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
19 import cloudpickle # type: ignore
20 from overrides import overrides
22 from ansi import bg, fg, underline, reset
25 from exec_utils import run_silently, cmd_in_background, cmd_with_timeout
26 from decorator_utils import singleton
27 import histogram as hist
29 logger = logging.getLogger(__name__)
31 parser = config.add_commandline_args(
32 f"Executors ({__file__})",
33 "Args related to processing executors."
36 '--executors_threadpool_size',
39 help='Number of threads in the default threadpool, leave unset for default',
43 '--executors_processpool_size',
46 help='Number of processes in the default processpool, leave unset for default',
50 '--executors_schedule_remote_backups',
52 action=argparse_utils.ActionNoYes,
53 help='Should we schedule duplicative backup work if a remote bundle is slow',
56 '--executors_max_bundle_failures',
60 help='Maximum number of failures before giving up on a bundle',
63 RSYNC = 'rsync -q --no-motd -W --ignore-existing --timeout=60 --size-only -z'
64 SSH = 'ssh -oForwardX11=no'
67 def make_cloud_pickle(fun, *args, **kwargs):
68 logger.debug(f"Making cloudpickled bundle at {fun.__name__}")
69 return cloudpickle.dumps((fun, args, kwargs))
72 class BaseExecutor(ABC):
73 def __init__(self, *, title=''):
76 self.histogram = hist.SimpleHistogram(
77 hist.SimpleHistogram.n_evenly_spaced_buckets(
86 **kwargs) -> fut.Future:
91 wait: bool = True) -> None:
94 def adjust_task_count(self, delta: int) -> None:
95 self.task_count += delta
96 logger.debug(f'Executor current task count is {self.task_count}')
99 class ThreadExecutor(BaseExecutor):
101 max_workers: Optional[int] = None):
104 if max_workers is not None:
105 workers = max_workers
106 elif 'executors_threadpool_size' in config.config:
107 workers = config.config['executors_threadpool_size']
108 logger.debug(f'Creating threadpool executor with {workers} workers')
109 self._thread_pool_executor = fut.ThreadPoolExecutor(
111 thread_name_prefix="thread_executor_helper"
114 def run_local_bundle(self, fun, *args, **kwargs):
115 logger.debug(f"Running local bundle at {fun.__name__}")
117 result = fun(*args, **kwargs)
119 self.adjust_task_count(-1)
120 duration = end - start
121 logger.debug(f"{fun.__name__} finished; used {duration:.1f}s")
122 self.histogram.add_item(duration)
129 **kwargs) -> fut.Future:
130 self.adjust_task_count(+1)
132 newargs.append(function)
135 return self._thread_pool_executor.submit(
136 self.run_local_bundle,
142 wait = True) -> None:
143 logger.debug(f'Shutting down threadpool executor {self.title}')
144 print(self.histogram)
145 self._thread_pool_executor.shutdown(wait)
148 class ProcessExecutor(BaseExecutor):
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 def run_cloud_pickle(self, pickle):
163 fun, args, kwargs = cloudpickle.loads(pickle)
164 logger.debug(f"Running pickled bundle at {fun.__name__}")
165 result = fun(*args, **kwargs)
166 self.adjust_task_count(-1)
173 **kwargs) -> fut.Future:
175 self.adjust_task_count(+1)
176 pickle = make_cloud_pickle(function, *args, **kwargs)
177 result = self._process_executor.submit(
178 self.run_cloud_pickle,
181 result.add_done_callback(
182 lambda _: self.histogram.add_item(
189 def shutdown(self, wait=True) -> None:
190 logger.debug(f'Shutting down processpool executor {self.title}')
191 self._process_executor.shutdown(wait)
192 print(self.histogram)
194 def __getstate__(self):
195 state = self.__dict__.copy()
196 state['_process_executor'] = None
200 class RemoteExecutorException(Exception):
201 """Thrown when a bundle cannot be executed despite several retries."""
206 class RemoteWorkerRecord:
213 return hash((self.username, self.machine))
216 return f'{self.username}@{self.machine}'
224 worker: Optional[RemoteWorkerRecord]
225 username: Optional[str]
226 machine: Optional[str]
235 src_bundle: BundleDetails
236 is_cancelled: threading.Event
238 backup_bundles: Optional[List[BundleDetails]]
242 class RemoteExecutorStatus:
243 def __init__(self, total_worker_count: int) -> None:
244 self.worker_count = total_worker_count
245 self.known_workers: Set[RemoteWorkerRecord] = set()
246 self.start_per_bundle: Dict[str, float] = defaultdict(float)
247 self.end_per_bundle: Dict[str, float] = defaultdict(float)
248 self.finished_bundle_timings_per_worker: Dict[
252 self.in_flight_bundles_by_worker: Dict[
256 self.bundle_details_by_uuid: Dict[str, BundleDetails] = {}
257 self.finished_bundle_timings: List[float] = []
258 self.last_periodic_dump: Optional[float] = None
259 self.total_bundles_submitted = 0
261 # Protects reads and modification using self. Also used
262 # as a memory fence for modifications to bundle.
263 self.lock = threading.Lock()
265 def record_acquire_worker(
267 worker: RemoteWorkerRecord,
271 self.record_acquire_worker_already_locked(
276 def record_acquire_worker_already_locked(
278 worker: RemoteWorkerRecord,
281 assert self.lock.locked()
282 self.known_workers.add(worker)
283 self.start_per_bundle[uuid] = None
284 x = self.in_flight_bundles_by_worker.get(worker, set())
286 self.in_flight_bundles_by_worker[worker] = x
288 def record_bundle_details(
290 details: BundleDetails) -> None:
292 self.record_bundle_details_already_locked(details)
294 def record_bundle_details_already_locked(
296 details: BundleDetails) -> None:
297 assert self.lock.locked()
298 self.bundle_details_by_uuid[details.uuid] = details
300 def record_release_worker(
302 worker: RemoteWorkerRecord,
307 self.record_release_worker_already_locked(
308 worker, uuid, was_cancelled
311 def record_release_worker_already_locked(
313 worker: RemoteWorkerRecord,
317 assert self.lock.locked()
319 self.end_per_bundle[uuid] = ts
320 self.in_flight_bundles_by_worker[worker].remove(uuid)
321 if not was_cancelled:
322 bundle_latency = ts - self.start_per_bundle[uuid]
323 x = self.finished_bundle_timings_per_worker.get(worker, list())
324 x.append(bundle_latency)
325 self.finished_bundle_timings_per_worker[worker] = x
326 self.finished_bundle_timings.append(bundle_latency)
328 def record_processing_began(self, uuid: str):
330 self.start_per_bundle[uuid] = time.time()
332 def total_in_flight(self) -> int:
333 assert self.lock.locked()
335 for worker in self.known_workers:
336 total_in_flight += len(self.in_flight_bundles_by_worker[worker])
337 return total_in_flight
339 def total_idle(self) -> int:
340 assert self.lock.locked()
341 return self.worker_count - self.total_in_flight()
344 assert self.lock.locked()
346 total_finished = len(self.finished_bundle_timings)
347 total_in_flight = self.total_in_flight()
348 ret = f'\n\n{underline()}Remote Executor Pool Status{reset()}: '
350 if len(self.finished_bundle_timings) > 1:
351 qall = numpy.quantile(self.finished_bundle_timings, [0.5, 0.95])
353 f'⏱=∀p50:{qall[0]:.1f}s, ∀p95:{qall[1]:.1f}s, '
354 f'✅={total_finished}/{self.total_bundles_submitted}, '
355 f'💻n={total_in_flight}/{self.worker_count}\n'
359 f' ✅={total_finished}/{self.total_bundles_submitted}, '
360 f'💻n={total_in_flight}/{self.worker_count}\n'
363 for worker in self.known_workers:
364 ret += f' {fg("lightning yellow")}{worker.machine}{reset()}: '
365 timings = self.finished_bundle_timings_per_worker.get(worker, [])
369 qworker = numpy.quantile(timings, [0.5, 0.95])
370 ret += f' 💻p50: {qworker[0]:.1f}s, 💻p95: {qworker[1]:.1f}s\n'
374 ret += f' ...finished {count} total bundle(s) so far\n'
375 in_flight = len(self.in_flight_bundles_by_worker[worker])
377 ret += f' ...{in_flight} bundles currently in flight:\n'
378 for bundle_uuid in self.in_flight_bundles_by_worker[worker]:
379 details = self.bundle_details_by_uuid.get(
383 pid = str(details.pid) if details is not None 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}): {bundle_uuid} for {sec:.1f}s so far '
388 ret += f' {bundle_uuid} setting up / copying data...'
391 if qworker is not None:
393 ret += f'{bg("red")}>💻p95{reset()} '
394 elif sec > qworker[0]:
395 ret += f'{fg("red")}>💻p50{reset()} '
397 if sec > qall[1] * 1.5:
398 ret += f'{bg("red")}!!!{reset()}'
399 if details is not None:
400 logger.debug(f'Flagging {details.uuid} for another backup')
401 details.super_slow = True
403 ret += f'{bg("red")}>∀p95{reset()} '
404 if details is not None:
405 logger.debug(f'Flagging {details.uuid} for a backup')
406 details.too_slow = True
408 ret += f'{fg("red")}>∀p50{reset()}'
412 def periodic_dump(self, total_bundles_submitted: int) -> None:
413 assert self.lock.locked()
414 self.total_bundles_submitted = total_bundles_submitted
417 self.last_periodic_dump is None
418 or ts - self.last_periodic_dump > 5.0
421 self.last_periodic_dump = ts
424 class RemoteWorkerSelectionPolicy(ABC):
425 def register_worker_pool(self, workers):
427 self.workers = workers
430 def is_worker_available(self) -> bool:
436 machine_to_avoid = None
437 ) -> Optional[RemoteWorkerRecord]:
441 class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
442 def is_worker_available(self) -> bool:
443 for worker in self.workers:
450 machine_to_avoid = None
451 ) -> Optional[RemoteWorkerRecord]:
453 for worker in self.workers:
454 for x in range(0, worker.count):
455 for y in range(0, worker.weight):
456 grabbag.append(worker)
458 for _ in range(0, 5):
459 random.shuffle(grabbag)
461 if worker.machine != machine_to_avoid or _ > 2:
464 logger.debug(f'Selected worker {worker}')
466 logger.warning("Couldn't find a worker; go fish.")
470 class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
471 def __init__(self) -> None:
474 def is_worker_available(self) -> bool:
475 for worker in self.workers:
482 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 logger.warning("Couldn't find a worker; go fish.")
503 class RemoteExecutor(BaseExecutor):
505 workers: List[RemoteWorkerRecord],
506 policy: RemoteWorkerSelectionPolicy) -> None:
508 self.workers = workers
510 self.worker_count = 0
511 for worker in self.workers:
512 self.worker_count += worker.count
513 if self.worker_count <= 0:
514 msg = f"We need somewhere to schedule work; count was {self.worker_count}"
516 raise RemoteExecutorException(msg)
517 self.policy.register_worker_pool(self.workers)
518 self.cv = threading.Condition()
519 logger.debug(f'Creating {self.worker_count} local threads, one per remote worker.')
520 self._helper_executor = fut.ThreadPoolExecutor(
521 thread_name_prefix="remote_executor_helper",
522 max_workers=self.worker_count,
524 self.status = RemoteExecutorStatus(self.worker_count)
525 self.total_bundles_submitted = 0
527 def bundle_prefix(self, bundle: BundleDetails) -> str:
534 fg('marigold yellow'),
537 fg('cornflower blue'),
538 fg('turquoise blue'),
540 fg('lavender purple'),
543 c = colorz[int(bundle.uuid[-2:], 16) % len(colorz)]
544 fname = bundle.fname if bundle.fname is not None else 'nofname'
545 machine = bundle.machine if bundle.machine is not None else 'nomachine'
546 return f'{c}{bundle.uuid[-8:]}/{fname}/{machine}{reset()}'
548 def is_worker_available(self) -> bool:
549 return self.policy.is_worker_available()
553 machine_to_avoid: str = None
554 ) -> Optional[RemoteWorkerRecord]:
555 return self.policy.acquire_worker(machine_to_avoid)
557 def find_available_worker_or_block(
559 machine_to_avoid: str = None
560 ) -> RemoteWorkerRecord:
562 while not self.is_worker_available():
564 worker = self.acquire_worker(machine_to_avoid)
565 if worker is not None:
567 msg = "We should never reach this point in the code"
571 def release_worker(self, worker: RemoteWorkerRecord) -> None:
572 logger.debug(f'Released worker {worker}')
577 def heartbeat(self) -> None:
578 with self.status.lock:
579 # Regular progress report
580 self.status.periodic_dump(self.total_bundles_submitted)
582 # Look for bundles to reschedule.
583 num_done = len(self.status.finished_bundle_timings)
584 if num_done > 7 or (num_done > 5 and self.is_worker_available()):
585 for worker, bundle_uuids in self.status.in_flight_bundles_by_worker.items():
586 for uuid in bundle_uuids:
587 bundle = self.status.bundle_details_by_uuid.get(uuid, None)
589 bundle is not None and
591 bundle.src_bundle is None and
592 config.config['executors_schedule_remote_backups']
594 self.consider_backup_for_bundle(bundle)
596 def consider_backup_for_bundle(self, bundle: BundleDetails) -> None:
597 assert self.status.lock.locked()
600 and len(bundle.backup_bundles) == 0 # one backup per
602 msg = f"*** Rescheduling {bundle.pid}/{bundle.uuid} ***"
604 self.schedule_backup_for_bundle(bundle)
608 and len(bundle.backup_bundles) < 2 # two backups in dire situations
609 and self.status.total_idle() > 4
611 msg = f"*** Rescheduling {bundle.pid}/{bundle.uuid} ***"
613 self.schedule_backup_for_bundle(bundle)
616 def check_if_cancelled(self, bundle: BundleDetails) -> bool:
617 with self.status.lock:
618 if bundle.is_cancelled.wait(timeout=0.0):
619 logger.debug(f'Bundle {bundle.uuid} is cancelled, bail out.')
620 bundle.was_cancelled = True
624 def launch(self, bundle: BundleDetails, override_avoid_machine=None) -> Any:
625 """Find a worker for bundle or block until one is available."""
626 self.adjust_task_count(+1)
628 hostname = bundle.hostname
629 avoid_machine = override_avoid_machine
630 is_original = bundle.src_bundle is None
632 # Try not to schedule a backup on the same host as the original.
633 if avoid_machine is None and bundle.src_bundle is not None:
634 avoid_machine = bundle.src_bundle.machine
636 while worker is None:
637 worker = self.find_available_worker_or_block(avoid_machine)
639 # Ok, found a worker.
640 bundle.worker = worker
641 machine = bundle.machine = worker.machine
642 username = bundle.username = worker.username
644 self.status.record_acquire_worker(worker, uuid)
645 logger.debug(f'{self.bundle_prefix(bundle)}: Running bundle on {worker}...')
647 # Before we do any work, make sure the bundle is still viable.
648 if self.check_if_cancelled(bundle):
650 return self.post_launch_work(bundle)
651 except Exception as e:
654 f'{self.bundle_prefix(bundle)}: bundle says it\'s cancelled upfront but no results?!'
656 assert bundle.worker is not None
657 self.status.record_release_worker(
662 self.release_worker(bundle.worker)
663 self.adjust_task_count(-1)
665 # Weird. We are the original owner of this
666 # bundle. For it to have been cancelled, a backup
667 # must have already started and completed before
668 # we even for started. Moreover, the backup says
669 # it is done but we can't find the results it
670 # should have copied over. Reschedule the whole
672 return self.emergency_retry_nasty_bundle(bundle)
674 # Expected(?). We're a backup and our bundle is
675 # cancelled before we even got started. Something
676 # went bad in post_launch_work (I acutually don't
677 # see what?) but probably not worth worrying
681 # Send input code / data to worker machine if it's not local.
682 if hostname not in machine:
684 cmd = f'{RSYNC} {bundle.code_file} {username}@{machine}:{bundle.code_file}'
685 start_ts = time.time()
686 logger.info(f"{self.bundle_prefix(bundle)}: Copying work to {worker} via {cmd}.")
688 xfer_latency = time.time() - start_ts
689 logger.info(f"{self.bundle_prefix(bundle)}: Copying done to {worker} in {xfer_latency:.1f}s.")
690 except Exception as e:
693 f'{self.bundle_prefix(bundle)}: failed to send instructions to worker machine?!?'
695 assert bundle.worker is not None
696 self.status.record_release_worker(
701 self.release_worker(bundle.worker)
702 self.adjust_task_count(-1)
704 # Weird. We tried to copy the code to the worker and it failed...
705 # And we're the original bundle. We have to retry.
706 return self.emergency_retry_nasty_bundle(bundle)
708 # This is actually expected; we're a backup.
709 # There's a race condition where someone else
710 # already finished the work and removed the source
711 # code file before we could copy it. No biggie.
714 # Kick off the work. Note that if this fails we let
715 # wait_for_process deal with it.
716 self.status.record_processing_began(uuid)
717 cmd = (f'{SSH} {bundle.username}@{bundle.machine} '
718 f'"source py39-venv/bin/activate &&'
719 f' /home/scott/lib/python_modules/remote_worker.py'
720 f' --code_file {bundle.code_file} --result_file {bundle.result_file}"')
721 logger.debug(f'{self.bundle_prefix(bundle)}: Executing {cmd} in the background to kick off work...')
722 p = cmd_in_background(cmd, silent=True)
723 bundle.pid = pid = p.pid
724 logger.debug(f'{self.bundle_prefix(bundle)}: Local ssh process pid={pid}; remote worker is {machine}.')
725 return self.wait_for_process(p, bundle, 0)
727 def wait_for_process(self, p: subprocess.Popen, bundle: BundleDetails, depth: int) -> Any:
729 machine = bundle.machine
734 f"I've gotten repeated errors waiting on this bundle; giving up on pid={pid}."
737 self.status.record_release_worker(
742 self.release_worker(bundle.worker)
743 self.adjust_task_count(-1)
744 return self.emergency_retry_nasty_bundle(bundle)
746 # Spin until either the ssh job we scheduled finishes the
747 # bundle or some backup worker signals that they finished it
752 except subprocess.TimeoutExpired:
754 if self.check_if_cancelled(bundle):
756 f'{self.bundle_prefix(bundle)}: another worker finished bundle, checking it out...'
761 f"{self.bundle_prefix(bundle)}: pid {pid} ({machine}) our ssh finished, checking it out..."
766 # If we get here we believe the bundle is done; either the ssh
767 # subprocess finished (hopefully successfully) or we noticed
768 # that some other worker seems to have completed the bundle
769 # and we're bailing out.
771 ret = self.post_launch_work(bundle)
772 if ret is not None and p is not None:
776 # Something went wrong; e.g. we could not copy the results
777 # back, cleanup after ourselves on the remote machine, or
778 # unpickle the results we got from the remove machine. If we
779 # still have an active ssh subprocess, keep waiting on it.
780 # Otherwise, time for an emergency reschedule.
781 except Exception as e:
783 logger.error(f'{self.bundle_prefix(bundle)}: Something unexpected just happened...')
786 f"{self.bundle_prefix(bundle)}: Failed to wrap up \"done\" bundle, re-waiting on active ssh."
788 return self.wait_for_process(p, bundle, depth + 1)
790 self.status.record_release_worker(
795 self.release_worker(bundle.worker)
796 self.adjust_task_count(-1)
797 return self.emergency_retry_nasty_bundle(bundle)
799 def post_launch_work(self, bundle: BundleDetails) -> Any:
800 with self.status.lock:
801 is_original = bundle.src_bundle is None
802 was_cancelled = bundle.was_cancelled
803 username = bundle.username
804 machine = bundle.machine
805 result_file = bundle.result_file
806 code_file = bundle.code_file
810 # Whether original or backup, if we finished first we must
811 # fetch the results if the computation happened on a
813 bundle.end_ts = time.time()
814 if not was_cancelled:
815 assert bundle.machine is not None
816 if bundle.hostname not in bundle.machine:
817 cmd = f'{RSYNC} {username}@{machine}:{result_file} {result_file} 2>/dev/null'
819 f"{self.bundle_prefix(bundle)}: Fetching results from {username}@{machine} via {cmd}"
822 # If either of these throw they are handled in
825 run_silently(f'{SSH} {username}@{machine}'
826 f' "/bin/rm -f {code_file} {result_file}"')
827 dur = bundle.end_ts - bundle.start_ts
828 self.histogram.add_item(dur)
830 # Only the original worker should unpickle the file contents
831 # though since it's the only one whose result matters. The
832 # original is also the only job that may delete result_file
833 # from disk. Note that the original may have been cancelled
834 # if one of the backups finished first; it still must read the
837 logger.debug(f"{self.bundle_prefix(bundle)}: Unpickling {result_file}.")
839 with open(f'{result_file}', 'rb') as rb:
840 serialized = rb.read()
841 result = cloudpickle.loads(serialized)
842 except Exception as e:
843 msg = f'Failed to load {result_file}, this is bad news.'
845 self.status.record_release_worker(
850 self.release_worker(bundle.worker)
852 # Re-raise the exception; the code in wait_for_process may
853 # decide to emergency_retry_nasty_bundle here.
857 f'Removing local (master) {code_file} and {result_file}.'
859 os.remove(f'{result_file}')
860 os.remove(f'{code_file}')
862 # Notify any backups that the original is done so they
863 # should stop ASAP. Do this whether or not we
864 # finished first since there could be more than one
866 if bundle.backup_bundles is not None:
867 for backup in bundle.backup_bundles:
869 f'{self.bundle_prefix(bundle)}: Notifying backup {backup.uuid} that it\'s cancelled'
871 backup.is_cancelled.set()
873 # This is a backup job and, by now, we have already fetched
874 # the bundle results.
876 # Backup results don't matter, they just need to leave the
877 # result file in the right place for their originals to
878 # read/unpickle later.
881 # Tell the original to stop if we finished first.
882 if not was_cancelled:
884 f'{self.bundle_prefix(bundle)}: Notifying original {bundle.src_bundle.uuid} we beat them to it.'
886 bundle.src_bundle.is_cancelled.set()
888 assert bundle.worker is not None
889 self.status.record_release_worker(
894 self.release_worker(bundle.worker)
895 self.adjust_task_count(-1)
898 def create_original_bundle(self, pickle, fname: str):
899 from string_utils import generate_uuid
900 uuid = generate_uuid(as_hex=True)
901 code_file = f'/tmp/{uuid}.code.bin'
902 result_file = f'/tmp/{uuid}.result.bin'
904 logger.debug(f'Writing pickled code to {code_file}')
905 with open(f'{code_file}', 'wb') as wb:
908 bundle = BundleDetails(
909 pickled_code = pickle,
915 hostname = platform.node(),
916 code_file = code_file,
917 result_file = result_file,
919 start_ts = time.time(),
924 is_cancelled = threading.Event(),
925 was_cancelled = False,
929 self.status.record_bundle_details(bundle)
930 logger.debug(f'{self.bundle_prefix(bundle)}: Created an original bundle')
933 def create_backup_bundle(self, src_bundle: BundleDetails):
934 assert src_bundle.backup_bundles is not None
935 n = len(src_bundle.backup_bundles)
936 uuid = src_bundle.uuid + f'_backup#{n}'
938 backup_bundle = BundleDetails(
939 pickled_code = src_bundle.pickled_code,
941 fname = src_bundle.fname,
945 hostname = src_bundle.hostname,
946 code_file = src_bundle.code_file,
947 result_file = src_bundle.result_file,
949 start_ts = time.time(),
953 src_bundle = src_bundle,
954 is_cancelled = threading.Event(),
955 was_cancelled = False,
956 backup_bundles = None, # backup backups not allowed
959 src_bundle.backup_bundles.append(backup_bundle)
960 self.status.record_bundle_details_already_locked(backup_bundle)
961 logger.debug(f'{self.bundle_prefix(bundle)}: Created a backup bundle')
964 def schedule_backup_for_bundle(self,
965 src_bundle: BundleDetails):
966 assert self.status.lock.locked()
967 backup_bundle = self.create_backup_bundle(src_bundle)
969 f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...'
971 self._helper_executor.submit(self.launch, backup_bundle)
973 # Results from backups don't matter; if they finish first
974 # they will move the result_file to this machine and let
975 # the original pick them up and unpickle them.
977 def emergency_retry_nasty_bundle(self, bundle: BundleDetails) -> fut.Future:
979 is_original = bundle.src_bundle is None
981 avoid_last_machine = bundle.machine
982 bundle.machine = None
983 bundle.username = None
984 bundle.failure_count += 1
990 if bundle.failure_count > retry_limit:
992 f'{self.bundle_prefix(bundle)}: Tried this bundle too many times already ({retry_limit}x); giving up.'
995 raise RemoteExecutorException(
996 f'{self.bundle_prefix(bundle)}: This bundle can\'t be completed despite several backups and retries'
999 logger.error(f'{self.bundle_prefix(bundle)}: At least it\'s only a backup; better luck with the others.')
1003 f'>>> Emergency rescheduling {self.bundle_prefix(bundle)} because of unexected errors (wtf?!) <<<'
1005 return self.launch(bundle, avoid_last_machine)
1011 **kwargs) -> fut.Future:
1012 pickle = make_cloud_pickle(function, *args, **kwargs)
1013 bundle = self.create_original_bundle(pickle, function.__name__)
1014 self.total_bundles_submitted += 1
1015 return self._helper_executor.submit(self.launch, bundle)
1018 def shutdown(self, wait=True) -> None:
1019 self._helper_executor.shutdown(wait)
1020 logging.debug(f'Shutting down RemoteExecutor {self.title}')
1021 print(self.histogram)
1025 class DefaultExecutors(object):
1027 self.thread_executor: Optional[ThreadExecutor] = None
1028 self.process_executor: Optional[ProcessExecutor] = None
1029 self.remote_executor: Optional[RemoteExecutor] = None
1031 def ping(self, host) -> bool:
1032 logger.debug(f'RUN> ping -c 1 {host}')
1034 x = cmd_with_timeout(
1035 f'ping -c 1 {host} >/dev/null 2>/dev/null',
1042 def thread_pool(self) -> ThreadExecutor:
1043 if self.thread_executor is None:
1044 self.thread_executor = ThreadExecutor()
1045 return self.thread_executor
1047 def process_pool(self) -> ProcessExecutor:
1048 if self.process_executor is None:
1049 self.process_executor = ProcessExecutor()
1050 return self.process_executor
1052 def remote_pool(self) -> RemoteExecutor:
1053 if self.remote_executor is None:
1054 logger.info('Looking for some helper machines...')
1055 pool: List[RemoteWorkerRecord] = []
1056 if self.ping('cheetah.house'):
1057 logger.info('Found cheetah.house')
1061 machine = 'cheetah.house',
1066 if self.ping('video.house'):
1067 logger.info('Found video.house')
1071 machine = 'video.house',
1076 if self.ping('wannabe.house'):
1077 logger.info('Found wannabe.house')
1081 machine = 'wannabe.house',
1086 if self.ping('meerkat.cabin'):
1087 logger.info('Found meerkat.cabin')
1091 machine = 'meerkat.cabin',
1096 if self.ping('backup.house'):
1097 logger.info('Found backup.house')
1101 machine = 'backup.house',
1106 if self.ping('kiosk.house'):
1107 logger.info('Found kiosk.house')
1111 machine = 'kiosk.house',
1116 if self.ping('puma.cabin'):
1117 logger.info('Found puma.cabin')
1121 machine = 'puma.cabin',
1127 # The controller machine has a lot to do; go easy on it.
1129 if record.machine == platform.node() and record.count > 1:
1130 logger.info(f'Reducing workload for {record.machine}.')
1133 policy = WeightedRandomRemoteWorkerSelectionPolicy()
1134 policy.register_worker_pool(pool)
1135 self.remote_executor = RemoteExecutor(pool, policy)
1136 return self.remote_executor