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=''):
78 self.histogram = hist.SimpleHistogram(
79 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 self.task_count += delta
92 logger.debug(f'Executor current task count is {self.task_count}')
95 class ThreadExecutor(BaseExecutor):
96 def __init__(self, max_workers: Optional[int] = None):
99 if max_workers is not None:
100 workers = max_workers
101 elif 'executors_threadpool_size' in config.config:
102 workers = config.config['executors_threadpool_size']
103 logger.debug(f'Creating threadpool executor with {workers} workers')
104 self._thread_pool_executor = fut.ThreadPoolExecutor(
105 max_workers=workers, thread_name_prefix="thread_executor_helper"
108 def run_local_bundle(self, fun, *args, **kwargs):
109 logger.debug(f"Running local bundle at {fun.__name__}")
111 result = fun(*args, **kwargs)
113 self.adjust_task_count(-1)
114 duration = end - start
115 logger.debug(f"{fun.__name__} finished; used {duration:.1f}s")
116 self.histogram.add_item(duration)
120 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
121 self.adjust_task_count(+1)
123 newargs.append(function)
126 return self._thread_pool_executor.submit(
127 self.run_local_bundle, *newargs, **kwargs
131 def shutdown(self, wait=True) -> None:
132 logger.debug(f'Shutting down threadpool executor {self.title}')
133 print(self.histogram)
134 self._thread_pool_executor.shutdown(wait)
137 class ProcessExecutor(BaseExecutor):
138 def __init__(self, max_workers=None):
141 if max_workers is not None:
142 workers = max_workers
143 elif 'executors_processpool_size' in config.config:
144 workers = config.config['executors_processpool_size']
145 logger.debug(f'Creating processpool executor with {workers} workers.')
146 self._process_executor = fut.ProcessPoolExecutor(
150 def run_cloud_pickle(self, pickle):
151 fun, args, kwargs = cloudpickle.loads(pickle)
152 logger.debug(f"Running pickled bundle at {fun.__name__}")
153 result = fun(*args, **kwargs)
154 self.adjust_task_count(-1)
158 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
160 self.adjust_task_count(+1)
161 pickle = make_cloud_pickle(function, *args, **kwargs)
162 result = self._process_executor.submit(self.run_cloud_pickle, pickle)
163 result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start))
167 def shutdown(self, wait=True) -> None:
168 logger.debug(f'Shutting down processpool executor {self.title}')
169 self._process_executor.shutdown(wait)
170 print(self.histogram)
172 def __getstate__(self):
173 state = self.__dict__.copy()
174 state['_process_executor'] = None
178 class RemoteExecutorException(Exception):
179 """Thrown when a bundle cannot be executed despite several retries."""
185 class RemoteWorkerRecord:
192 return hash((self.username, self.machine))
195 return f'{self.username}@{self.machine}'
203 worker: Optional[RemoteWorkerRecord]
204 username: Optional[str]
205 machine: Optional[str]
212 slower_than_local_p95: bool
213 slower_than_global_p95: bool
214 src_bundle: BundleDetails
215 is_cancelled: threading.Event
217 backup_bundles: Optional[List[BundleDetails]]
222 if uuid[-9:-2] == '_backup':
224 suffix = f'{uuid[-6:]}_b{self.uuid[-1:]}'
234 fg('marigold yellow'),
237 fg('cornflower blue'),
238 fg('turquoise blue'),
240 fg('lavender purple'),
243 c = colorz[int(uuid[-2:], 16) % len(colorz)]
244 fname = self.fname if self.fname is not None else 'nofname'
245 machine = self.machine if self.machine is not None else 'nomachine'
246 return f'{c}{suffix}/{fname}/{machine}{reset()}'
249 class RemoteExecutorStatus:
250 def __init__(self, total_worker_count: int) -> None:
251 self.worker_count: int = total_worker_count
252 self.known_workers: Set[RemoteWorkerRecord] = set()
253 self.start_time: float = time.time()
254 self.start_per_bundle: Dict[str, float] = defaultdict(float)
255 self.end_per_bundle: Dict[str, float] = defaultdict(float)
256 self.finished_bundle_timings_per_worker: Dict[
257 RemoteWorkerRecord, List[float]
259 self.in_flight_bundles_by_worker: Dict[RemoteWorkerRecord, Set[str]] = {}
260 self.bundle_details_by_uuid: Dict[str, BundleDetails] = {}
261 self.finished_bundle_timings: List[float] = []
262 self.last_periodic_dump: Optional[float] = None
263 self.total_bundles_submitted: int = 0
265 # Protects reads and modification using self. Also used
266 # as a memory fence for modifications to bundle.
267 self.lock: threading.Lock = threading.Lock()
269 def record_acquire_worker(self, worker: RemoteWorkerRecord, uuid: str) -> None:
271 self.record_acquire_worker_already_locked(worker, uuid)
273 def record_acquire_worker_already_locked(
274 self, worker: RemoteWorkerRecord, uuid: str
276 assert self.lock.locked()
277 self.known_workers.add(worker)
278 self.start_per_bundle[uuid] = None
279 x = self.in_flight_bundles_by_worker.get(worker, set())
281 self.in_flight_bundles_by_worker[worker] = x
283 def record_bundle_details(self, details: BundleDetails) -> None:
285 self.record_bundle_details_already_locked(details)
287 def record_bundle_details_already_locked(self, details: BundleDetails) -> None:
288 assert self.lock.locked()
289 self.bundle_details_by_uuid[details.uuid] = details
291 def record_release_worker(
293 worker: RemoteWorkerRecord,
298 self.record_release_worker_already_locked(worker, uuid, was_cancelled)
300 def record_release_worker_already_locked(
302 worker: RemoteWorkerRecord,
306 assert self.lock.locked()
308 self.end_per_bundle[uuid] = ts
309 self.in_flight_bundles_by_worker[worker].remove(uuid)
310 if not was_cancelled:
311 bundle_latency = ts - self.start_per_bundle[uuid]
312 x = self.finished_bundle_timings_per_worker.get(worker, list())
313 x.append(bundle_latency)
314 self.finished_bundle_timings_per_worker[worker] = x
315 self.finished_bundle_timings.append(bundle_latency)
317 def record_processing_began(self, uuid: str):
319 self.start_per_bundle[uuid] = time.time()
321 def total_in_flight(self) -> int:
322 assert self.lock.locked()
324 for worker in self.known_workers:
325 total_in_flight += len(self.in_flight_bundles_by_worker[worker])
326 return total_in_flight
328 def total_idle(self) -> int:
329 assert self.lock.locked()
330 return self.worker_count - self.total_in_flight()
333 assert self.lock.locked()
335 total_finished = len(self.finished_bundle_timings)
336 total_in_flight = self.total_in_flight()
337 ret = f'\n\n{underline()}Remote Executor Pool Status{reset()}: '
339 if len(self.finished_bundle_timings) > 1:
340 qall = numpy.quantile(self.finished_bundle_timings, [0.5, 0.95])
342 f'⏱=∀p50:{qall[0]:.1f}s, ∀p95:{qall[1]:.1f}s, total={ts-self.start_time:.1f}s, '
343 f'✅={total_finished}/{self.total_bundles_submitted}, '
344 f'💻n={total_in_flight}/{self.worker_count}\n'
348 f'⏱={ts-self.start_time:.1f}s, '
349 f'✅={total_finished}/{self.total_bundles_submitted}, '
350 f'💻n={total_in_flight}/{self.worker_count}\n'
353 for worker in self.known_workers:
354 ret += f' {fg("lightning yellow")}{worker.machine}{reset()}: '
355 timings = self.finished_bundle_timings_per_worker.get(worker, [])
359 qworker = numpy.quantile(timings, [0.5, 0.95])
360 ret += f' 💻p50: {qworker[0]:.1f}s, 💻p95: {qworker[1]:.1f}s\n'
364 ret += f' ...finished {count} total bundle(s) so far\n'
365 in_flight = len(self.in_flight_bundles_by_worker[worker])
367 ret += f' ...{in_flight} bundles currently in flight:\n'
368 for bundle_uuid in self.in_flight_bundles_by_worker[worker]:
369 details = self.bundle_details_by_uuid.get(bundle_uuid, None)
370 pid = str(details.pid) if (details and details.pid != 0) else "TBD"
371 if self.start_per_bundle[bundle_uuid] is not None:
372 sec = ts - self.start_per_bundle[bundle_uuid]
373 ret += f' (pid={pid}): {details} for {sec:.1f}s so far '
375 ret += f' {details} setting up / copying data...'
378 if qworker is not None:
380 ret += f'{bg("red")}>💻p95{reset()} '
381 if details is not None:
382 details.slower_than_local_p95 = True
384 if details is not None:
385 details.slower_than_local_p95 = False
389 ret += f'{bg("red")}>∀p95{reset()} '
390 if details is not None:
391 details.slower_than_global_p95 = True
393 details.slower_than_global_p95 = False
397 def periodic_dump(self, total_bundles_submitted: int) -> None:
398 assert self.lock.locked()
399 self.total_bundles_submitted = total_bundles_submitted
401 if self.last_periodic_dump is None or ts - self.last_periodic_dump > 5.0:
403 self.last_periodic_dump = ts
406 class RemoteWorkerSelectionPolicy(ABC):
407 def register_worker_pool(self, workers):
408 self.workers = workers
411 def is_worker_available(self) -> bool:
415 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
419 class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
421 def is_worker_available(self) -> bool:
422 for worker in self.workers:
428 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
430 for worker in self.workers:
431 if worker.machine != machine_to_avoid:
433 for _ in range(worker.count * worker.weight):
434 grabbag.append(worker)
436 if len(grabbag) == 0:
438 f'There are no available workers that avoid {machine_to_avoid}...'
440 for worker in self.workers:
442 for _ in range(worker.count * worker.weight):
443 grabbag.append(worker)
445 if len(grabbag) == 0:
446 logger.warning('There are no available workers?!')
449 worker = random.sample(grabbag, 1)[0]
450 assert worker.count > 0
452 logger.debug(f'Chose worker {worker}')
456 class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
457 def __init__(self) -> None:
461 def is_worker_available(self) -> bool:
462 for worker in self.workers:
469 self, machine_to_avoid: str = None
470 ) -> Optional[RemoteWorkerRecord]:
473 worker = self.workers[x]
477 if x >= len(self.workers):
480 logger.debug(f'Selected worker {worker}')
483 if x >= len(self.workers):
486 msg = 'Unexpectedly could not find a worker, retrying...'
491 class RemoteExecutor(BaseExecutor):
494 workers: List[RemoteWorkerRecord],
495 policy: RemoteWorkerSelectionPolicy,
498 self.workers = workers
500 self.worker_count = 0
501 for worker in self.workers:
502 self.worker_count += worker.count
503 if self.worker_count <= 0:
504 msg = f"We need somewhere to schedule work; count was {self.worker_count}"
506 raise RemoteExecutorException(msg)
507 self.policy.register_worker_pool(self.workers)
508 self.cv = threading.Condition()
510 f'Creating {self.worker_count} local threads, one per remote worker.'
512 self._helper_executor = fut.ThreadPoolExecutor(
513 thread_name_prefix="remote_executor_helper",
514 max_workers=self.worker_count,
516 self.status = RemoteExecutorStatus(self.worker_count)
517 self.total_bundles_submitted = 0
518 self.backup_lock = threading.Lock()
519 self.last_backup = None
521 self.heartbeat_thread,
522 self.heartbeat_stop_event,
523 ) = self.run_periodic_heartbeat()
526 def run_periodic_heartbeat(self, stop_event: threading.Event) -> None:
527 while not stop_event.is_set():
529 logger.debug('Running periodic heartbeat code...')
531 logger.debug('Periodic heartbeat thread shutting down.')
533 def heartbeat(self) -> None:
534 # Note: this is invoked on a background thread, not an
535 # executor thread. Be careful what you do with it b/c it
536 # needs to get back and dump status again periodically.
537 with self.status.lock:
538 self.status.periodic_dump(self.total_bundles_submitted)
540 # Look for bundles to reschedule via executor.submit
541 if config.config['executors_schedule_remote_backups']:
542 self.maybe_schedule_backup_bundles()
544 def maybe_schedule_backup_bundles(self):
545 assert self.status.lock.locked()
546 num_done = len(self.status.finished_bundle_timings)
547 num_idle_workers = self.worker_count - self.task_count
551 and num_idle_workers > 1
552 and (self.last_backup is None or (now - self.last_backup > 9.0))
553 and self.backup_lock.acquire(blocking=False)
556 assert self.backup_lock.locked()
558 bundle_to_backup = None
563 ) in self.status.in_flight_bundles_by_worker.items():
565 # Prefer to schedule backups of bundles running on
568 for record in self.workers:
569 if worker.machine == record.machine:
570 base_score = float(record.weight)
571 base_score = 1.0 / base_score
573 base_score = int(base_score)
576 for uuid in bundle_uuids:
577 bundle = self.status.bundle_details_by_uuid.get(uuid, None)
580 and bundle.src_bundle is None
581 and bundle.backup_bundles is not None
585 # Schedule backups of bundles running
586 # longer; especially those that are
588 start_ts = self.status.start_per_bundle[uuid]
589 if start_ts is not None:
590 runtime = now - start_ts
593 f'score[{bundle}] => {score} # latency boost'
596 if bundle.slower_than_local_p95:
599 f'score[{bundle}] => {score} # >worker p95'
602 if bundle.slower_than_global_p95:
605 f'score[{bundle}] => {score} # >global p95'
608 # Prefer backups of bundles that don't
609 # have backups already.
610 backup_count = len(bundle.backup_bundles)
611 if backup_count == 0:
613 elif backup_count == 1:
615 elif backup_count == 2:
620 f'score[{bundle}] => {score} # {backup_count} dup backup factor'
624 best_score is None or score > best_score
626 bundle_to_backup = bundle
627 assert bundle is not None
628 assert bundle.backup_bundles is not None
629 assert bundle.src_bundle is None
632 # Note: this is all still happening on the heartbeat
633 # runner thread. That's ok because
634 # schedule_backup_for_bundle uses the executor to
635 # submit the bundle again which will cause it to be
636 # picked up by a worker thread and allow this thread
637 # to return to run future heartbeats.
638 if bundle_to_backup is not None:
639 self.last_backup = now
641 f'=====> SCHEDULING BACKUP {bundle_to_backup} (score={best_score:.1f}) <====='
643 self.schedule_backup_for_bundle(bundle_to_backup)
645 self.backup_lock.release()
647 def is_worker_available(self) -> bool:
648 return self.policy.is_worker_available()
651 self, machine_to_avoid: str = None
652 ) -> Optional[RemoteWorkerRecord]:
653 return self.policy.acquire_worker(machine_to_avoid)
655 def find_available_worker_or_block(
656 self, machine_to_avoid: str = None
657 ) -> RemoteWorkerRecord:
659 while not self.is_worker_available():
661 worker = self.acquire_worker(machine_to_avoid)
662 if worker is not None:
664 msg = "We should never reach this point in the code"
668 def release_worker(self, bundle: BundleDetails, *, was_cancelled=True) -> None:
669 worker = bundle.worker
670 assert worker is not None
671 logger.debug(f'Released worker {worker}')
672 self.status.record_release_worker(
680 self.adjust_task_count(-1)
682 def check_if_cancelled(self, bundle: BundleDetails) -> bool:
683 with self.status.lock:
684 if bundle.is_cancelled.wait(timeout=0.0):
685 logger.debug(f'Bundle {bundle.uuid} is cancelled, bail out.')
686 bundle.was_cancelled = True
690 def launch(self, bundle: BundleDetails, override_avoid_machine=None) -> Any:
691 """Find a worker for bundle or block until one is available."""
692 self.adjust_task_count(+1)
694 hostname = bundle.hostname
695 avoid_machine = override_avoid_machine
696 is_original = bundle.src_bundle is None
698 # Try not to schedule a backup on the same host as the original.
699 if avoid_machine is None and bundle.src_bundle is not None:
700 avoid_machine = bundle.src_bundle.machine
702 while worker is None:
703 worker = self.find_available_worker_or_block(avoid_machine)
706 # Ok, found a worker.
707 bundle.worker = worker
708 machine = bundle.machine = worker.machine
709 username = bundle.username = worker.username
710 self.status.record_acquire_worker(worker, uuid)
711 logger.debug(f'{bundle}: Running bundle on {worker}...')
713 # Before we do any work, make sure the bundle is still viable.
714 # It may have been some time between when it was submitted and
715 # now due to lack of worker availability and someone else may
716 # have already finished it.
717 if self.check_if_cancelled(bundle):
719 return self.process_work_result(bundle)
720 except Exception as e:
722 f'{bundle}: bundle says it\'s cancelled upfront but no results?!'
724 self.release_worker(bundle)
726 # Weird. We are the original owner of this
727 # bundle. For it to have been cancelled, a backup
728 # must have already started and completed before
729 # we even for started. Moreover, the backup says
730 # it is done but we can't find the results it
731 # should have copied over. Reschedule the whole
735 f'{bundle}: We are the original owner thread and yet there are '
736 + 'no results for this bundle. This is unexpected and bad.'
738 return self.emergency_retry_nasty_bundle(bundle)
740 # Expected(?). We're a backup and our bundle is
741 # cancelled before we even got started. Something
742 # went bad in process_work_result (I acutually don't
743 # see what?) but probably not worth worrying
744 # about. Let the original thread worry about
745 # either finding the results or complaining about
749 # Send input code / data to worker machine if it's not local.
750 if hostname not in machine:
753 f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}'
755 start_ts = time.time()
756 logger.info(f"{bundle}: Copying work to {worker} via {cmd}.")
758 xfer_latency = time.time() - start_ts
759 logger.debug(f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s.")
760 except Exception as e:
761 self.release_worker(bundle)
763 # Weird. We tried to copy the code to the worker and it failed...
764 # And we're the original bundle. We have to retry.
767 f"{bundle}: Failed to send instructions to the worker machine?! "
768 + "This is not expected; we\'re the original bundle so this shouldn\'t "
769 + "be a race condition. Attempting an emergency retry..."
771 return self.emergency_retry_nasty_bundle(bundle)
773 # This is actually expected; we're a backup.
774 # There's a race condition where someone else
775 # already finished the work and removed the source
776 # code file before we could copy it. No biggie.
777 msg = f'{bundle}: Failed to send instructions to the worker machine... '
778 msg += 'We\'re a backup and this may be caused by the original (or some '
779 msg += 'other backup) already finishing this work. Ignoring this.'
783 # Kick off the work. Note that if this fails we let
784 # wait_for_process deal with it.
785 self.status.record_processing_began(uuid)
787 f'{SSH} {bundle.username}@{bundle.machine} '
788 f'"source py38-venv/bin/activate &&'
789 f' /home/scott/lib/python_modules/remote_worker.py'
790 f' --code_file {bundle.code_file} --result_file {bundle.result_file}"'
792 logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...')
793 p = cmd_in_background(cmd, silent=True)
796 f'{bundle}: Local ssh process pid={p.pid}; remote worker is {machine}.'
798 return self.wait_for_process(p, bundle, 0)
800 def wait_for_process(
801 self, p: subprocess.Popen, bundle: BundleDetails, depth: int
803 machine = bundle.machine
807 f"I've gotten repeated errors waiting on this bundle; giving up on pid={pid}."
810 self.release_worker(bundle)
811 return self.emergency_retry_nasty_bundle(bundle)
813 # Spin until either the ssh job we scheduled finishes the
814 # bundle or some backup worker signals that they finished it
819 except subprocess.TimeoutExpired:
820 if self.check_if_cancelled(bundle):
822 f'{bundle}: looks like another worker finished bundle...'
826 logger.info(f"{bundle}: pid {pid} ({machine}) is finished!")
830 # If we get here we believe the bundle is done; either the ssh
831 # subprocess finished (hopefully successfully) or we noticed
832 # that some other worker seems to have completed the bundle
833 # and we're bailing out.
835 ret = self.process_work_result(bundle)
836 if ret is not None and p is not None:
840 # Something went wrong; e.g. we could not copy the results
841 # back, cleanup after ourselves on the remote machine, or
842 # unpickle the results we got from the remove machine. If we
843 # still have an active ssh subprocess, keep waiting on it.
844 # Otherwise, time for an emergency reschedule.
845 except Exception as e:
847 logger.error(f'{bundle}: Something unexpected just happened...')
849 msg = f"{bundle}: Failed to wrap up \"done\" bundle, re-waiting on active ssh."
851 return self.wait_for_process(p, bundle, depth + 1)
853 self.release_worker(bundle)
854 return self.emergency_retry_nasty_bundle(bundle)
856 def process_work_result(self, bundle: BundleDetails) -> Any:
857 with self.status.lock:
858 is_original = bundle.src_bundle is None
859 was_cancelled = bundle.was_cancelled
860 username = bundle.username
861 machine = bundle.machine
862 result_file = bundle.result_file
863 code_file = bundle.code_file
865 # Whether original or backup, if we finished first we must
866 # fetch the results if the computation happened on a
868 bundle.end_ts = time.time()
869 if not was_cancelled:
870 assert bundle.machine is not None
871 if bundle.hostname not in bundle.machine:
872 cmd = f'{SCP} {username}@{machine}:{result_file} {result_file} 2>/dev/null'
874 f"{bundle}: Fetching results back from {username}@{machine} via {cmd}"
877 # If either of these throw they are handled in
883 except Exception as e:
891 f'{SSH} {username}@{machine}'
892 f' "/bin/rm -f {code_file} {result_file}"'
895 f'Fetching results back took {time.time() - bundle.end_ts:.1f}s.'
897 dur = bundle.end_ts - bundle.start_ts
898 self.histogram.add_item(dur)
900 # Only the original worker should unpickle the file contents
901 # though since it's the only one whose result matters. The
902 # original is also the only job that may delete result_file
903 # from disk. Note that the original may have been cancelled
904 # if one of the backups finished first; it still must read the
907 logger.debug(f"{bundle}: Unpickling {result_file}.")
909 with open(result_file, 'rb') as rb:
910 serialized = rb.read()
911 result = cloudpickle.loads(serialized)
912 except Exception as e:
914 msg = f'Failed to load {result_file}... this is bad news.'
916 self.release_worker(bundle)
918 # Re-raise the exception; the code in wait_for_process may
919 # decide to emergency_retry_nasty_bundle here.
921 logger.debug(f'Removing local (master) {code_file} and {result_file}.')
922 os.remove(f'{result_file}')
923 os.remove(f'{code_file}')
925 # Notify any backups that the original is done so they
926 # should stop ASAP. Do this whether or not we
927 # finished first since there could be more than one
929 if bundle.backup_bundles is not None:
930 for backup in bundle.backup_bundles:
932 f'{bundle}: Notifying backup {backup.uuid} that it\'s cancelled'
934 backup.is_cancelled.set()
936 # This is a backup job and, by now, we have already fetched
937 # the bundle results.
939 # Backup results don't matter, they just need to leave the
940 # result file in the right place for their originals to
941 # read/unpickle later.
944 # Tell the original to stop if we finished first.
945 if not was_cancelled:
947 f'{bundle}: Notifying original {bundle.src_bundle.uuid} we beat them to it.'
949 bundle.src_bundle.is_cancelled.set()
950 self.release_worker(bundle, was_cancelled=was_cancelled)
953 def create_original_bundle(self, pickle, fname: str):
954 from string_utils import generate_uuid
956 uuid = generate_uuid(omit_dashes=True)
957 code_file = f'/tmp/{uuid}.code.bin'
958 result_file = f'/tmp/{uuid}.result.bin'
960 logger.debug(f'Writing pickled code to {code_file}')
961 with open(f'{code_file}', 'wb') as wb:
964 bundle = BundleDetails(
971 hostname=platform.node(),
973 result_file=result_file,
975 start_ts=time.time(),
977 slower_than_local_p95=False,
978 slower_than_global_p95=False,
980 is_cancelled=threading.Event(),
985 self.status.record_bundle_details(bundle)
986 logger.debug(f'{bundle}: Created an original bundle')
989 def create_backup_bundle(self, src_bundle: BundleDetails):
990 assert src_bundle.backup_bundles is not None
991 n = len(src_bundle.backup_bundles)
992 uuid = src_bundle.uuid + f'_backup#{n}'
994 backup_bundle = BundleDetails(
995 pickled_code=src_bundle.pickled_code,
997 fname=src_bundle.fname,
1001 hostname=src_bundle.hostname,
1002 code_file=src_bundle.code_file,
1003 result_file=src_bundle.result_file,
1005 start_ts=time.time(),
1007 slower_than_local_p95=False,
1008 slower_than_global_p95=False,
1009 src_bundle=src_bundle,
1010 is_cancelled=threading.Event(),
1011 was_cancelled=False,
1012 backup_bundles=None, # backup backups not allowed
1015 src_bundle.backup_bundles.append(backup_bundle)
1016 self.status.record_bundle_details_already_locked(backup_bundle)
1017 logger.debug(f'{backup_bundle}: Created a backup bundle')
1018 return backup_bundle
1020 def schedule_backup_for_bundle(self, src_bundle: BundleDetails):
1021 assert self.status.lock.locked()
1022 assert src_bundle is not None
1023 backup_bundle = self.create_backup_bundle(src_bundle)
1025 f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...'
1027 self._helper_executor.submit(self.launch, backup_bundle)
1029 # Results from backups don't matter; if they finish first
1030 # they will move the result_file to this machine and let
1031 # the original pick them up and unpickle them.
1033 def emergency_retry_nasty_bundle(self, bundle: BundleDetails) -> fut.Future:
1034 is_original = bundle.src_bundle is None
1035 bundle.worker = None
1036 avoid_last_machine = bundle.machine
1037 bundle.machine = None
1038 bundle.username = None
1039 bundle.failure_count += 1
1045 if bundle.failure_count > retry_limit:
1047 f'{bundle}: Tried this bundle too many times already ({retry_limit}x); giving up.'
1050 raise RemoteExecutorException(
1051 f'{bundle}: This bundle can\'t be completed despite several backups and retries'
1055 f'{bundle}: At least it\'s only a backup; better luck with the others.'
1059 msg = f'>>> Emergency rescheduling {bundle} because of unexected errors (wtf?!) <<<'
1062 return self.launch(bundle, avoid_last_machine)
1065 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
1066 pickle = make_cloud_pickle(function, *args, **kwargs)
1067 bundle = self.create_original_bundle(pickle, function.__name__)
1068 self.total_bundles_submitted += 1
1069 return self._helper_executor.submit(self.launch, bundle)
1072 def shutdown(self, wait=True) -> None:
1073 logging.debug(f'Shutting down RemoteExecutor {self.title}')
1074 self.heartbeat_stop_event.set()
1075 self.heartbeat_thread.join()
1076 self._helper_executor.shutdown(wait)
1077 print(self.histogram)
1081 class DefaultExecutors(object):
1083 self.thread_executor: Optional[ThreadExecutor] = None
1084 self.process_executor: Optional[ProcessExecutor] = None
1085 self.remote_executor: Optional[RemoteExecutor] = None
1087 def ping(self, host) -> bool:
1088 logger.debug(f'RUN> ping -c 1 {host}')
1090 x = cmd_with_timeout(
1091 f'ping -c 1 {host} >/dev/null 2>/dev/null', timeout_seconds=1.0
1097 def thread_pool(self) -> ThreadExecutor:
1098 if self.thread_executor is None:
1099 self.thread_executor = ThreadExecutor()
1100 return self.thread_executor
1102 def process_pool(self) -> ProcessExecutor:
1103 if self.process_executor is None:
1104 self.process_executor = ProcessExecutor()
1105 return self.process_executor
1107 def remote_pool(self) -> RemoteExecutor:
1108 if self.remote_executor is None:
1109 logger.info('Looking for some helper machines...')
1110 pool: List[RemoteWorkerRecord] = []
1111 if self.ping('cheetah.house'):
1112 logger.info('Found cheetah.house')
1116 machine='cheetah.house',
1121 if self.ping('meerkat.cabin'):
1122 logger.info('Found meerkat.cabin')
1126 machine='meerkat.cabin',
1131 if self.ping('wannabe.house'):
1132 logger.info('Found wannabe.house')
1136 machine='wannabe.house',
1141 if self.ping('puma.cabin'):
1142 logger.info('Found puma.cabin')
1146 machine='puma.cabin',
1151 if self.ping('backup.house'):
1152 logger.info('Found backup.house')
1156 machine='backup.house',
1162 # The controller machine has a lot to do; go easy on it.
1164 if record.machine == platform.node() and record.count > 1:
1165 logger.info(f'Reducing workload for {record.machine}.')
1168 policy = WeightedRandomRemoteWorkerSelectionPolicy()
1169 policy.register_worker_pool(pool)
1170 self.remote_executor = RemoteExecutor(pool, policy)
1171 return self.remote_executor
1173 def shutdown(self) -> None:
1174 if self.thread_executor is not None:
1175 self.thread_executor.shutdown()
1176 self.thread_executor = None
1177 if self.process_executor is not None:
1178 self.process_executor.shutdown()
1179 self.process_executor = None
1180 if self.remote_executor is not None:
1181 self.remote_executor.shutdown()
1182 self.remote_executor = None