug('Queueing job, now having {} queued'.format(len(self.queued_source_jobs)))
self._queue_job(job)
else:
# Keep the job in a temporary queue, do no dispatch!
fastr.log.debug('Putting soure/sink job {} in source queue due to limitations!'.format(job.id))
self.source_job_queue.append(job)
if self.SUPPORTS_HOLD_RELEASE:
# Submit the job in a hold state
fastr.log.debug('Queue SourceJob {} as hold'.format(job.status))
job.status = JobState.hold
self._queue_job(job)
else:
self._queue_job(job)
[docs] def queue_job(self, job):
"""
Add a job to the execution queue
:param Job job: job to add
"""
if not self.accepting:
return
if isinstance(job, list):
for j in job:
self.queue_job(j)
return
# Register the job for tracking
self.register_job(job)
# Save the job (and initial result) to file before serializing
iohelpers.save_gpickle(job.commandfile, job)
# If the job has been preset to done, immediately send for the callback
if job.status in [JobState.execution_done, JobState.execution_failed]:
self.job_finished(job)
return
# Dispatch the job to the plugin
self._dispatch_job(job)
[docs] def cancel_job(self, job):
"""
Cancel a job previously queued
:param job: job to cancel
"""
if not isinstance(job, Job):
try:
job = self.job_dict[job]
except KeyError:
fastr.log.warning('Job {} is no longer under processing, cannot cancel!'.format(job))
return
# Make sure the job is not using a source/sink slot
try:
with self.source_queue_lock:
self.source_job_queue.remove(job)
except ValueError:
pass # Job was not in the source_job_queue and that is fine
# Check if job is not already being cancelled
if job.status == JobState.cancelled:
return
if job.status == JobState.hold and not self.SUPPORTS_CANCEL and self.SUPPORTS_HOLD_RELEASE:
# This is a corner case where the job is scheduled and held, but
# somehow cannot be cancelled. We need to release the Job and let
# it crash itself.
self.release_job(job)
if not self.SUPPORTS_CANCEL and job.status not in [JobState.created, JobState.hold]:
fastr.log.warning('Job is already queued or running and cannot be cancelled anymore!')
return
fastr.log.debug('Cancelling {}'.format(job.id))
# Cancel job
old_status = job.status
job.status = self.job_status[job.id] = JobState.cancelled
# If supported, send out the actual cancellation command
# If the job status is created, it means it never got properly
# accepted by the plugin and thus only needs the callback, but
# not the actual cancellation
if self.SUPPORTS_CANCEL and old_status != JobState.created:
self._cancel_job(job)
fastr.log.debug('Removing {} from jobdict'.format(job.id))
self.job_archive[job.id] = job
# Initialize provenance and write the final state of the job to
# the logfile
job.provenance.init_provenance(job)
job.write()
try:
del self.job_dict[job.id]
except KeyError:
pass
# Calling callback for cancelling
fastr.log.debug('Calling cancelled for {}'.format(job.id))
if self._cancelled_callback is not None:
self._cancelled_callback(job)
# Cancel all children (some systems might have automatically cancelled them)
fastr.log.debug('Cancelling children for {}'.format(job.id))
if not self.CANCELS_DEPENDENCIES:
if job.id in self.held_queue:
fastr.log.debug('Found children....')
held_queue = self.held_queue[job.id]
for dependent_job in held_queue:
fastr.log.debug('Checking sub {}'.format(dependent_job))
if dependent_job in self.job_dict and dependent_job in self.job_status and not self.job_status[dependent_job].done:
fastr.log.debug('Cancelling sub {}'.format(dependent_job))
self.cancel_job(dependent_job)
else:
fastr.log.debug('No children....')
[docs] def hold_job(self, job):
if self.SUPPORTS_HOLD_RELEASE:
if not isinstance(job, Job):
if job not in self.job_dict:
fastr.log.warning('Job {} is no longer under processing, cannot release!'.format(job))
return
try:
job = self.job_dict[job]
except KeyError:
fastr.log.warning('Job {} is no longer under processing, cannot release!'.format(job))
return
job.status = JobState.hold
self._hold_job(job)
else:
raise exceptions.FastrNotImplementedError('Cannot handle hold/release by default yet!')
[docs] def release_job(self, job):
"""
Release a job that has been put on hold
:param job: job to release
"""
if not isinstance(job, Job):
try:
job = self.job_dict[job]
except KeyError:
fastr.log.warning('Job {} is no longer under processing, cannot release!'.format(job))
return
job.status = JobState.queued
if self.SUPPORTS_HOLD_RELEASE:
# Job is already queued, but held so it need to be released
self._release_job(job)
else:
# Job was never queued in the first place
self.queue_job(job)
[docs] def job_finished(self, job, errors=None, blocking=False):
"""
The default callback that is called when a Job finishes. This will
create a new thread that handles the actual callback.
:param Job job: the job that finished
:param errors: optional list of errors encountered
:param bool blocking: if blocking, do not create threads
:return:
"""
if isinstance(job, (SourceJob, SinkJob)) and self.source_job_limit > 0:
with self.source_queue_lock:
self.queued_source_jobs.remove(job.id)
if len(self.source_job_queue) > 0:
fastr.log.debug('Taking source job from the queue')
new_source_job = self.source_job_queue.popleft()
while not new_source_job.status.idle:
fastr.log.warning('Discarding non-idle queued source/sink job {} ({})'.format(
new_source_job,
new_source_job.status)
)
new_source_job = self.source_job_queue.popleft()
fastr.log.debug('Selected {} to run'.format(new_source_job.id))
self.queued_source_jobs.append(new_source_job.id)
if self.SUPPORTS_HOLD_RELEASE:
self.release_job(new_source_job)
else:
self._queue_job(new_source_job)
fastr.log.debug('New number of queued source jobs: {}'.format(
len(self.queued_source_jobs)
))
if not blocking:
self.callback_queue.put((job, errors))
else:
self._job_finished_body(job, errors)
[docs] def process_callbacks(self):
while self.processing_callbacks:
try:
job, errors = self.callback_queue.get(block=True, timeout=2)
self._job_finished_body(job, errors)
except Queue.Empty:
pass
fastr.log.info('Callback processing thread ended!')
def _job_finished_body(self, job, errors):
"""
The actual callback that is executed in a separate thread. This
method handles the collection of the result, the release of depending
jobs and calling the user defined callback.
:param Job job: the job that finished
:param errors: optional list of errors encountered
"""
fastr.log.debug('ExecutorInterface._job_finished_callback called for {}'.format(job))
self.job_status[job.id] = JobState.processing_callback
if errors is None:
errors = []
# The Job finished should always log the errors rather than
# crashing the whole execution system
# pylint: disable=bare-except
try:
try:
if filesynchelper_enabled():
FileSyncHelper().wait_for_job(job.logurl)
job = iohelpers.load_gpickle(job.logfile)
except EOFError:
errors.append(
exceptions.FastrResultFileNotFound(
('Could not read job result file {}, assuming '
'the job crashed during output write.').format(job.logfile)).excerpt())
job.status = JobState.failed
except IOError:
errors.append(
exceptions.FastrResultFileNotFound(
('Could not find/read job result file {}, assuming '
'the job crashed before it created output.').format(job.logfile)).excerpt())
job.status = JobState.failed
if self._status_callback is not None:
self._status_callback(job)
job.status_callback = self._status_callback
except:
exc_type, _, trace = sys.exc_info()
exc_info = traceback.format_exc()
trace = traceback.extract_tb(trace, 1)[0]
fastr.log.error('Encountered exception ({}) during execution:\n{}'.format(exc_type.__name__, exc_info))
errors.append((exc_type.__name__, exc_info, trace[0], trace[1]))
job.status = JobState.execution_failed
# The execution plugin found that there were errors, append them
if len(errors) > 0:
job.errors.extend(errors)
# Resave the job with the updated error list
fastr.log.info('Found errors from execution plugin, updating pickle {}'.format(job.logfile))
iohelpers.save_gpickle(job.logfile, job)
# Strip provenance information from the job
del job.provenance
result = job
fastr.log.debug('Finished {} with status {}'.format(job.id, job.status))
job_id = result.id
# Make sure the status is either finished or failed
if result.status == JobState.execution_done:
result.status = JobState.finished
else:
result.status = JobState.failed
# Set the job status so the hold jobs will be release properly
self.job_status[job_id] = result.status
if job_id in self.job_dict:
self.job_dict[job_id].status = result.status
# Do the callbacks before releasing the other jobs
# Extra subclass callback
fastr.log.debug('Subclass callback')
try:
self._job_finished(result)
except:
exc_type, _, _ = sys.exc_info()
exc_info = traceback.format_exc()
fastr.log.error('Encountered exception ({}) during callback {}._job_finished:\n{}'.format(exc_type.__name__, type(self).__name__, exc_info))
# Extra callback from object
fastr.log.debug('Calling callback for {}'.format(job_id))
if self._finished_callback is not None:
try:
self._finished_callback(result)
except:
if isinstance(self._finished_callback, functools.partial):
args = self._finished_callback.args + tuple('{}={}'.format(k, v) for k, v in self._finished_callback.keywords.items())
callback_name = '{f.__module__}.{f.func_name}({a})'.format(f=self._finished_callback.func, a=','.join(args))
elif isinstance(self._finished_callback, types.FunctionType):
callback_name = '{f.__module__}.{f.func_name}'.format(f=self._finished_callback)
elif isinstance(self._finished_callback, types.MethodType):
callback_name = '{m.__module__}.{m.im_class.__name__}.{m.im_func.func_name}'.format(m=self._finished_callback)
else:
callback_name = repr(self._finished_callback)
exc_type, _, _ = sys.exc_info()
exc_info = traceback.format_exc()
fastr.log.error('Encountered exception ({}) during callback {}:\n{}'.format(exc_type.__name__, callback_name, exc_info))
else:
fastr.log.debug('No callback specified')
# The ProcessPoolExecutor has to track job dependencies itself, so
# therefor we have to check for jobs depending on the finished job
if job_id in self.held_queue:
fastr.log.debug('Signaling depending jobs {}'.format(self.held_queue[job_id]))
ready_jobs = []
for held_job_id in self.held_queue[job_id]:
action = self.check_job_requirements(held_job_id)
if action == JobAction.queue:
# Re-assign managed dict member
if held_job_id in self.job_dict:
held_job = self.job_dict[held_job_id]
self.job_dict[held_job_id] = held_job
fastr.log.debug('Job {} is now ready to be submitted'.format(held_job_id))
# If ready, flag job for removal from held queue and send
# to pool queue to be executed
ready_jobs.append(held_job_id)
if not self.SUPPORTS_DEPENDENCY:
self.release_job(held_job_id)
elif action == JobAction.cancel:
fastr.log.debug('Job {} will be cancelled'.format(held_job_id))
ready_jobs.append(held_job_id)
self.cancel_job(held_job_id)
else:
fastr.log.debug('Job {} still has unmet dependencies'.format(held_job_id))
# Remove jobs that no longer need to be held from held_queue
for ready_job_id in ready_jobs:
job = self.get_job(ready_job_id)
for hold_id in job.hold_jobs:
with self.held_queue_lock: # First lock to avoid any possible mutations
if hold_id in self.held_queue:
# remove from held_queue. because of the managed dict,
# we need to replace the value (not update)
required_job = self.held_queue[hold_id]
try:
required_job.remove(ready_job_id)
except ValueError:
fastr.log.warning('Could not remove {} from dependencies {} of {}'.format(ready_job_id, required_job, hold_id))
with self.held_queue_lock:
del self.held_queue[job_id]
# Move the job to archive (to keep the number of working jobs limited
# in the future the archive can be moved to a db/disk if needed
fastr.log.debug('Archiving job {} with status {}'.format(job_id, result.status))
try:
del self.job_status[job_id]
except KeyError:
pass
self.job_archive[job_id] = result
try:
del self.job_dict[job_id]
except KeyError:
pass
fastr.log.debug('Done archiving')
[docs] def get_job(self, job_id):
try:
return self.job_dict[job_id]
except KeyError:
try:
return self.job_archive[job_id]
except:
raise exceptions.FastrKeyError('Could not find job {}'.format(job_id))
[docs] def get_status(self, job):
if not isinstance(job, Job):
job = self.get_job(job)
return job.status
@abstractmethod
def _queue_job(self, job):
"""
Method that a subclass implements to actually queue a Job for execution
:param job: job to queue
.. note:: If ``SUPPORT_HOLD_RELEASE=True`` the plugin should check
``job.status`` to see if it is ``JobStatus.hold``. If
this is the case it should submit the job in a held
state to be released at a later time. If
``SUPPORT_HOLD_RELEASE=False`` the status be ignored.
"""
def _cancel_job(self, job):
"""
Method that a subclass implements to actually cancel a Job
:param job: job to cancel
"""
if self.SUPPORTS_CANCEL:
raise exceptions.FastrPluginCapabilityNotImplemented('The plugin should have reimplemented _cancel_job!')
else:
fastr.log.warning('Cannot remove job {} from a {} queue!'.format(job.id,
type(self).__name__))
def _hold_job(self, job):
"""
Method that a subclass implements to hold a queued/running Job, only
required if ``SUPPORT_HOLD_RELEASE=True``
:param job: Job to hold
"""
if self.SUPPORTS_HOLD_RELEASE:
raise exceptions.FastrPluginCapabilityNotImplemented(
'Cannot hold job {}: the plugin should have reimplemented _hold_job!'.format(
job.id
)
)
def _release_job(self, job):
"""
Method that a subclass implements to actually release a job, only
required if ``SUPPORT_HOLD_RELEASE=True``
:param job: job to release from hold
"""
if self.SUPPORTS_HOLD_RELEASE:
raise exceptions.FastrPluginCapabilityNotImplemented(
'Cannot release job {}: the plugin should have reimplemented _release_job!'.format(
job.id
)
)
def _job_finished(self, job):
"""
Method that a subclass can implement to add to the default callback.
It will be called by ``_job_finished_body`` right before the user
defined callback will be called.
:param job: Job that resulted from the execution
"""
[docs] def show_jobs(self, req_status=None):
"""
List the queued jobs, possible filtered by status
:param req_status: requested status to filter on
:return: list of jobs
"""
if isinstance(req_status, basestring):
req_status = JobState[req_status]
if not isinstance(req_status, JobState):
return []
results = []
for key, status in self.job_status.items():
if req_status is None or status == req_status:
results.append(self.get_job(key))
return results
[docs] def check_job_status(self, job_id):
"""
Get the status of a specified job
:param job_id: the target job
:return: the status of the job (or None if job not found)
"""
try:
return self.get_status(job_id)
except exceptions.FastrKeyError:
return None
[docs] def check_job_requirements(self, job_id):
"""
Check if the requirements for a job are fulfilled.
:param job_id: job to check
:return: directive what should happen with the job
:rtype: JobAction
"""
job = self.get_job(job_id)
if job.hold_jobs is None or len(job.hold_jobs) == 0:
return JobAction.queue
status_list = [self.check_job_status(jid) for jid in job.hold_jobs]
if all(x.done for x in status_list):
if all(x == JobState.finished for x in status_list):
return JobAction.queue
else:
return JobAction.cancel
else:
return JobAction.hold
[docs]class ExecutionPluginManager(PluginSubManager):
"""
Container holding all the ExecutionPlugins known to the Fastr system
"""
[docs] def __init__(self):
"""
Initialize a ExecutionPluginManager and load plugins.
:param path: path to search for plugins
:param recursive: flag for searching recursively
:return: newly created ExecutionPluginManager
"""
super(ExecutionPluginManager, self).__init__(parent=fastr.plugin_manager,
plugin_class=ExecutionPlugin)
@property
def _instantiate(self):
"""
Indicate that the plugins should not instantiated before being stored
"""
return False