X-Git-Url: https://wannabe.guru.org/gitweb/?a=blobdiff_plain;f=executors.py;h=47b4a89a88d693d535ed2e036c6288829505a005;hb=f2b4fe83f6fc853a68653bd5e3d9fe0648c3d105;hp=c91c2a6af535502ad31db6bb233fa9145d43dd4a;hpb=eedcbd4f64af13ec2098508c3d839a60f7e9ffce;p=python_utils.git diff --git a/executors.py b/executors.py index c91c2a6..47b4a89 100644 --- a/executors.py +++ b/executors.py @@ -32,15 +32,14 @@ from thread_utils import background_thread logger = logging.getLogger(__name__) parser = config.add_commandline_args( - f"Executors ({__file__})", - "Args related to processing executors." + f"Executors ({__file__})", "Args related to processing executors." ) parser.add_argument( '--executors_threadpool_size', type=int, metavar='#THREADS', help='Number of threads in the default threadpool, leave unset for default', - default=None + default=None, ) parser.add_argument( '--executors_processpool_size', @@ -64,7 +63,7 @@ parser.add_argument( ) SSH = '/usr/bin/ssh -oForwardX11=no' -SCP = '/usr/bin/scp' +SCP = '/usr/bin/scp -C' def make_cloud_pickle(fun, *args, **kwargs): @@ -75,33 +74,51 @@ def make_cloud_pickle(fun, *args, **kwargs): class BaseExecutor(ABC): def __init__(self, *, title=''): self.title = title - self.task_count = 0 self.histogram = hist.SimpleHistogram( - hist.SimpleHistogram.n_evenly_spaced_buckets( - int(0), int(500), 50 - ) + hist.SimpleHistogram.n_evenly_spaced_buckets(int(0), int(500), 50) ) + self.task_count = 0 @abstractmethod - def submit(self, - function: Callable, - *args, - **kwargs) -> fut.Future: + def submit(self, function: Callable, *args, **kwargs) -> fut.Future: pass @abstractmethod - def shutdown(self, - wait: bool = True) -> None: + def shutdown(self, wait: bool = True) -> None: pass + def shutdown_if_idle(self) -> bool: + """Shutdown the executor and return True if the executor is idle + (i.e. there are no pending or active tasks). Return False + otherwise. Note: this should only be called by the launcher + process. + + """ + if self.task_count == 0: + self.shutdown() + return True + return False + def adjust_task_count(self, delta: int) -> None: + """Change the task count. Note: do not call this method from a + worker, it should only be called by the launcher process / + thread / machine. + + """ self.task_count += delta - logger.debug(f'Executor current task count is {self.task_count}') + logger.debug(f'Adjusted task count by {delta} to {self.task_count}') + + def get_task_count(self) -> int: + """Change the task count. Note: do not call this method from a + worker, it should only be called by the launcher process / + thread / machine. + + """ + return self.task_count class ThreadExecutor(BaseExecutor): - def __init__(self, - max_workers: Optional[int] = None): + def __init__(self, max_workers: Optional[int] = None): super().__init__() workers = None if max_workers is not None: @@ -110,47 +127,44 @@ class ThreadExecutor(BaseExecutor): workers = config.config['executors_threadpool_size'] logger.debug(f'Creating threadpool executor with {workers} workers') self._thread_pool_executor = fut.ThreadPoolExecutor( - max_workers=workers, - thread_name_prefix="thread_executor_helper" + max_workers=workers, thread_name_prefix="thread_executor_helper" ) + self.already_shutdown = False + # This is run on a different thread; do not adjust task count here. def run_local_bundle(self, fun, *args, **kwargs): logger.debug(f"Running local bundle at {fun.__name__}") - start = time.time() result = fun(*args, **kwargs) - end = time.time() - self.adjust_task_count(-1) - duration = end - start - logger.debug(f"{fun.__name__} finished; used {duration:.1f}s") - self.histogram.add_item(duration) return result @overrides - def submit(self, - function: Callable, - *args, - **kwargs) -> fut.Future: + def submit(self, function: Callable, *args, **kwargs) -> fut.Future: + if self.already_shutdown: + raise Exception('Submitted work after shutdown.') self.adjust_task_count(+1) newargs = [] newargs.append(function) for arg in args: newargs.append(arg) - return self._thread_pool_executor.submit( - self.run_local_bundle, - *newargs, - **kwargs) + start = time.time() + result = self._thread_pool_executor.submit( + self.run_local_bundle, *newargs, **kwargs + ) + result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start)) + result.add_done_callback(lambda _: self.adjust_task_count(-1)) + return result @overrides - def shutdown(self, - wait = True) -> None: - logger.debug(f'Shutting down threadpool executor {self.title}') - print(self.histogram) - self._thread_pool_executor.shutdown(wait) + def shutdown(self, wait=True) -> None: + if not self.already_shutdown: + logger.debug(f'Shutting down threadpool executor {self.title}') + print(self.histogram) + self._thread_pool_executor.shutdown(wait) + self.already_shutdown = True class ProcessExecutor(BaseExecutor): - def __init__(self, - max_workers=None): + def __init__(self, max_workers=None): super().__init__() workers = None if max_workers is not None: @@ -161,38 +175,34 @@ class ProcessExecutor(BaseExecutor): self._process_executor = fut.ProcessPoolExecutor( max_workers=workers, ) + self.already_shutdown = False + # This is run in another process; do not adjust task count here. def run_cloud_pickle(self, pickle): fun, args, kwargs = cloudpickle.loads(pickle) logger.debug(f"Running pickled bundle at {fun.__name__}") result = fun(*args, **kwargs) - self.adjust_task_count(-1) return result @overrides - def submit(self, - function: Callable, - *args, - **kwargs) -> fut.Future: + def submit(self, function: Callable, *args, **kwargs) -> fut.Future: + if self.already_shutdown: + raise Exception('Submitted work after shutdown.') start = time.time() self.adjust_task_count(+1) pickle = make_cloud_pickle(function, *args, **kwargs) - result = self._process_executor.submit( - self.run_cloud_pickle, - pickle - ) - result.add_done_callback( - lambda _: self.histogram.add_item( - time.time() - start - ) - ) + result = self._process_executor.submit(self.run_cloud_pickle, pickle) + result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start)) + result.add_done_callback(lambda _: self.adjust_task_count(-1)) return result @overrides def shutdown(self, wait=True) -> None: - logger.debug(f'Shutting down processpool executor {self.title}') - self._process_executor.shutdown(wait) - print(self.histogram) + if not self.already_shutdown: + logger.debug(f'Shutting down processpool executor {self.title}') + self._process_executor.shutdown(wait) + print(self.histogram) + self.already_shutdown = True def __getstate__(self): state = self.__dict__.copy() @@ -202,6 +212,7 @@ class ProcessExecutor(BaseExecutor): class RemoteExecutorException(Exception): """Thrown when a bundle cannot be executed despite several retries.""" + pass @@ -272,42 +283,30 @@ class BundleDetails: class RemoteExecutorStatus: def __init__(self, total_worker_count: int) -> None: - self.worker_count = total_worker_count + self.worker_count: int = total_worker_count self.known_workers: Set[RemoteWorkerRecord] = set() + self.start_time: float = time.time() self.start_per_bundle: Dict[str, float] = defaultdict(float) self.end_per_bundle: Dict[str, float] = defaultdict(float) self.finished_bundle_timings_per_worker: Dict[ - RemoteWorkerRecord, - List[float] - ] = {} - self.in_flight_bundles_by_worker: Dict[ - RemoteWorkerRecord, - Set[str] + RemoteWorkerRecord, List[float] ] = {} + self.in_flight_bundles_by_worker: Dict[RemoteWorkerRecord, Set[str]] = {} self.bundle_details_by_uuid: Dict[str, BundleDetails] = {} self.finished_bundle_timings: List[float] = [] self.last_periodic_dump: Optional[float] = None - self.total_bundles_submitted = 0 + self.total_bundles_submitted: int = 0 # Protects reads and modification using self. Also used # as a memory fence for modifications to bundle. - self.lock = threading.Lock() + self.lock: threading.Lock = threading.Lock() - def record_acquire_worker( - self, - worker: RemoteWorkerRecord, - uuid: str - ) -> None: + def record_acquire_worker(self, worker: RemoteWorkerRecord, uuid: str) -> None: with self.lock: - self.record_acquire_worker_already_locked( - worker, - uuid - ) + self.record_acquire_worker_already_locked(worker, uuid) def record_acquire_worker_already_locked( - self, - worker: RemoteWorkerRecord, - uuid: str + self, worker: RemoteWorkerRecord, uuid: str ) -> None: assert self.lock.locked() self.known_workers.add(worker) @@ -316,34 +315,28 @@ class RemoteExecutorStatus: x.add(uuid) self.in_flight_bundles_by_worker[worker] = x - def record_bundle_details( - self, - details: BundleDetails) -> None: + def record_bundle_details(self, details: BundleDetails) -> None: with self.lock: self.record_bundle_details_already_locked(details) - def record_bundle_details_already_locked( - self, - details: BundleDetails) -> None: + def record_bundle_details_already_locked(self, details: BundleDetails) -> None: assert self.lock.locked() self.bundle_details_by_uuid[details.uuid] = details def record_release_worker( - self, - worker: RemoteWorkerRecord, - uuid: str, - was_cancelled: bool, + self, + worker: RemoteWorkerRecord, + uuid: str, + was_cancelled: bool, ) -> None: with self.lock: - self.record_release_worker_already_locked( - worker, uuid, was_cancelled - ) + self.record_release_worker_already_locked(worker, uuid, was_cancelled) def record_release_worker_already_locked( - self, - worker: RemoteWorkerRecord, - uuid: str, - was_cancelled: bool, + self, + worker: RemoteWorkerRecord, + uuid: str, + was_cancelled: bool, ) -> None: assert self.lock.locked() ts = time.time() @@ -381,13 +374,14 @@ class RemoteExecutorStatus: if len(self.finished_bundle_timings) > 1: qall = numpy.quantile(self.finished_bundle_timings, [0.5, 0.95]) ret += ( - f'⏱=∀p50:{qall[0]:.1f}s, ∀p95:{qall[1]:.1f}s, ' + f'⏱=∀p50:{qall[0]:.1f}s, ∀p95:{qall[1]:.1f}s, total={ts-self.start_time:.1f}s, ' f'✅={total_finished}/{self.total_bundles_submitted}, ' f'💻n={total_in_flight}/{self.worker_count}\n' ) else: ret += ( - f' ✅={total_finished}/{self.total_bundles_submitted}, ' + f'⏱={ts-self.start_time:.1f}s, ' + f'✅={total_finished}/{self.total_bundles_submitted}, ' f'💻n={total_in_flight}/{self.worker_count}\n' ) @@ -407,10 +401,7 @@ class RemoteExecutorStatus: if in_flight > 0: ret += f' ...{in_flight} bundles currently in flight:\n' for bundle_uuid in self.in_flight_bundles_by_worker[worker]: - details = self.bundle_details_by_uuid.get( - bundle_uuid, - None - ) + details = self.bundle_details_by_uuid.get(bundle_uuid, None) pid = str(details.pid) if (details and details.pid != 0) else "TBD" if self.start_per_bundle[bundle_uuid] is not None: sec = ts - self.start_per_bundle[bundle_uuid] @@ -442,10 +433,7 @@ class RemoteExecutorStatus: assert self.lock.locked() self.total_bundles_submitted = total_bundles_submitted ts = time.time() - if ( - self.last_periodic_dump is None - or ts - self.last_periodic_dump > 5.0 - ): + if self.last_periodic_dump is None or ts - self.last_periodic_dump > 5.0: print(self) self.last_periodic_dump = ts @@ -459,10 +447,7 @@ class RemoteWorkerSelectionPolicy(ABC): pass @abstractmethod - def acquire_worker( - self, - machine_to_avoid = None - ) -> Optional[RemoteWorkerRecord]: + def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]: pass @@ -475,27 +460,32 @@ class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): return False @overrides - def acquire_worker( - self, - machine_to_avoid = None - ) -> Optional[RemoteWorkerRecord]: + def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]: grabbag = [] for worker in self.workers: - for x in range(0, worker.count): - for y in range(0, worker.weight): - grabbag.append(worker) - - for _ in range(0, 5): - random.shuffle(grabbag) - worker = grabbag[0] - if worker.machine != machine_to_avoid or _ > 2: + if worker.machine != machine_to_avoid: if worker.count > 0: - worker.count -= 1 - logger.debug(f'Selected worker {worker}') - return worker - msg = 'Unexpectedly could not find a worker, retrying...' - logger.warning(msg) - return None + for _ in range(worker.count * worker.weight): + grabbag.append(worker) + + if len(grabbag) == 0: + logger.debug( + f'There are no available workers that avoid {machine_to_avoid}...' + ) + for worker in self.workers: + if worker.count > 0: + for _ in range(worker.count * worker.weight): + grabbag.append(worker) + + if len(grabbag) == 0: + logger.warning('There are no available workers?!') + return None + + worker = random.sample(grabbag, 1)[0] + assert worker.count > 0 + worker.count -= 1 + logger.debug(f'Chose worker {worker}') + return worker class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): @@ -511,8 +501,7 @@ class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): @overrides def acquire_worker( - self, - machine_to_avoid: str = None + self, machine_to_avoid: str = None ) -> Optional[RemoteWorkerRecord]: x = self.index while True: @@ -535,9 +524,11 @@ class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy): class RemoteExecutor(BaseExecutor): - def __init__(self, - workers: List[RemoteWorkerRecord], - policy: RemoteWorkerSelectionPolicy) -> None: + def __init__( + self, + workers: List[RemoteWorkerRecord], + policy: RemoteWorkerSelectionPolicy, + ) -> None: super().__init__() self.workers = workers self.policy = policy @@ -550,7 +541,9 @@ class RemoteExecutor(BaseExecutor): raise RemoteExecutorException(msg) self.policy.register_worker_pool(self.workers) self.cv = threading.Condition() - logger.debug(f'Creating {self.worker_count} local threads, one per remote worker.') + logger.debug( + f'Creating {self.worker_count} local threads, one per remote worker.' + ) self._helper_executor = fut.ThreadPoolExecutor( thread_name_prefix="remote_executor_helper", max_workers=self.worker_count, @@ -559,10 +552,14 @@ class RemoteExecutor(BaseExecutor): self.total_bundles_submitted = 0 self.backup_lock = threading.Lock() self.last_backup = None - (self.heartbeat_thread, self.heartbeat_stop_event) = self.run_periodic_heartbeat() + ( + self.heartbeat_thread, + self.heartbeat_stop_event, + ) = self.run_periodic_heartbeat() + self.already_shutdown = False @background_thread - def run_periodic_heartbeat(self, stop_event) -> None: + def run_periodic_heartbeat(self, stop_event: threading.Event) -> None: while not stop_event.is_set(): time.sleep(5.0) logger.debug('Running periodic heartbeat code...') @@ -570,8 +567,10 @@ class RemoteExecutor(BaseExecutor): logger.debug('Periodic heartbeat thread shutting down.') def heartbeat(self) -> None: + # Note: this is invoked on a background thread, not an + # executor thread. Be careful what you do with it b/c it + # needs to get back and dump status again periodically. with self.status.lock: - # Dump regular progress report self.status.periodic_dump(self.total_bundles_submitted) # Look for bundles to reschedule via executor.submit @@ -584,17 +583,20 @@ class RemoteExecutor(BaseExecutor): num_idle_workers = self.worker_count - self.task_count now = time.time() if ( - num_done > 2 - and num_idle_workers > 1 - and (self.last_backup is None or (now - self.last_backup > 6.0)) - and self.backup_lock.acquire(blocking=False) + num_done > 2 + and num_idle_workers > 1 + and (self.last_backup is None or (now - self.last_backup > 9.0)) + and self.backup_lock.acquire(blocking=False) ): try: assert self.backup_lock.locked() bundle_to_backup = None best_score = None - for worker, bundle_uuids in self.status.in_flight_bundles_by_worker.items(): + for ( + worker, + bundle_uuids, + ) in self.status.in_flight_bundles_by_worker.items(): # Prefer to schedule backups of bundles running on # slower machines. @@ -610,9 +612,9 @@ class RemoteExecutor(BaseExecutor): for uuid in bundle_uuids: bundle = self.status.bundle_details_by_uuid.get(uuid, None) if ( - bundle is not None - and bundle.src_bundle is None - and bundle.backup_bundles is not None + bundle is not None + and bundle.src_bundle is None + and bundle.backup_bundles is not None ): score = base_score @@ -623,15 +625,21 @@ class RemoteExecutor(BaseExecutor): if start_ts is not None: runtime = now - start_ts score += runtime - logger.debug(f'score[{bundle}] => {score} # latency boost') + logger.debug( + f'score[{bundle}] => {score} # latency boost' + ) if bundle.slower_than_local_p95: score += runtime / 2 - logger.debug(f'score[{bundle}] => {score} # >worker p95') + logger.debug( + f'score[{bundle}] => {score} # >worker p95' + ) if bundle.slower_than_global_p95: score += runtime / 4 - logger.debug(f'score[{bundle}] => {score} # >global p95') + logger.debug( + f'score[{bundle}] => {score} # >global p95' + ) # Prefer backups of bundles that don't # have backups already. @@ -648,9 +656,8 @@ class RemoteExecutor(BaseExecutor): f'score[{bundle}] => {score} # {backup_count} dup backup factor' ) - if ( - score != 0 - and (best_score is None or score > best_score) + if score != 0 and ( + best_score is None or score > best_score ): bundle_to_backup = bundle assert bundle is not None @@ -677,14 +684,12 @@ class RemoteExecutor(BaseExecutor): return self.policy.is_worker_available() def acquire_worker( - self, - machine_to_avoid: str = None + self, machine_to_avoid: str = None ) -> Optional[RemoteWorkerRecord]: return self.policy.acquire_worker(machine_to_avoid) def find_available_worker_or_block( - self, - machine_to_avoid: str = None + self, machine_to_avoid: str = None ) -> RemoteWorkerRecord: with self.cv: while not self.is_worker_available(): @@ -696,12 +701,7 @@ class RemoteExecutor(BaseExecutor): logger.critical(msg) raise Exception(msg) - def release_worker( - self, - bundle: BundleDetails, - *, - was_cancelled=True - ) -> None: + def release_worker(self, bundle: BundleDetails, *, was_cancelled=True) -> None: worker = bundle.worker assert worker is not None logger.debug(f'Released worker {worker}') @@ -768,8 +768,8 @@ class RemoteExecutor(BaseExecutor): # thing. logger.exception(e) logger.error( - f'{bundle}: We are the original owner thread and yet there are ' + - 'no results for this bundle. This is unexpected and bad.' + f'{bundle}: We are the original owner thread and yet there are ' + + 'no results for this bundle. This is unexpected and bad.' ) return self.emergency_retry_nasty_bundle(bundle) else: @@ -785,12 +785,14 @@ class RemoteExecutor(BaseExecutor): # Send input code / data to worker machine if it's not local. if hostname not in machine: try: - cmd = f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}' + cmd = ( + f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}' + ) start_ts = time.time() logger.info(f"{bundle}: Copying work to {worker} via {cmd}.") run_silently(cmd) xfer_latency = time.time() - start_ts - logger.info(f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s.") + logger.debug(f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s.") except Exception as e: self.release_worker(bundle) if is_original: @@ -798,9 +800,9 @@ class RemoteExecutor(BaseExecutor): # And we're the original bundle. We have to retry. logger.exception(e) logger.error( - f"{bundle}: Failed to send instructions to the worker machine?! " + - "This is not expected; we\'re the original bundle so this shouldn\'t " + - "be a race condition. Attempting an emergency retry..." + f"{bundle}: Failed to send instructions to the worker machine?! " + + "This is not expected; we\'re the original bundle so this shouldn\'t " + + "be a race condition. Attempting an emergency retry..." ) return self.emergency_retry_nasty_bundle(bundle) else: @@ -817,17 +819,23 @@ class RemoteExecutor(BaseExecutor): # Kick off the work. Note that if this fails we let # wait_for_process deal with it. self.status.record_processing_began(uuid) - cmd = (f'{SSH} {bundle.username}@{bundle.machine} ' - f'"source py38-venv/bin/activate &&' - f' /home/scott/lib/python_modules/remote_worker.py' - f' --code_file {bundle.code_file} --result_file {bundle.result_file}"') + cmd = ( + f'{SSH} {bundle.username}@{bundle.machine} ' + f'"source py38-venv/bin/activate &&' + f' /home/scott/lib/python_modules/remote_worker.py' + f' --code_file {bundle.code_file} --result_file {bundle.result_file}"' + ) logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...') p = cmd_in_background(cmd, silent=True) bundle.pid = p.pid - logger.debug(f'{bundle}: Local ssh process pid={p.pid}; remote worker is {machine}.') + logger.debug( + f'{bundle}: Local ssh process pid={p.pid}; remote worker is {machine}.' + ) return self.wait_for_process(p, bundle, 0) - def wait_for_process(self, p: subprocess.Popen, bundle: BundleDetails, depth: int) -> Any: + def wait_for_process( + self, p: subprocess.Popen, bundle: BundleDetails, depth: int + ) -> Any: machine = bundle.machine pid = p.pid if depth > 3: @@ -847,13 +855,11 @@ class RemoteExecutor(BaseExecutor): except subprocess.TimeoutExpired: if self.check_if_cancelled(bundle): logger.info( - f'{bundle}: another worker finished bundle, checking it out...' + f'{bundle}: looks like another worker finished bundle...' ) break else: - logger.info( - f"{bundle}: pid {pid} ({machine}) is finished!" - ) + logger.info(f"{bundle}: pid {pid} ({machine}) is finished!") p = None break @@ -913,13 +919,17 @@ class RemoteExecutor(BaseExecutor): except Exception as e: attempts += 1 if attempts >= 3: - raise(e) + raise (e) else: break - run_silently(f'{SSH} {username}@{machine}' - f' "/bin/rm -f {code_file} {result_file}"') - logger.debug(f'Fetching results back took {time.time() - bundle.end_ts:.1f}s.') + run_silently( + f'{SSH} {username}@{machine}' + f' "/bin/rm -f {code_file} {result_file}"' + ) + logger.debug( + f'Fetching results back took {time.time() - bundle.end_ts:.1f}s.' + ) dur = bundle.end_ts - bundle.start_ts self.histogram.add_item(dur) @@ -944,9 +954,7 @@ class RemoteExecutor(BaseExecutor): # Re-raise the exception; the code in wait_for_process may # decide to emergency_retry_nasty_bundle here. raise Exception(e) - logger.debug( - f'Removing local (master) {code_file} and {result_file}.' - ) + logger.debug(f'Removing local (master) {code_file} and {result_file}.') os.remove(f'{result_file}') os.remove(f'{code_file}') @@ -980,6 +988,7 @@ class RemoteExecutor(BaseExecutor): def create_original_bundle(self, pickle, fname: str): from string_utils import generate_uuid + uuid = generate_uuid(omit_dashes=True) code_file = f'/tmp/{uuid}.code.bin' result_file = f'/tmp/{uuid}.result.bin' @@ -989,25 +998,25 @@ class RemoteExecutor(BaseExecutor): wb.write(pickle) bundle = BundleDetails( - pickled_code = pickle, - uuid = uuid, - fname = fname, - worker = None, - username = None, - machine = None, - hostname = platform.node(), - code_file = code_file, - result_file = result_file, - pid = 0, - start_ts = time.time(), - end_ts = 0.0, - slower_than_local_p95 = False, - slower_than_global_p95 = False, - src_bundle = None, - is_cancelled = threading.Event(), - was_cancelled = False, - backup_bundles = [], - failure_count = 0, + pickled_code=pickle, + uuid=uuid, + fname=fname, + worker=None, + username=None, + machine=None, + hostname=platform.node(), + code_file=code_file, + result_file=result_file, + pid=0, + start_ts=time.time(), + end_ts=0.0, + slower_than_local_p95=False, + slower_than_global_p95=False, + src_bundle=None, + is_cancelled=threading.Event(), + was_cancelled=False, + backup_bundles=[], + failure_count=0, ) self.status.record_bundle_details(bundle) logger.debug(f'{bundle}: Created an original bundle') @@ -1019,33 +1028,32 @@ class RemoteExecutor(BaseExecutor): uuid = src_bundle.uuid + f'_backup#{n}' backup_bundle = BundleDetails( - pickled_code = src_bundle.pickled_code, - uuid = uuid, - fname = src_bundle.fname, - worker = None, - username = None, - machine = None, - hostname = src_bundle.hostname, - code_file = src_bundle.code_file, - result_file = src_bundle.result_file, - pid = 0, - start_ts = time.time(), - end_ts = 0.0, - slower_than_local_p95 = False, - slower_than_global_p95 = False, - src_bundle = src_bundle, - is_cancelled = threading.Event(), - was_cancelled = False, - backup_bundles = None, # backup backups not allowed - failure_count = 0, + pickled_code=src_bundle.pickled_code, + uuid=uuid, + fname=src_bundle.fname, + worker=None, + username=None, + machine=None, + hostname=src_bundle.hostname, + code_file=src_bundle.code_file, + result_file=src_bundle.result_file, + pid=0, + start_ts=time.time(), + end_ts=0.0, + slower_than_local_p95=False, + slower_than_global_p95=False, + src_bundle=src_bundle, + is_cancelled=threading.Event(), + was_cancelled=False, + backup_bundles=None, # backup backups not allowed + failure_count=0, ) src_bundle.backup_bundles.append(backup_bundle) self.status.record_bundle_details_already_locked(backup_bundle) logger.debug(f'{backup_bundle}: Created a backup bundle') return backup_bundle - def schedule_backup_for_bundle(self, - src_bundle: BundleDetails): + def schedule_backup_for_bundle(self, src_bundle: BundleDetails): assert self.status.lock.locked() assert src_bundle is not None backup_bundle = self.create_backup_bundle(src_bundle) @@ -1079,7 +1087,9 @@ class RemoteExecutor(BaseExecutor): f'{bundle}: This bundle can\'t be completed despite several backups and retries' ) else: - logger.error(f'{bundle}: At least it\'s only a backup; better luck with the others.') + logger.error( + f'{bundle}: At least it\'s only a backup; better luck with the others.' + ) return None else: msg = f'>>> Emergency rescheduling {bundle} because of unexected errors (wtf?!) <<<' @@ -1088,10 +1098,9 @@ class RemoteExecutor(BaseExecutor): return self.launch(bundle, avoid_last_machine) @overrides - def submit(self, - function: Callable, - *args, - **kwargs) -> fut.Future: + def submit(self, function: Callable, *args, **kwargs) -> fut.Future: + if self.already_shutdown: + raise Exception('Submitted work after shutdown.') pickle = make_cloud_pickle(function, *args, **kwargs) bundle = self.create_original_bundle(pickle, function.__name__) self.total_bundles_submitted += 1 @@ -1099,11 +1108,13 @@ class RemoteExecutor(BaseExecutor): @overrides def shutdown(self, wait=True) -> None: - logging.debug(f'Shutting down RemoteExecutor {self.title}') - self.heartbeat_stop_event.set() - self.heartbeat_thread.join() - self._helper_executor.shutdown(wait) - print(self.histogram) + if not self.already_shutdown: + logging.debug(f'Shutting down RemoteExecutor {self.title}') + self.heartbeat_stop_event.set() + self.heartbeat_thread.join() + self._helper_executor.shutdown(wait) + print(self.histogram) + self.already_shutdown = True @singleton @@ -1117,8 +1128,7 @@ class DefaultExecutors(object): logger.debug(f'RUN> ping -c 1 {host}') try: x = cmd_with_timeout( - f'ping -c 1 {host} >/dev/null 2>/dev/null', - timeout_seconds=1.0 + f'ping -c 1 {host} >/dev/null 2>/dev/null', timeout_seconds=1.0 ) return x == 0 except Exception: @@ -1142,50 +1152,50 @@ class DefaultExecutors(object): logger.info('Found cheetah.house') pool.append( RemoteWorkerRecord( - username = 'scott', - machine = 'cheetah.house', - weight = 25, - count = 6, + username='scott', + machine='cheetah.house', + weight=30, + count=6, ), ) if self.ping('meerkat.cabin'): logger.info('Found meerkat.cabin') pool.append( RemoteWorkerRecord( - username = 'scott', - machine = 'meerkat.cabin', - weight = 12, - count = 2, + username='scott', + machine='meerkat.cabin', + weight=12, + count=2, ), ) if self.ping('wannabe.house'): logger.info('Found wannabe.house') pool.append( RemoteWorkerRecord( - username = 'scott', - machine = 'wannabe.house', - weight = 30, - count = 10, + username='scott', + machine='wannabe.house', + weight=25, + count=10, ), ) if self.ping('puma.cabin'): logger.info('Found puma.cabin') pool.append( RemoteWorkerRecord( - username = 'scott', - machine = 'puma.cabin', - weight = 25, - count = 6, + username='scott', + machine='puma.cabin', + weight=30, + count=6, ), ) if self.ping('backup.house'): logger.info('Found backup.house') pool.append( RemoteWorkerRecord( - username = 'scott', - machine = 'backup.house', - weight = 7, - count = 2, + username='scott', + machine='backup.house', + weight=8, + count=2, ), )