3 from __future__ import annotations
5 from abc import ABC, abstractmethod
6 import concurrent.futures as fut
7 from collections import defaultdict
8 from dataclasses import dataclass
17 from typing import Any, Callable, Dict, List, Optional, Set
20 import cloudpickle # type: ignore
21 from overrides import overrides
23 from ansi import bg, fg, underline, reset
26 from decorator_utils import singleton
27 from exec_utils import run_silently, cmd_in_background, cmd_with_timeout
28 import histogram as hist
29 from thread_utils import background_thread
32 logger = logging.getLogger(__name__)
34 parser = config.add_commandline_args(
35 f"Executors ({__file__})", "Args related to processing executors."
38 '--executors_threadpool_size',
41 help='Number of threads in the default threadpool, leave unset for default',
45 '--executors_processpool_size',
48 help='Number of processes in the default processpool, leave unset for default',
52 '--executors_schedule_remote_backups',
54 action=argparse_utils.ActionNoYes,
55 help='Should we schedule duplicative backup work if a remote bundle is slow',
58 '--executors_max_bundle_failures',
62 help='Maximum number of failures before giving up on a bundle',
65 SSH = '/usr/bin/ssh -oForwardX11=no'
66 SCP = '/usr/bin/scp -C'
69 def make_cloud_pickle(fun, *args, **kwargs):
70 logger.debug(f"Making cloudpickled bundle at {fun.__name__}")
71 return cloudpickle.dumps((fun, args, kwargs))
74 class BaseExecutor(ABC):
75 def __init__(self, *, title=''):
77 self.histogram = hist.SimpleHistogram(
78 hist.SimpleHistogram.n_evenly_spaced_buckets(int(0), int(500), 50)
83 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
87 def shutdown(self, wait: bool = True) -> None:
90 def adjust_task_count(self, delta: int) -> None:
91 """Change the task count. Note: do not call this method from a
92 worker, it should only be called by the launcher process /
96 self.task_count += delta
97 print(f'(adjusted task count by {delta} to {self.task_count})')
98 logger.debug(f'Executor current task count is {self.task_count}')
100 def get_task_count(self) -> int:
101 """Change the task count. Note: do not call this method from a
102 worker, it should only be called by the launcher process /
106 return self.task_count
109 class ThreadExecutor(BaseExecutor):
110 def __init__(self, max_workers: Optional[int] = None):
113 if max_workers is not None:
114 workers = max_workers
115 elif 'executors_threadpool_size' in config.config:
116 workers = config.config['executors_threadpool_size']
117 logger.debug(f'Creating threadpool executor with {workers} workers')
118 self._thread_pool_executor = fut.ThreadPoolExecutor(
119 max_workers=workers, thread_name_prefix="thread_executor_helper"
122 # This is run on a different thread; do not adjust task count here.
123 def run_local_bundle(self, fun, *args, **kwargs):
124 logger.debug(f"Running local bundle at {fun.__name__}")
125 result = fun(*args, **kwargs)
129 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
130 self.adjust_task_count(+1)
132 newargs.append(function)
136 result = self._thread_pool_executor.submit(
137 self.run_local_bundle, *newargs, **kwargs
139 result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start))
140 result.add_done_callback(lambda _: self.adjust_task_count(-1))
143 def shutdown(self, wait=True) -> None:
144 logger.debug(f'Shutting down threadpool executor {self.title}')
145 print(self.histogram)
146 self._thread_pool_executor.shutdown(wait)
149 class ProcessExecutor(BaseExecutor):
150 def __init__(self, max_workers=None):
153 if max_workers is not None:
154 workers = max_workers
155 elif 'executors_processpool_size' in config.config:
156 workers = config.config['executors_processpool_size']
157 logger.debug(f'Creating processpool executor with {workers} workers.')
158 self._process_executor = fut.ProcessPoolExecutor(
162 # This is run in another process; do not adjust task count here.
163 def run_cloud_pickle(self, pickle):
164 fun, args, kwargs = cloudpickle.loads(pickle)
165 logger.debug(f"Running pickled bundle at {fun.__name__}")
166 result = fun(*args, **kwargs)
170 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
172 self.adjust_task_count(+1)
173 pickle = make_cloud_pickle(function, *args, **kwargs)
174 result = self._process_executor.submit(self.run_cloud_pickle, pickle)
175 result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start))
176 result.add_done_callback(lambda _: self.adjust_task_count(-1))
180 def shutdown(self, wait=True) -> None:
181 logger.debug(f'Shutting down processpool executor {self.title}')
182 self._process_executor.shutdown(wait)
183 print(self.histogram)
185 def __getstate__(self):
186 state = self.__dict__.copy()
187 state['_process_executor'] = None
191 class RemoteExecutorException(Exception):
192 """Thrown when a bundle cannot be executed despite several retries."""
198 class RemoteWorkerRecord:
205 return hash((self.username, self.machine))
208 return f'{self.username}@{self.machine}'
216 worker: Optional[RemoteWorkerRecord]
217 username: Optional[str]
218 machine: Optional[str]
225 slower_than_local_p95: bool
226 slower_than_global_p95: bool
227 src_bundle: BundleDetails
228 is_cancelled: threading.Event
230 backup_bundles: Optional[List[BundleDetails]]
235 if uuid[-9:-2] == '_backup':
237 suffix = f'{uuid[-6:]}_b{self.uuid[-1:]}'
247 fg('marigold yellow'),
250 fg('cornflower blue'),
251 fg('turquoise blue'),
253 fg('lavender purple'),
256 c = colorz[int(uuid[-2:], 16) % len(colorz)]
257 fname = self.fname if self.fname is not None else 'nofname'
258 machine = self.machine if self.machine is not None else 'nomachine'
259 return f'{c}{suffix}/{fname}/{machine}{reset()}'
262 class RemoteExecutorStatus:
263 def __init__(self, total_worker_count: int) -> None:
264 self.worker_count: int = total_worker_count
265 self.known_workers: Set[RemoteWorkerRecord] = set()
266 self.start_time: float = time.time()
267 self.start_per_bundle: Dict[str, float] = defaultdict(float)
268 self.end_per_bundle: Dict[str, float] = defaultdict(float)
269 self.finished_bundle_timings_per_worker: Dict[
270 RemoteWorkerRecord, List[float]
272 self.in_flight_bundles_by_worker: Dict[RemoteWorkerRecord, Set[str]] = {}
273 self.bundle_details_by_uuid: Dict[str, BundleDetails] = {}
274 self.finished_bundle_timings: List[float] = []
275 self.last_periodic_dump: Optional[float] = None
276 self.total_bundles_submitted: int = 0
278 # Protects reads and modification using self. Also used
279 # as a memory fence for modifications to bundle.
280 self.lock: threading.Lock = threading.Lock()
282 def record_acquire_worker(self, worker: RemoteWorkerRecord, uuid: str) -> None:
284 self.record_acquire_worker_already_locked(worker, uuid)
286 def record_acquire_worker_already_locked(
287 self, worker: RemoteWorkerRecord, uuid: str
289 assert self.lock.locked()
290 self.known_workers.add(worker)
291 self.start_per_bundle[uuid] = None
292 x = self.in_flight_bundles_by_worker.get(worker, set())
294 self.in_flight_bundles_by_worker[worker] = x
296 def record_bundle_details(self, details: BundleDetails) -> None:
298 self.record_bundle_details_already_locked(details)
300 def record_bundle_details_already_locked(self, details: BundleDetails) -> None:
301 assert self.lock.locked()
302 self.bundle_details_by_uuid[details.uuid] = details
304 def record_release_worker(
306 worker: RemoteWorkerRecord,
311 self.record_release_worker_already_locked(worker, uuid, was_cancelled)
313 def record_release_worker_already_locked(
315 worker: RemoteWorkerRecord,
319 assert self.lock.locked()
321 self.end_per_bundle[uuid] = ts
322 self.in_flight_bundles_by_worker[worker].remove(uuid)
323 if not was_cancelled:
324 bundle_latency = ts - self.start_per_bundle[uuid]
325 x = self.finished_bundle_timings_per_worker.get(worker, list())
326 x.append(bundle_latency)
327 self.finished_bundle_timings_per_worker[worker] = x
328 self.finished_bundle_timings.append(bundle_latency)
330 def record_processing_began(self, uuid: str):
332 self.start_per_bundle[uuid] = time.time()
334 def total_in_flight(self) -> int:
335 assert self.lock.locked()
337 for worker in self.known_workers:
338 total_in_flight += len(self.in_flight_bundles_by_worker[worker])
339 return total_in_flight
341 def total_idle(self) -> int:
342 assert self.lock.locked()
343 return self.worker_count - self.total_in_flight()
346 assert self.lock.locked()
348 total_finished = len(self.finished_bundle_timings)
349 total_in_flight = self.total_in_flight()
350 ret = f'\n\n{underline()}Remote Executor Pool Status{reset()}: '
352 if len(self.finished_bundle_timings) > 1:
353 qall = numpy.quantile(self.finished_bundle_timings, [0.5, 0.95])
355 f'⏱=∀p50:{qall[0]:.1f}s, ∀p95:{qall[1]:.1f}s, total={ts-self.start_time:.1f}s, '
356 f'✅={total_finished}/{self.total_bundles_submitted}, '
357 f'💻n={total_in_flight}/{self.worker_count}\n'
361 f'⏱={ts-self.start_time:.1f}s, '
362 f'✅={total_finished}/{self.total_bundles_submitted}, '
363 f'💻n={total_in_flight}/{self.worker_count}\n'
366 for worker in self.known_workers:
367 ret += f' {fg("lightning yellow")}{worker.machine}{reset()}: '
368 timings = self.finished_bundle_timings_per_worker.get(worker, [])
372 qworker = numpy.quantile(timings, [0.5, 0.95])
373 ret += f' 💻p50: {qworker[0]:.1f}s, 💻p95: {qworker[1]:.1f}s\n'
377 ret += f' ...finished {count} total bundle(s) so far\n'
378 in_flight = len(self.in_flight_bundles_by_worker[worker])
380 ret += f' ...{in_flight} bundles currently in flight:\n'
381 for bundle_uuid in self.in_flight_bundles_by_worker[worker]:
382 details = self.bundle_details_by_uuid.get(bundle_uuid, None)
383 pid = str(details.pid) if (details and details.pid != 0) else "TBD"
384 if self.start_per_bundle[bundle_uuid] is not None:
385 sec = ts - self.start_per_bundle[bundle_uuid]
386 ret += f' (pid={pid}): {details} for {sec:.1f}s so far '
388 ret += f' {details} setting up / copying data...'
391 if qworker is not None:
393 ret += f'{bg("red")}>💻p95{reset()} '
394 if details is not None:
395 details.slower_than_local_p95 = True
397 if details is not None:
398 details.slower_than_local_p95 = False
402 ret += f'{bg("red")}>∀p95{reset()} '
403 if details is not None:
404 details.slower_than_global_p95 = True
406 details.slower_than_global_p95 = False
410 def periodic_dump(self, total_bundles_submitted: int) -> None:
411 assert self.lock.locked()
412 self.total_bundles_submitted = total_bundles_submitted
414 if self.last_periodic_dump is None or ts - self.last_periodic_dump > 5.0:
416 self.last_periodic_dump = ts
419 class RemoteWorkerSelectionPolicy(ABC):
420 def register_worker_pool(self, workers):
421 self.workers = workers
424 def is_worker_available(self) -> bool:
428 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
432 class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
434 def is_worker_available(self) -> bool:
435 for worker in self.workers:
441 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
443 for worker in self.workers:
444 if worker.machine != machine_to_avoid:
446 for _ in range(worker.count * worker.weight):
447 grabbag.append(worker)
449 if len(grabbag) == 0:
451 f'There are no available workers that avoid {machine_to_avoid}...'
453 for worker in self.workers:
455 for _ in range(worker.count * worker.weight):
456 grabbag.append(worker)
458 if len(grabbag) == 0:
459 logger.warning('There are no available workers?!')
462 worker = random.sample(grabbag, 1)[0]
463 assert worker.count > 0
465 logger.debug(f'Chose worker {worker}')
469 class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
470 def __init__(self) -> None:
474 def is_worker_available(self) -> bool:
475 for worker in self.workers:
482 self, machine_to_avoid: str = None
483 ) -> Optional[RemoteWorkerRecord]:
486 worker = self.workers[x]
490 if x >= len(self.workers):
493 logger.debug(f'Selected worker {worker}')
496 if x >= len(self.workers):
499 msg = 'Unexpectedly could not find a worker, retrying...'
504 class RemoteExecutor(BaseExecutor):
507 workers: List[RemoteWorkerRecord],
508 policy: RemoteWorkerSelectionPolicy,
511 self.workers = workers
513 self.worker_count = 0
514 for worker in self.workers:
515 self.worker_count += worker.count
516 if self.worker_count <= 0:
517 msg = f"We need somewhere to schedule work; count was {self.worker_count}"
519 raise RemoteExecutorException(msg)
520 self.policy.register_worker_pool(self.workers)
521 self.cv = threading.Condition()
523 f'Creating {self.worker_count} local threads, one per remote worker.'
525 self._helper_executor = fut.ThreadPoolExecutor(
526 thread_name_prefix="remote_executor_helper",
527 max_workers=self.worker_count,
529 self.status = RemoteExecutorStatus(self.worker_count)
530 self.total_bundles_submitted = 0
531 self.backup_lock = threading.Lock()
532 self.last_backup = None
534 self.heartbeat_thread,
535 self.heartbeat_stop_event,
536 ) = self.run_periodic_heartbeat()
539 def run_periodic_heartbeat(self, stop_event: threading.Event) -> None:
540 while not stop_event.is_set():
542 logger.debug('Running periodic heartbeat code...')
544 logger.debug('Periodic heartbeat thread shutting down.')
546 def heartbeat(self) -> None:
547 # Note: this is invoked on a background thread, not an
548 # executor thread. Be careful what you do with it b/c it
549 # needs to get back and dump status again periodically.
550 with self.status.lock:
551 self.status.periodic_dump(self.total_bundles_submitted)
553 # Look for bundles to reschedule via executor.submit
554 if config.config['executors_schedule_remote_backups']:
555 self.maybe_schedule_backup_bundles()
557 def maybe_schedule_backup_bundles(self):
558 assert self.status.lock.locked()
559 num_done = len(self.status.finished_bundle_timings)
560 num_idle_workers = self.worker_count - self.task_count
564 and num_idle_workers > 1
565 and (self.last_backup is None or (now - self.last_backup > 9.0))
566 and self.backup_lock.acquire(blocking=False)
569 assert self.backup_lock.locked()
571 bundle_to_backup = None
576 ) in self.status.in_flight_bundles_by_worker.items():
578 # Prefer to schedule backups of bundles running on
581 for record in self.workers:
582 if worker.machine == record.machine:
583 base_score = float(record.weight)
584 base_score = 1.0 / base_score
586 base_score = int(base_score)
589 for uuid in bundle_uuids:
590 bundle = self.status.bundle_details_by_uuid.get(uuid, None)
593 and bundle.src_bundle is None
594 and bundle.backup_bundles is not None
598 # Schedule backups of bundles running
599 # longer; especially those that are
601 start_ts = self.status.start_per_bundle[uuid]
602 if start_ts is not None:
603 runtime = now - start_ts
606 f'score[{bundle}] => {score} # latency boost'
609 if bundle.slower_than_local_p95:
612 f'score[{bundle}] => {score} # >worker p95'
615 if bundle.slower_than_global_p95:
618 f'score[{bundle}] => {score} # >global p95'
621 # Prefer backups of bundles that don't
622 # have backups already.
623 backup_count = len(bundle.backup_bundles)
624 if backup_count == 0:
626 elif backup_count == 1:
628 elif backup_count == 2:
633 f'score[{bundle}] => {score} # {backup_count} dup backup factor'
637 best_score is None or score > best_score
639 bundle_to_backup = bundle
640 assert bundle is not None
641 assert bundle.backup_bundles is not None
642 assert bundle.src_bundle is None
645 # Note: this is all still happening on the heartbeat
646 # runner thread. That's ok because
647 # schedule_backup_for_bundle uses the executor to
648 # submit the bundle again which will cause it to be
649 # picked up by a worker thread and allow this thread
650 # to return to run future heartbeats.
651 if bundle_to_backup is not None:
652 self.last_backup = now
654 f'=====> SCHEDULING BACKUP {bundle_to_backup} (score={best_score:.1f}) <====='
656 self.schedule_backup_for_bundle(bundle_to_backup)
658 self.backup_lock.release()
660 def is_worker_available(self) -> bool:
661 return self.policy.is_worker_available()
664 self, machine_to_avoid: str = None
665 ) -> Optional[RemoteWorkerRecord]:
666 return self.policy.acquire_worker(machine_to_avoid)
668 def find_available_worker_or_block(
669 self, machine_to_avoid: str = None
670 ) -> RemoteWorkerRecord:
672 while not self.is_worker_available():
674 worker = self.acquire_worker(machine_to_avoid)
675 if worker is not None:
677 msg = "We should never reach this point in the code"
681 def release_worker(self, bundle: BundleDetails, *, was_cancelled=True) -> None:
682 worker = bundle.worker
683 assert worker is not None
684 logger.debug(f'Released worker {worker}')
685 self.status.record_release_worker(
693 self.adjust_task_count(-1)
695 def check_if_cancelled(self, bundle: BundleDetails) -> bool:
696 with self.status.lock:
697 if bundle.is_cancelled.wait(timeout=0.0):
698 logger.debug(f'Bundle {bundle.uuid} is cancelled, bail out.')
699 bundle.was_cancelled = True
703 def launch(self, bundle: BundleDetails, override_avoid_machine=None) -> Any:
704 """Find a worker for bundle or block until one is available."""
705 self.adjust_task_count(+1)
707 hostname = bundle.hostname
708 avoid_machine = override_avoid_machine
709 is_original = bundle.src_bundle is None
711 # Try not to schedule a backup on the same host as the original.
712 if avoid_machine is None and bundle.src_bundle is not None:
713 avoid_machine = bundle.src_bundle.machine
715 while worker is None:
716 worker = self.find_available_worker_or_block(avoid_machine)
719 # Ok, found a worker.
720 bundle.worker = worker
721 machine = bundle.machine = worker.machine
722 username = bundle.username = worker.username
723 self.status.record_acquire_worker(worker, uuid)
724 logger.debug(f'{bundle}: Running bundle on {worker}...')
726 # Before we do any work, make sure the bundle is still viable.
727 # It may have been some time between when it was submitted and
728 # now due to lack of worker availability and someone else may
729 # have already finished it.
730 if self.check_if_cancelled(bundle):
732 return self.process_work_result(bundle)
733 except Exception as e:
735 f'{bundle}: bundle says it\'s cancelled upfront but no results?!'
737 self.release_worker(bundle)
739 # Weird. We are the original owner of this
740 # bundle. For it to have been cancelled, a backup
741 # must have already started and completed before
742 # we even for started. Moreover, the backup says
743 # it is done but we can't find the results it
744 # should have copied over. Reschedule the whole
748 f'{bundle}: We are the original owner thread and yet there are '
749 + 'no results for this bundle. This is unexpected and bad.'
751 return self.emergency_retry_nasty_bundle(bundle)
753 # Expected(?). We're a backup and our bundle is
754 # cancelled before we even got started. Something
755 # went bad in process_work_result (I acutually don't
756 # see what?) but probably not worth worrying
757 # about. Let the original thread worry about
758 # either finding the results or complaining about
762 # Send input code / data to worker machine if it's not local.
763 if hostname not in machine:
766 f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}'
768 start_ts = time.time()
769 logger.info(f"{bundle}: Copying work to {worker} via {cmd}.")
771 xfer_latency = time.time() - start_ts
772 logger.debug(f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s.")
773 except Exception as e:
774 self.release_worker(bundle)
776 # Weird. We tried to copy the code to the worker and it failed...
777 # And we're the original bundle. We have to retry.
780 f"{bundle}: Failed to send instructions to the worker machine?! "
781 + "This is not expected; we\'re the original bundle so this shouldn\'t "
782 + "be a race condition. Attempting an emergency retry..."
784 return self.emergency_retry_nasty_bundle(bundle)
786 # This is actually expected; we're a backup.
787 # There's a race condition where someone else
788 # already finished the work and removed the source
789 # code file before we could copy it. No biggie.
790 msg = f'{bundle}: Failed to send instructions to the worker machine... '
791 msg += 'We\'re a backup and this may be caused by the original (or some '
792 msg += 'other backup) already finishing this work. Ignoring this.'
796 # Kick off the work. Note that if this fails we let
797 # wait_for_process deal with it.
798 self.status.record_processing_began(uuid)
800 f'{SSH} {bundle.username}@{bundle.machine} '
801 f'"source py38-venv/bin/activate &&'
802 f' /home/scott/lib/python_modules/remote_worker.py'
803 f' --code_file {bundle.code_file} --result_file {bundle.result_file}"'
805 logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...')
806 p = cmd_in_background(cmd, silent=True)
809 f'{bundle}: Local ssh process pid={p.pid}; remote worker is {machine}.'
811 return self.wait_for_process(p, bundle, 0)
813 def wait_for_process(
814 self, p: subprocess.Popen, bundle: BundleDetails, depth: int
816 machine = bundle.machine
820 f"I've gotten repeated errors waiting on this bundle; giving up on pid={pid}."
823 self.release_worker(bundle)
824 return self.emergency_retry_nasty_bundle(bundle)
826 # Spin until either the ssh job we scheduled finishes the
827 # bundle or some backup worker signals that they finished it
832 except subprocess.TimeoutExpired:
833 if self.check_if_cancelled(bundle):
835 f'{bundle}: looks like another worker finished bundle...'
839 logger.info(f"{bundle}: pid {pid} ({machine}) is finished!")
843 # If we get here we believe the bundle is done; either the ssh
844 # subprocess finished (hopefully successfully) or we noticed
845 # that some other worker seems to have completed the bundle
846 # and we're bailing out.
848 ret = self.process_work_result(bundle)
849 if ret is not None and p is not None:
853 # Something went wrong; e.g. we could not copy the results
854 # back, cleanup after ourselves on the remote machine, or
855 # unpickle the results we got from the remove machine. If we
856 # still have an active ssh subprocess, keep waiting on it.
857 # Otherwise, time for an emergency reschedule.
858 except Exception as e:
860 logger.error(f'{bundle}: Something unexpected just happened...')
862 msg = f"{bundle}: Failed to wrap up \"done\" bundle, re-waiting on active ssh."
864 return self.wait_for_process(p, bundle, depth + 1)
866 self.release_worker(bundle)
867 return self.emergency_retry_nasty_bundle(bundle)
869 def process_work_result(self, bundle: BundleDetails) -> Any:
870 with self.status.lock:
871 is_original = bundle.src_bundle is None
872 was_cancelled = bundle.was_cancelled
873 username = bundle.username
874 machine = bundle.machine
875 result_file = bundle.result_file
876 code_file = bundle.code_file
878 # Whether original or backup, if we finished first we must
879 # fetch the results if the computation happened on a
881 bundle.end_ts = time.time()
882 if not was_cancelled:
883 assert bundle.machine is not None
884 if bundle.hostname not in bundle.machine:
885 cmd = f'{SCP} {username}@{machine}:{result_file} {result_file} 2>/dev/null'
887 f"{bundle}: Fetching results back from {username}@{machine} via {cmd}"
890 # If either of these throw they are handled in
896 except Exception as e:
904 f'{SSH} {username}@{machine}'
905 f' "/bin/rm -f {code_file} {result_file}"'
908 f'Fetching results back took {time.time() - bundle.end_ts:.1f}s.'
910 dur = bundle.end_ts - bundle.start_ts
911 self.histogram.add_item(dur)
913 # Only the original worker should unpickle the file contents
914 # though since it's the only one whose result matters. The
915 # original is also the only job that may delete result_file
916 # from disk. Note that the original may have been cancelled
917 # if one of the backups finished first; it still must read the
920 logger.debug(f"{bundle}: Unpickling {result_file}.")
922 with open(result_file, 'rb') as rb:
923 serialized = rb.read()
924 result = cloudpickle.loads(serialized)
925 except Exception as e:
927 msg = f'Failed to load {result_file}... this is bad news.'
929 self.release_worker(bundle)
931 # Re-raise the exception; the code in wait_for_process may
932 # decide to emergency_retry_nasty_bundle here.
934 logger.debug(f'Removing local (master) {code_file} and {result_file}.')
935 os.remove(f'{result_file}')
936 os.remove(f'{code_file}')
938 # Notify any backups that the original is done so they
939 # should stop ASAP. Do this whether or not we
940 # finished first since there could be more than one
942 if bundle.backup_bundles is not None:
943 for backup in bundle.backup_bundles:
945 f'{bundle}: Notifying backup {backup.uuid} that it\'s cancelled'
947 backup.is_cancelled.set()
949 # This is a backup job and, by now, we have already fetched
950 # the bundle results.
952 # Backup results don't matter, they just need to leave the
953 # result file in the right place for their originals to
954 # read/unpickle later.
957 # Tell the original to stop if we finished first.
958 if not was_cancelled:
960 f'{bundle}: Notifying original {bundle.src_bundle.uuid} we beat them to it.'
962 bundle.src_bundle.is_cancelled.set()
963 self.release_worker(bundle, was_cancelled=was_cancelled)
966 def create_original_bundle(self, pickle, fname: str):
967 from string_utils import generate_uuid
969 uuid = generate_uuid(omit_dashes=True)
970 code_file = f'/tmp/{uuid}.code.bin'
971 result_file = f'/tmp/{uuid}.result.bin'
973 logger.debug(f'Writing pickled code to {code_file}')
974 with open(f'{code_file}', 'wb') as wb:
977 bundle = BundleDetails(
984 hostname=platform.node(),
986 result_file=result_file,
988 start_ts=time.time(),
990 slower_than_local_p95=False,
991 slower_than_global_p95=False,
993 is_cancelled=threading.Event(),
998 self.status.record_bundle_details(bundle)
999 logger.debug(f'{bundle}: Created an original bundle')
1002 def create_backup_bundle(self, src_bundle: BundleDetails):
1003 assert src_bundle.backup_bundles is not None
1004 n = len(src_bundle.backup_bundles)
1005 uuid = src_bundle.uuid + f'_backup#{n}'
1007 backup_bundle = BundleDetails(
1008 pickled_code=src_bundle.pickled_code,
1010 fname=src_bundle.fname,
1014 hostname=src_bundle.hostname,
1015 code_file=src_bundle.code_file,
1016 result_file=src_bundle.result_file,
1018 start_ts=time.time(),
1020 slower_than_local_p95=False,
1021 slower_than_global_p95=False,
1022 src_bundle=src_bundle,
1023 is_cancelled=threading.Event(),
1024 was_cancelled=False,
1025 backup_bundles=None, # backup backups not allowed
1028 src_bundle.backup_bundles.append(backup_bundle)
1029 self.status.record_bundle_details_already_locked(backup_bundle)
1030 logger.debug(f'{backup_bundle}: Created a backup bundle')
1031 return backup_bundle
1033 def schedule_backup_for_bundle(self, src_bundle: BundleDetails):
1034 assert self.status.lock.locked()
1035 assert src_bundle is not None
1036 backup_bundle = self.create_backup_bundle(src_bundle)
1038 f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...'
1040 self._helper_executor.submit(self.launch, backup_bundle)
1042 # Results from backups don't matter; if they finish first
1043 # they will move the result_file to this machine and let
1044 # the original pick them up and unpickle them.
1046 def emergency_retry_nasty_bundle(self, bundle: BundleDetails) -> fut.Future:
1047 is_original = bundle.src_bundle is None
1048 bundle.worker = None
1049 avoid_last_machine = bundle.machine
1050 bundle.machine = None
1051 bundle.username = None
1052 bundle.failure_count += 1
1058 if bundle.failure_count > retry_limit:
1060 f'{bundle}: Tried this bundle too many times already ({retry_limit}x); giving up.'
1063 raise RemoteExecutorException(
1064 f'{bundle}: This bundle can\'t be completed despite several backups and retries'
1068 f'{bundle}: At least it\'s only a backup; better luck with the others.'
1072 msg = f'>>> Emergency rescheduling {bundle} because of unexected errors (wtf?!) <<<'
1075 return self.launch(bundle, avoid_last_machine)
1078 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
1079 pickle = make_cloud_pickle(function, *args, **kwargs)
1080 bundle = self.create_original_bundle(pickle, function.__name__)
1081 self.total_bundles_submitted += 1
1082 return self._helper_executor.submit(self.launch, bundle)
1085 def shutdown(self, wait=True) -> None:
1086 logging.debug(f'Shutting down RemoteExecutor {self.title}')
1087 self.heartbeat_stop_event.set()
1088 self.heartbeat_thread.join()
1089 self._helper_executor.shutdown(wait)
1090 print(self.histogram)
1094 class DefaultExecutors(object):
1096 self.thread_executor: Optional[ThreadExecutor] = None
1097 self.process_executor: Optional[ProcessExecutor] = None
1098 self.remote_executor: Optional[RemoteExecutor] = None
1100 def ping(self, host) -> bool:
1101 logger.debug(f'RUN> ping -c 1 {host}')
1103 x = cmd_with_timeout(
1104 f'ping -c 1 {host} >/dev/null 2>/dev/null', timeout_seconds=1.0
1110 def thread_pool(self) -> ThreadExecutor:
1111 if self.thread_executor is None:
1112 self.thread_executor = ThreadExecutor()
1113 return self.thread_executor
1115 def process_pool(self) -> ProcessExecutor:
1116 if self.process_executor is None:
1117 self.process_executor = ProcessExecutor()
1118 return self.process_executor
1120 def remote_pool(self) -> RemoteExecutor:
1121 if self.remote_executor is None:
1122 logger.info('Looking for some helper machines...')
1123 pool: List[RemoteWorkerRecord] = []
1124 if self.ping('cheetah.house'):
1125 logger.info('Found cheetah.house')
1129 machine='cheetah.house',
1134 if self.ping('meerkat.cabin'):
1135 logger.info('Found meerkat.cabin')
1139 machine='meerkat.cabin',
1144 if self.ping('wannabe.house'):
1145 logger.info('Found wannabe.house')
1149 machine='wannabe.house',
1154 if self.ping('puma.cabin'):
1155 logger.info('Found puma.cabin')
1159 machine='puma.cabin',
1164 if self.ping('backup.house'):
1165 logger.info('Found backup.house')
1169 machine='backup.house',
1175 # The controller machine has a lot to do; go easy on it.
1177 if record.machine == platform.node() and record.count > 1:
1178 logger.info(f'Reducing workload for {record.machine}.')
1181 policy = WeightedRandomRemoteWorkerSelectionPolicy()
1182 policy.register_worker_pool(pool)
1183 self.remote_executor = RemoteExecutor(pool, policy)
1184 return self.remote_executor
1186 def shutdown(self) -> None:
1187 if self.thread_executor is not None:
1188 self.thread_executor.shutdown()
1189 self.thread_executor = None
1190 if self.process_executor is not None:
1191 self.process_executor.shutdown()
1192 self.process_executor = None
1193 if self.remote_executor is not None:
1194 self.remote_executor.shutdown()
1195 self.remote_executor = None