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 shutdown_if_idle(self) -> bool:
91 """Shutdown the executor and return True if the executor is idle
92 (i.e. there are no pending or active tasks). Return False
93 otherwise. Note: this should only be called by the launcher
97 if self.task_count == 0:
102 def adjust_task_count(self, delta: int) -> None:
103 """Change the task count. Note: do not call this method from a
104 worker, it should only be called by the launcher process /
108 self.task_count += delta
109 logger.debug(f'Adjusted task count by {delta} to {self.task_count}')
111 def get_task_count(self) -> int:
112 """Change the task count. Note: do not call this method from a
113 worker, it should only be called by the launcher process /
117 return self.task_count
120 class ThreadExecutor(BaseExecutor):
121 def __init__(self, max_workers: Optional[int] = None):
124 if max_workers is not None:
125 workers = max_workers
126 elif 'executors_threadpool_size' in config.config:
127 workers = config.config['executors_threadpool_size']
128 logger.debug(f'Creating threadpool executor with {workers} workers')
129 self._thread_pool_executor = fut.ThreadPoolExecutor(
130 max_workers=workers, thread_name_prefix="thread_executor_helper"
132 self.already_shutdown = False
134 # This is run on a different thread; do not adjust task count here.
135 def run_local_bundle(self, fun, *args, **kwargs):
136 logger.debug(f"Running local bundle at {fun.__name__}")
137 result = fun(*args, **kwargs)
141 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
142 if self.already_shutdown:
143 raise Exception('Submitted work after shutdown.')
144 self.adjust_task_count(+1)
146 newargs.append(function)
150 result = self._thread_pool_executor.submit(
151 self.run_local_bundle, *newargs, **kwargs
153 result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start))
154 result.add_done_callback(lambda _: self.adjust_task_count(-1))
157 def shutdown(self, wait=True) -> None:
158 if not self.already_shutdown:
159 logger.debug(f'Shutting down threadpool executor {self.title}')
160 print(self.histogram)
161 self._thread_pool_executor.shutdown(wait)
162 self.already_shutdown = True
165 class ProcessExecutor(BaseExecutor):
166 def __init__(self, max_workers=None):
169 if max_workers is not None:
170 workers = max_workers
171 elif 'executors_processpool_size' in config.config:
172 workers = config.config['executors_processpool_size']
173 logger.debug(f'Creating processpool executor with {workers} workers.')
174 self._process_executor = fut.ProcessPoolExecutor(
177 self.already_shutdown = False
179 # This is run in another process; do not adjust task count here.
180 def run_cloud_pickle(self, pickle):
181 fun, args, kwargs = cloudpickle.loads(pickle)
182 logger.debug(f"Running pickled bundle at {fun.__name__}")
183 result = fun(*args, **kwargs)
187 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
188 if self.already_shutdown:
189 raise Exception('Submitted work after shutdown.')
191 self.adjust_task_count(+1)
192 pickle = make_cloud_pickle(function, *args, **kwargs)
193 result = self._process_executor.submit(self.run_cloud_pickle, pickle)
194 result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start))
195 result.add_done_callback(lambda _: self.adjust_task_count(-1))
199 def shutdown(self, wait=True) -> None:
200 if not self.already_shutdown:
201 logger.debug(f'Shutting down processpool executor {self.title}')
202 self._process_executor.shutdown(wait)
203 print(self.histogram)
204 self.already_shutdown = True
206 def __getstate__(self):
207 state = self.__dict__.copy()
208 state['_process_executor'] = None
212 class RemoteExecutorException(Exception):
213 """Thrown when a bundle cannot be executed despite several retries."""
219 class RemoteWorkerRecord:
226 return hash((self.username, self.machine))
229 return f'{self.username}@{self.machine}'
237 worker: Optional[RemoteWorkerRecord]
238 username: Optional[str]
239 machine: Optional[str]
246 slower_than_local_p95: bool
247 slower_than_global_p95: bool
248 src_bundle: BundleDetails
249 is_cancelled: threading.Event
251 backup_bundles: Optional[List[BundleDetails]]
256 if uuid[-9:-2] == '_backup':
258 suffix = f'{uuid[-6:]}_b{self.uuid[-1:]}'
268 fg('marigold yellow'),
271 fg('cornflower blue'),
272 fg('turquoise blue'),
274 fg('lavender purple'),
277 c = colorz[int(uuid[-2:], 16) % len(colorz)]
278 fname = self.fname if self.fname is not None else 'nofname'
279 machine = self.machine if self.machine is not None else 'nomachine'
280 return f'{c}{suffix}/{fname}/{machine}{reset()}'
283 class RemoteExecutorStatus:
284 def __init__(self, total_worker_count: int) -> None:
285 self.worker_count: int = total_worker_count
286 self.known_workers: Set[RemoteWorkerRecord] = set()
287 self.start_time: float = time.time()
288 self.start_per_bundle: Dict[str, float] = defaultdict(float)
289 self.end_per_bundle: Dict[str, float] = defaultdict(float)
290 self.finished_bundle_timings_per_worker: Dict[
291 RemoteWorkerRecord, List[float]
293 self.in_flight_bundles_by_worker: Dict[RemoteWorkerRecord, Set[str]] = {}
294 self.bundle_details_by_uuid: Dict[str, BundleDetails] = {}
295 self.finished_bundle_timings: List[float] = []
296 self.last_periodic_dump: Optional[float] = None
297 self.total_bundles_submitted: int = 0
299 # Protects reads and modification using self. Also used
300 # as a memory fence for modifications to bundle.
301 self.lock: threading.Lock = threading.Lock()
303 def record_acquire_worker(self, worker: RemoteWorkerRecord, uuid: str) -> None:
305 self.record_acquire_worker_already_locked(worker, uuid)
307 def record_acquire_worker_already_locked(
308 self, worker: RemoteWorkerRecord, uuid: str
310 assert self.lock.locked()
311 self.known_workers.add(worker)
312 self.start_per_bundle[uuid] = None
313 x = self.in_flight_bundles_by_worker.get(worker, set())
315 self.in_flight_bundles_by_worker[worker] = x
317 def record_bundle_details(self, details: BundleDetails) -> None:
319 self.record_bundle_details_already_locked(details)
321 def record_bundle_details_already_locked(self, details: BundleDetails) -> None:
322 assert self.lock.locked()
323 self.bundle_details_by_uuid[details.uuid] = details
325 def record_release_worker(
327 worker: RemoteWorkerRecord,
332 self.record_release_worker_already_locked(worker, uuid, was_cancelled)
334 def record_release_worker_already_locked(
336 worker: RemoteWorkerRecord,
340 assert self.lock.locked()
342 self.end_per_bundle[uuid] = ts
343 self.in_flight_bundles_by_worker[worker].remove(uuid)
344 if not was_cancelled:
345 bundle_latency = ts - self.start_per_bundle[uuid]
346 x = self.finished_bundle_timings_per_worker.get(worker, list())
347 x.append(bundle_latency)
348 self.finished_bundle_timings_per_worker[worker] = x
349 self.finished_bundle_timings.append(bundle_latency)
351 def record_processing_began(self, uuid: str):
353 self.start_per_bundle[uuid] = time.time()
355 def total_in_flight(self) -> int:
356 assert self.lock.locked()
358 for worker in self.known_workers:
359 total_in_flight += len(self.in_flight_bundles_by_worker[worker])
360 return total_in_flight
362 def total_idle(self) -> int:
363 assert self.lock.locked()
364 return self.worker_count - self.total_in_flight()
367 assert self.lock.locked()
369 total_finished = len(self.finished_bundle_timings)
370 total_in_flight = self.total_in_flight()
371 ret = f'\n\n{underline()}Remote Executor Pool Status{reset()}: '
373 if len(self.finished_bundle_timings) > 1:
374 qall = numpy.quantile(self.finished_bundle_timings, [0.5, 0.95])
376 f'⏱=∀p50:{qall[0]:.1f}s, ∀p95:{qall[1]:.1f}s, total={ts-self.start_time:.1f}s, '
377 f'✅={total_finished}/{self.total_bundles_submitted}, '
378 f'💻n={total_in_flight}/{self.worker_count}\n'
382 f'⏱={ts-self.start_time:.1f}s, '
383 f'✅={total_finished}/{self.total_bundles_submitted}, '
384 f'💻n={total_in_flight}/{self.worker_count}\n'
387 for worker in self.known_workers:
388 ret += f' {fg("lightning yellow")}{worker.machine}{reset()}: '
389 timings = self.finished_bundle_timings_per_worker.get(worker, [])
393 qworker = numpy.quantile(timings, [0.5, 0.95])
394 ret += f' 💻p50: {qworker[0]:.1f}s, 💻p95: {qworker[1]:.1f}s\n'
398 ret += f' ...finished {count} total bundle(s) so far\n'
399 in_flight = len(self.in_flight_bundles_by_worker[worker])
401 ret += f' ...{in_flight} bundles currently in flight:\n'
402 for bundle_uuid in self.in_flight_bundles_by_worker[worker]:
403 details = self.bundle_details_by_uuid.get(bundle_uuid, None)
404 pid = str(details.pid) if (details and details.pid != 0) else "TBD"
405 if self.start_per_bundle[bundle_uuid] is not None:
406 sec = ts - self.start_per_bundle[bundle_uuid]
407 ret += f' (pid={pid}): {details} for {sec:.1f}s so far '
409 ret += f' {details} setting up / copying data...'
412 if qworker is not None:
414 ret += f'{bg("red")}>💻p95{reset()} '
415 if details is not None:
416 details.slower_than_local_p95 = True
418 if details is not None:
419 details.slower_than_local_p95 = False
423 ret += f'{bg("red")}>∀p95{reset()} '
424 if details is not None:
425 details.slower_than_global_p95 = True
427 details.slower_than_global_p95 = False
431 def periodic_dump(self, total_bundles_submitted: int) -> None:
432 assert self.lock.locked()
433 self.total_bundles_submitted = total_bundles_submitted
435 if self.last_periodic_dump is None or ts - self.last_periodic_dump > 5.0:
437 self.last_periodic_dump = ts
440 class RemoteWorkerSelectionPolicy(ABC):
441 def register_worker_pool(self, workers):
442 self.workers = workers
445 def is_worker_available(self) -> bool:
449 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
453 class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
455 def is_worker_available(self) -> bool:
456 for worker in self.workers:
462 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
464 for worker in self.workers:
465 if worker.machine != machine_to_avoid:
467 for _ in range(worker.count * worker.weight):
468 grabbag.append(worker)
470 if len(grabbag) == 0:
472 f'There are no available workers that avoid {machine_to_avoid}...'
474 for worker in self.workers:
476 for _ in range(worker.count * worker.weight):
477 grabbag.append(worker)
479 if len(grabbag) == 0:
480 logger.warning('There are no available workers?!')
483 worker = random.sample(grabbag, 1)[0]
484 assert worker.count > 0
486 logger.debug(f'Chose worker {worker}')
490 class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
491 def __init__(self) -> None:
495 def is_worker_available(self) -> bool:
496 for worker in self.workers:
503 self, machine_to_avoid: str = None
504 ) -> Optional[RemoteWorkerRecord]:
507 worker = self.workers[x]
511 if x >= len(self.workers):
514 logger.debug(f'Selected worker {worker}')
517 if x >= len(self.workers):
520 msg = 'Unexpectedly could not find a worker, retrying...'
525 class RemoteExecutor(BaseExecutor):
528 workers: List[RemoteWorkerRecord],
529 policy: RemoteWorkerSelectionPolicy,
532 self.workers = workers
534 self.worker_count = 0
535 for worker in self.workers:
536 self.worker_count += worker.count
537 if self.worker_count <= 0:
538 msg = f"We need somewhere to schedule work; count was {self.worker_count}"
540 raise RemoteExecutorException(msg)
541 self.policy.register_worker_pool(self.workers)
542 self.cv = threading.Condition()
544 f'Creating {self.worker_count} local threads, one per remote worker.'
546 self._helper_executor = fut.ThreadPoolExecutor(
547 thread_name_prefix="remote_executor_helper",
548 max_workers=self.worker_count,
550 self.status = RemoteExecutorStatus(self.worker_count)
551 self.total_bundles_submitted = 0
552 self.backup_lock = threading.Lock()
553 self.last_backup = None
555 self.heartbeat_thread,
556 self.heartbeat_stop_event,
557 ) = self.run_periodic_heartbeat()
558 self.already_shutdown = False
561 def run_periodic_heartbeat(self, stop_event: threading.Event) -> None:
562 while not stop_event.is_set():
564 logger.debug('Running periodic heartbeat code...')
566 logger.debug('Periodic heartbeat thread shutting down.')
568 def heartbeat(self) -> None:
569 # Note: this is invoked on a background thread, not an
570 # executor thread. Be careful what you do with it b/c it
571 # needs to get back and dump status again periodically.
572 with self.status.lock:
573 self.status.periodic_dump(self.total_bundles_submitted)
575 # Look for bundles to reschedule via executor.submit
576 if config.config['executors_schedule_remote_backups']:
577 self.maybe_schedule_backup_bundles()
579 def maybe_schedule_backup_bundles(self):
580 assert self.status.lock.locked()
581 num_done = len(self.status.finished_bundle_timings)
582 num_idle_workers = self.worker_count - self.task_count
586 and num_idle_workers > 1
587 and (self.last_backup is None or (now - self.last_backup > 9.0))
588 and self.backup_lock.acquire(blocking=False)
591 assert self.backup_lock.locked()
593 bundle_to_backup = None
598 ) in self.status.in_flight_bundles_by_worker.items():
600 # Prefer to schedule backups of bundles running on
603 for record in self.workers:
604 if worker.machine == record.machine:
605 base_score = float(record.weight)
606 base_score = 1.0 / base_score
608 base_score = int(base_score)
611 for uuid in bundle_uuids:
612 bundle = self.status.bundle_details_by_uuid.get(uuid, None)
615 and bundle.src_bundle is None
616 and bundle.backup_bundles is not None
620 # Schedule backups of bundles running
621 # longer; especially those that are
623 start_ts = self.status.start_per_bundle[uuid]
624 if start_ts is not None:
625 runtime = now - start_ts
628 f'score[{bundle}] => {score} # latency boost'
631 if bundle.slower_than_local_p95:
634 f'score[{bundle}] => {score} # >worker p95'
637 if bundle.slower_than_global_p95:
640 f'score[{bundle}] => {score} # >global p95'
643 # Prefer backups of bundles that don't
644 # have backups already.
645 backup_count = len(bundle.backup_bundles)
646 if backup_count == 0:
648 elif backup_count == 1:
650 elif backup_count == 2:
655 f'score[{bundle}] => {score} # {backup_count} dup backup factor'
659 best_score is None or score > best_score
661 bundle_to_backup = bundle
662 assert bundle is not None
663 assert bundle.backup_bundles is not None
664 assert bundle.src_bundle is None
667 # Note: this is all still happening on the heartbeat
668 # runner thread. That's ok because
669 # schedule_backup_for_bundle uses the executor to
670 # submit the bundle again which will cause it to be
671 # picked up by a worker thread and allow this thread
672 # to return to run future heartbeats.
673 if bundle_to_backup is not None:
674 self.last_backup = now
676 f'=====> SCHEDULING BACKUP {bundle_to_backup} (score={best_score:.1f}) <====='
678 self.schedule_backup_for_bundle(bundle_to_backup)
680 self.backup_lock.release()
682 def is_worker_available(self) -> bool:
683 return self.policy.is_worker_available()
686 self, machine_to_avoid: str = None
687 ) -> Optional[RemoteWorkerRecord]:
688 return self.policy.acquire_worker(machine_to_avoid)
690 def find_available_worker_or_block(
691 self, machine_to_avoid: str = None
692 ) -> RemoteWorkerRecord:
694 while not self.is_worker_available():
696 worker = self.acquire_worker(machine_to_avoid)
697 if worker is not None:
699 msg = "We should never reach this point in the code"
703 def release_worker(self, bundle: BundleDetails, *, was_cancelled=True) -> None:
704 worker = bundle.worker
705 assert worker is not None
706 logger.debug(f'Released worker {worker}')
707 self.status.record_release_worker(
715 self.adjust_task_count(-1)
717 def check_if_cancelled(self, bundle: BundleDetails) -> bool:
718 with self.status.lock:
719 if bundle.is_cancelled.wait(timeout=0.0):
720 logger.debug(f'Bundle {bundle.uuid} is cancelled, bail out.')
721 bundle.was_cancelled = True
725 def launch(self, bundle: BundleDetails, override_avoid_machine=None) -> Any:
726 """Find a worker for bundle or block until one is available."""
727 self.adjust_task_count(+1)
729 hostname = bundle.hostname
730 avoid_machine = override_avoid_machine
731 is_original = bundle.src_bundle is None
733 # Try not to schedule a backup on the same host as the original.
734 if avoid_machine is None and bundle.src_bundle is not None:
735 avoid_machine = bundle.src_bundle.machine
737 while worker is None:
738 worker = self.find_available_worker_or_block(avoid_machine)
741 # Ok, found a worker.
742 bundle.worker = worker
743 machine = bundle.machine = worker.machine
744 username = bundle.username = worker.username
745 self.status.record_acquire_worker(worker, uuid)
746 logger.debug(f'{bundle}: Running bundle on {worker}...')
748 # Before we do any work, make sure the bundle is still viable.
749 # It may have been some time between when it was submitted and
750 # now due to lack of worker availability and someone else may
751 # have already finished it.
752 if self.check_if_cancelled(bundle):
754 return self.process_work_result(bundle)
755 except Exception as e:
757 f'{bundle}: bundle says it\'s cancelled upfront but no results?!'
759 self.release_worker(bundle)
761 # Weird. We are the original owner of this
762 # bundle. For it to have been cancelled, a backup
763 # must have already started and completed before
764 # we even for started. Moreover, the backup says
765 # it is done but we can't find the results it
766 # should have copied over. Reschedule the whole
770 f'{bundle}: We are the original owner thread and yet there are '
771 + 'no results for this bundle. This is unexpected and bad.'
773 return self.emergency_retry_nasty_bundle(bundle)
775 # Expected(?). We're a backup and our bundle is
776 # cancelled before we even got started. Something
777 # went bad in process_work_result (I acutually don't
778 # see what?) but probably not worth worrying
779 # about. Let the original thread worry about
780 # either finding the results or complaining about
784 # Send input code / data to worker machine if it's not local.
785 if hostname not in machine:
788 f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}'
790 start_ts = time.time()
791 logger.info(f"{bundle}: Copying work to {worker} via {cmd}.")
793 xfer_latency = time.time() - start_ts
794 logger.debug(f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s.")
795 except Exception as e:
796 self.release_worker(bundle)
798 # Weird. We tried to copy the code to the worker and it failed...
799 # And we're the original bundle. We have to retry.
802 f"{bundle}: Failed to send instructions to the worker machine?! "
803 + "This is not expected; we\'re the original bundle so this shouldn\'t "
804 + "be a race condition. Attempting an emergency retry..."
806 return self.emergency_retry_nasty_bundle(bundle)
808 # This is actually expected; we're a backup.
809 # There's a race condition where someone else
810 # already finished the work and removed the source
811 # code file before we could copy it. No biggie.
812 msg = f'{bundle}: Failed to send instructions to the worker machine... '
813 msg += 'We\'re a backup and this may be caused by the original (or some '
814 msg += 'other backup) already finishing this work. Ignoring this.'
818 # Kick off the work. Note that if this fails we let
819 # wait_for_process deal with it.
820 self.status.record_processing_began(uuid)
822 f'{SSH} {bundle.username}@{bundle.machine} '
823 f'"source py38-venv/bin/activate &&'
824 f' /home/scott/lib/python_modules/remote_worker.py'
825 f' --code_file {bundle.code_file} --result_file {bundle.result_file}"'
827 logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...')
828 p = cmd_in_background(cmd, silent=True)
831 f'{bundle}: Local ssh process pid={p.pid}; remote worker is {machine}.'
833 return self.wait_for_process(p, bundle, 0)
835 def wait_for_process(
836 self, p: subprocess.Popen, bundle: BundleDetails, depth: int
838 machine = bundle.machine
842 f"I've gotten repeated errors waiting on this bundle; giving up on pid={pid}."
845 self.release_worker(bundle)
846 return self.emergency_retry_nasty_bundle(bundle)
848 # Spin until either the ssh job we scheduled finishes the
849 # bundle or some backup worker signals that they finished it
854 except subprocess.TimeoutExpired:
855 if self.check_if_cancelled(bundle):
857 f'{bundle}: looks like another worker finished bundle...'
861 logger.info(f"{bundle}: pid {pid} ({machine}) is finished!")
865 # If we get here we believe the bundle is done; either the ssh
866 # subprocess finished (hopefully successfully) or we noticed
867 # that some other worker seems to have completed the bundle
868 # and we're bailing out.
870 ret = self.process_work_result(bundle)
871 if ret is not None and p is not None:
875 # Something went wrong; e.g. we could not copy the results
876 # back, cleanup after ourselves on the remote machine, or
877 # unpickle the results we got from the remove machine. If we
878 # still have an active ssh subprocess, keep waiting on it.
879 # Otherwise, time for an emergency reschedule.
880 except Exception as e:
882 logger.error(f'{bundle}: Something unexpected just happened...')
884 msg = f"{bundle}: Failed to wrap up \"done\" bundle, re-waiting on active ssh."
886 return self.wait_for_process(p, bundle, depth + 1)
888 self.release_worker(bundle)
889 return self.emergency_retry_nasty_bundle(bundle)
891 def process_work_result(self, bundle: BundleDetails) -> Any:
892 with self.status.lock:
893 is_original = bundle.src_bundle is None
894 was_cancelled = bundle.was_cancelled
895 username = bundle.username
896 machine = bundle.machine
897 result_file = bundle.result_file
898 code_file = bundle.code_file
900 # Whether original or backup, if we finished first we must
901 # fetch the results if the computation happened on a
903 bundle.end_ts = time.time()
904 if not was_cancelled:
905 assert bundle.machine is not None
906 if bundle.hostname not in bundle.machine:
907 cmd = f'{SCP} {username}@{machine}:{result_file} {result_file} 2>/dev/null'
909 f"{bundle}: Fetching results back from {username}@{machine} via {cmd}"
912 # If either of these throw they are handled in
918 except Exception as e:
926 f'{SSH} {username}@{machine}'
927 f' "/bin/rm -f {code_file} {result_file}"'
930 f'Fetching results back took {time.time() - bundle.end_ts:.1f}s.'
932 dur = bundle.end_ts - bundle.start_ts
933 self.histogram.add_item(dur)
935 # Only the original worker should unpickle the file contents
936 # though since it's the only one whose result matters. The
937 # original is also the only job that may delete result_file
938 # from disk. Note that the original may have been cancelled
939 # if one of the backups finished first; it still must read the
942 logger.debug(f"{bundle}: Unpickling {result_file}.")
944 with open(result_file, 'rb') as rb:
945 serialized = rb.read()
946 result = cloudpickle.loads(serialized)
947 except Exception as e:
949 msg = f'Failed to load {result_file}... this is bad news.'
951 self.release_worker(bundle)
953 # Re-raise the exception; the code in wait_for_process may
954 # decide to emergency_retry_nasty_bundle here.
956 logger.debug(f'Removing local (master) {code_file} and {result_file}.')
957 os.remove(f'{result_file}')
958 os.remove(f'{code_file}')
960 # Notify any backups that the original is done so they
961 # should stop ASAP. Do this whether or not we
962 # finished first since there could be more than one
964 if bundle.backup_bundles is not None:
965 for backup in bundle.backup_bundles:
967 f'{bundle}: Notifying backup {backup.uuid} that it\'s cancelled'
969 backup.is_cancelled.set()
971 # This is a backup job and, by now, we have already fetched
972 # the bundle results.
974 # Backup results don't matter, they just need to leave the
975 # result file in the right place for their originals to
976 # read/unpickle later.
979 # Tell the original to stop if we finished first.
980 if not was_cancelled:
982 f'{bundle}: Notifying original {bundle.src_bundle.uuid} we beat them to it.'
984 bundle.src_bundle.is_cancelled.set()
985 self.release_worker(bundle, was_cancelled=was_cancelled)
988 def create_original_bundle(self, pickle, fname: str):
989 from string_utils import generate_uuid
991 uuid = generate_uuid(omit_dashes=True)
992 code_file = f'/tmp/{uuid}.code.bin'
993 result_file = f'/tmp/{uuid}.result.bin'
995 logger.debug(f'Writing pickled code to {code_file}')
996 with open(f'{code_file}', 'wb') as wb:
999 bundle = BundleDetails(
1000 pickled_code=pickle,
1006 hostname=platform.node(),
1007 code_file=code_file,
1008 result_file=result_file,
1010 start_ts=time.time(),
1012 slower_than_local_p95=False,
1013 slower_than_global_p95=False,
1015 is_cancelled=threading.Event(),
1016 was_cancelled=False,
1020 self.status.record_bundle_details(bundle)
1021 logger.debug(f'{bundle}: Created an original bundle')
1024 def create_backup_bundle(self, src_bundle: BundleDetails):
1025 assert src_bundle.backup_bundles is not None
1026 n = len(src_bundle.backup_bundles)
1027 uuid = src_bundle.uuid + f'_backup#{n}'
1029 backup_bundle = BundleDetails(
1030 pickled_code=src_bundle.pickled_code,
1032 fname=src_bundle.fname,
1036 hostname=src_bundle.hostname,
1037 code_file=src_bundle.code_file,
1038 result_file=src_bundle.result_file,
1040 start_ts=time.time(),
1042 slower_than_local_p95=False,
1043 slower_than_global_p95=False,
1044 src_bundle=src_bundle,
1045 is_cancelled=threading.Event(),
1046 was_cancelled=False,
1047 backup_bundles=None, # backup backups not allowed
1050 src_bundle.backup_bundles.append(backup_bundle)
1051 self.status.record_bundle_details_already_locked(backup_bundle)
1052 logger.debug(f'{backup_bundle}: Created a backup bundle')
1053 return backup_bundle
1055 def schedule_backup_for_bundle(self, src_bundle: BundleDetails):
1056 assert self.status.lock.locked()
1057 assert src_bundle is not None
1058 backup_bundle = self.create_backup_bundle(src_bundle)
1060 f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...'
1062 self._helper_executor.submit(self.launch, backup_bundle)
1064 # Results from backups don't matter; if they finish first
1065 # they will move the result_file to this machine and let
1066 # the original pick them up and unpickle them.
1068 def emergency_retry_nasty_bundle(self, bundle: BundleDetails) -> fut.Future:
1069 is_original = bundle.src_bundle is None
1070 bundle.worker = None
1071 avoid_last_machine = bundle.machine
1072 bundle.machine = None
1073 bundle.username = None
1074 bundle.failure_count += 1
1080 if bundle.failure_count > retry_limit:
1082 f'{bundle}: Tried this bundle too many times already ({retry_limit}x); giving up.'
1085 raise RemoteExecutorException(
1086 f'{bundle}: This bundle can\'t be completed despite several backups and retries'
1090 f'{bundle}: At least it\'s only a backup; better luck with the others.'
1094 msg = f'>>> Emergency rescheduling {bundle} because of unexected errors (wtf?!) <<<'
1097 return self.launch(bundle, avoid_last_machine)
1100 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
1101 if self.already_shutdown:
1102 raise Exception('Submitted work after shutdown.')
1103 pickle = make_cloud_pickle(function, *args, **kwargs)
1104 bundle = self.create_original_bundle(pickle, function.__name__)
1105 self.total_bundles_submitted += 1
1106 return self._helper_executor.submit(self.launch, bundle)
1109 def shutdown(self, wait=True) -> None:
1110 if not self.already_shutdown:
1111 logging.debug(f'Shutting down RemoteExecutor {self.title}')
1112 self.heartbeat_stop_event.set()
1113 self.heartbeat_thread.join()
1114 self._helper_executor.shutdown(wait)
1115 print(self.histogram)
1116 self.already_shutdown = True
1120 class DefaultExecutors(object):
1122 self.thread_executor: Optional[ThreadExecutor] = None
1123 self.process_executor: Optional[ProcessExecutor] = None
1124 self.remote_executor: Optional[RemoteExecutor] = None
1126 def ping(self, host) -> bool:
1127 logger.debug(f'RUN> ping -c 1 {host}')
1129 x = cmd_with_timeout(
1130 f'ping -c 1 {host} >/dev/null 2>/dev/null', timeout_seconds=1.0
1136 def thread_pool(self) -> ThreadExecutor:
1137 if self.thread_executor is None:
1138 self.thread_executor = ThreadExecutor()
1139 return self.thread_executor
1141 def process_pool(self) -> ProcessExecutor:
1142 if self.process_executor is None:
1143 self.process_executor = ProcessExecutor()
1144 return self.process_executor
1146 def remote_pool(self) -> RemoteExecutor:
1147 if self.remote_executor is None:
1148 logger.info('Looking for some helper machines...')
1149 pool: List[RemoteWorkerRecord] = []
1150 if self.ping('cheetah.house'):
1151 logger.info('Found cheetah.house')
1155 machine='cheetah.house',
1160 if self.ping('meerkat.cabin'):
1161 logger.info('Found meerkat.cabin')
1165 machine='meerkat.cabin',
1170 if self.ping('wannabe.house'):
1171 logger.info('Found wannabe.house')
1175 machine='wannabe.house',
1180 if self.ping('puma.cabin'):
1181 logger.info('Found puma.cabin')
1185 machine='puma.cabin',
1190 if self.ping('backup.house'):
1191 logger.info('Found backup.house')
1195 machine='backup.house',
1201 # The controller machine has a lot to do; go easy on it.
1203 if record.machine == platform.node() and record.count > 1:
1204 logger.info(f'Reducing workload for {record.machine}.')
1207 policy = WeightedRandomRemoteWorkerSelectionPolicy()
1208 policy.register_worker_pool(pool)
1209 self.remote_executor = RemoteExecutor(pool, policy)
1210 return self.remote_executor
1212 def shutdown(self) -> None:
1213 if self.thread_executor is not None:
1214 self.thread_executor.shutdown()
1215 self.thread_executor = None
1216 if self.process_executor is not None:
1217 self.process_executor.shutdown()
1218 self.process_executor = None
1219 if self.remote_executor is not None:
1220 self.remote_executor.shutdown()
1221 self.remote_executor = None