# 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. import abc import contextlib import functools import itertools import threading from oslo_utils import excutils from oslo_utils import timeutils from taskflow.conductors import base from taskflow import exceptions as excp from taskflow.listeners import logging as logging_listener from taskflow import logging from taskflow import states from taskflow.types import timing as tt from taskflow.utils import iter_utils from taskflow.utils import misc LOG = logging.getLogger(__name__) class ExecutorConductor(base.Conductor, metaclass=abc.ABCMeta): """Dispatches jobs from blocking :py:meth:`.run` method to some executor. This conductor iterates over jobs in the provided jobboard (waiting for the given timeout if no jobs exist) and attempts to claim them, work on those jobs using an executor (potentially blocking further work from being claimed and consumed) and then consume those work units after completion. This process will repeat until the conductor has been stopped or other critical error occurs. NOTE(harlowja): consumption occurs even if a engine fails to run due to a atom failure. This is only skipped when an execution failure or a storage failure occurs which are *usually* correctable by re-running on a different conductor (storage failures and execution failures may be transient issues that can be worked around by later execution). If a job after completing can not be consumed or abandoned the conductor relies upon the jobboard capabilities to automatically abandon these jobs. """ LOG = None """ Logger that will be used for listening to events (if none then the module level logger will be used instead). """ REFRESH_PERIODICITY = 30 """ Every 30 seconds the jobboard will be resynced (if for some reason a watch or set of watches was not received) using the `ensure_fresh` option to ensure this (for supporting jobboard backends only). """ #: Default timeout used to idle/wait when no jobs have been found. WAIT_TIMEOUT = 0.5 MAX_SIMULTANEOUS_JOBS = -1 """ Default maximum number of jobs that can be in progress at the same time. Negative or zero values imply no limit (do note that if a executor is used that is built on a queue, as most are, that this will imply that the queue will contain a potentially large & unfinished backlog of submitted jobs). This *may* get better someday if https://bugs.python.org/issue22737 is ever implemented and released. """ #: Exceptions that will **not** cause consumption to occur. NO_CONSUME_EXCEPTIONS = tuple([ excp.ExecutionFailure, excp.StorageFailure, ]) _event_factory = threading.Event """This attribute *can* be overridden by subclasses (for example if an eventlet *green* event works better for the conductor user).""" EVENTS_EMITTED = tuple([ 'compilation_start', 'compilation_end', 'preparation_start', 'preparation_end', 'validation_start', 'validation_end', 'running_start', 'running_end', 'job_consumed', 'job_abandoned', ]) """Events will be emitted for each of the events above. The event is emitted to listeners registered with the conductor. """ def __init__(self, name, jobboard, persistence=None, engine=None, engine_options=None, wait_timeout=None, log=None, max_simultaneous_jobs=None): if max_simultaneous_jobs is None: max_simultaneous_jobs = MAX_SIMULTANEOUS_JOBS super().__init__( name, jobboard, persistence=persistence, engine=engine, engine_options=engine_options) self._wait_timeout = tt.convert_to_timeout( value=wait_timeout, default_value=self.WAIT_TIMEOUT, event_factory=self._event_factory) self._dead = self._event_factory() self._log = misc.pick_first_not_none(log, self.LOG, LOG) self._max_simultaneous_jobs = int( misc.pick_first_not_none(max_simultaneous_jobs, self.MAX_SIMULTANEOUS_JOBS)) self._dispatched = set() def _executor_factory(self): """Creates an executor to be used during dispatching.""" raise excp.NotImplementedError("This method must be implemented but" " it has not been") def stop(self): self._wait_timeout.interrupt() # Inherit the docs, so we can reference them in our class docstring, # if we don't do this sphinx gets confused... stop.__doc__ = base.Conductor.stop.__doc__ @property def dispatching(self): """Whether or not the dispatching loop is still dispatching.""" return not self._dead.is_set() def _listeners_from_job(self, job, engine): listeners = super()._listeners_from_job(job, engine) listeners.append(logging_listener.LoggingListener(engine, log=self._log)) return listeners def _dispatch_job(self, job): engine = self._engine_from_job(job) listeners = self._listeners_from_job(job, engine) with contextlib.ExitStack() as stack: for listener in listeners: stack.enter_context(listener) self._log.debug("Dispatching engine for job '%s'", job) consume = True details = { 'job': job, 'engine': engine, 'conductor': self, } def _run_engine(): has_suspended = False for _state in engine.run_iter(): if not has_suspended and self._wait_timeout.is_stopped(): self._log.info("Conductor stopped, requesting " "suspension of engine running " "job %s", job) engine.suspend() has_suspended = True try: for stage_func, event_name in [(engine.compile, 'compilation'), (engine.prepare, 'preparation'), (engine.validate, 'validation'), (_run_engine, 'running')]: self._notifier.notify("%s_start" % event_name, details) stage_func() self._notifier.notify("%s_end" % event_name, details) except excp.WrappedFailure as e: if all(f.check(*self.NO_CONSUME_EXCEPTIONS) for f in e): consume = False if self._log.isEnabledFor(logging.WARNING): if consume: self._log.warn( "Job execution failed (consumption being" " skipped): %s [%s failures]", job, len(e)) else: self._log.warn( "Job execution failed (consumption" " proceeding): %s [%s failures]", job, len(e)) # Show the failure/s + traceback (if possible)... for i, f in enumerate(e): self._log.warn("%s. %s", i + 1, f.pformat(traceback=True)) except self.NO_CONSUME_EXCEPTIONS: self._log.warn("Job execution failed (consumption being" " skipped): %s", job, exc_info=True) consume = False except Exception: self._log.warn( "Job execution failed (consumption proceeding): %s", job, exc_info=True) else: if engine.storage.get_flow_state() == states.SUSPENDED: self._log.info("Job execution was suspended: %s", job) consume = False else: self._log.info("Job completed successfully: %s", job) return consume def _try_finish_job(self, job, consume, trash=False): try: if consume: self._jobboard.consume(job, self._name) self._notifier.notify("job_consumed", { 'job': job, 'conductor': self, 'persistence': self._persistence, }) elif trash: self._jobboard.trash(job, self._name) self._notifier.notify("job_trashed", { 'job': job, 'conductor': self, 'persistence': self._persistence, }) else: self._jobboard.abandon(job, self._name) self._notifier.notify("job_abandoned", { 'job': job, 'conductor': self, 'persistence': self._persistence, }) except (excp.JobFailure, excp.NotFound): if consume: self._log.warn("Failed job consumption: %s", job, exc_info=True) else: self._log.warn("Failed job abandonment: %s", job, exc_info=True) def _on_job_done(self, job, fut): consume = False trash = False try: consume = fut.result() except KeyboardInterrupt: with excutils.save_and_reraise_exception(): self._log.warn("Job dispatching interrupted: %s", job) except Exception: self._log.warn("Job dispatching failed: %s", job, exc_info=True) trash = True try: self._try_finish_job(job, consume, trash) finally: self._dispatched.discard(fut) def _can_claim_more_jobs(self, job): if self._wait_timeout.is_stopped(): return False if self._max_simultaneous_jobs <= 0: return True if len(self._dispatched) >= self._max_simultaneous_jobs: return False else: return True def _run_until_dead(self, executor, max_dispatches=None): total_dispatched = 0 if max_dispatches is None: # NOTE(TheSriram): if max_dispatches is not set, # then the conductor will run indefinitely, and not # stop after 'n' number of dispatches max_dispatches = -1 dispatch_gen = iter_utils.iter_forever(max_dispatches) is_stopped = self._wait_timeout.is_stopped try: # Don't even do any work in the first place... if max_dispatches == 0: raise StopIteration fresh_period = timeutils.StopWatch( duration=self.REFRESH_PERIODICITY) fresh_period.start() while not is_stopped(): any_dispatched = False if fresh_period.expired(): ensure_fresh = True fresh_period.restart() else: ensure_fresh = False job_it = itertools.takewhile( self._can_claim_more_jobs, self._jobboard.iterjobs(ensure_fresh=ensure_fresh)) for job in job_it: self._log.debug("Trying to claim job: %s", job) try: self._jobboard.claim(job, self._name) except (excp.UnclaimableJob, excp.NotFound): self._log.debug("Job already claimed or" " consumed: %s", job) else: try: fut = executor.submit(self._dispatch_job, job) except RuntimeError: with excutils.save_and_reraise_exception(): self._log.warn("Job dispatch submitting" " failed: %s", job) self._try_finish_job(job, False) else: fut.job = job self._dispatched.add(fut) any_dispatched = True fut.add_done_callback( functools.partial(self._on_job_done, job)) total_dispatched = next(dispatch_gen) if not any_dispatched and not is_stopped(): self._wait_timeout.wait() except StopIteration: # This will be raised from 'dispatch_gen' if it reaches its # max dispatch number (which implies we should do no more work). with excutils.save_and_reraise_exception(): if max_dispatches >= 0 and total_dispatched >= max_dispatches: self._log.info("Maximum dispatch limit of %s reached", max_dispatches) def run(self, max_dispatches=None): self._dead.clear() self._dispatched.clear() try: self._jobboard.register_entity(self.conductor) with self._executor_factory() as executor: self._run_until_dead(executor, max_dispatches=max_dispatches) except StopIteration: pass except KeyboardInterrupt: with excutils.save_and_reraise_exception(): self._log.warn("Job dispatching interrupted") finally: self._dead.set() # Inherit the docs, so we can reference them in our class docstring, # if we don't do this sphinx gets confused... run.__doc__ = base.Conductor.run.__doc__ def wait(self, timeout=None): return self._dead.wait(timeout) # Inherit the docs, so we can reference them in our class docstring, # if we don't do this sphinx gets confused... wait.__doc__ = base.Conductor.wait.__doc__