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))
158 def shutdown(self, wait=True) -> None:
159 if not self.already_shutdown:
160 logger.debug(f'Shutting down threadpool executor {self.title}')
161 print(self.histogram.__repr__(label_formatter='%d'))
162 self._thread_pool_executor.shutdown(wait)
163 self.already_shutdown = True
166 class ProcessExecutor(BaseExecutor):
167 def __init__(self, max_workers=None):
170 if max_workers is not None:
171 workers = max_workers
172 elif 'executors_processpool_size' in config.config:
173 workers = config.config['executors_processpool_size']
174 logger.debug(f'Creating processpool executor with {workers} workers.')
175 self._process_executor = fut.ProcessPoolExecutor(
178 self.already_shutdown = False
180 # This is run in another process; do not adjust task count here.
181 def run_cloud_pickle(self, pickle):
182 fun, args, kwargs = cloudpickle.loads(pickle)
183 logger.debug(f"Running pickled bundle at {fun.__name__}")
184 result = fun(*args, **kwargs)
188 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
189 if self.already_shutdown:
190 raise Exception('Submitted work after shutdown.')
192 self.adjust_task_count(+1)
193 pickle = make_cloud_pickle(function, *args, **kwargs)
194 result = self._process_executor.submit(self.run_cloud_pickle, pickle)
195 result.add_done_callback(lambda _: self.histogram.add_item(time.time() - start))
196 result.add_done_callback(lambda _: self.adjust_task_count(-1))
200 def shutdown(self, wait=True) -> None:
201 if not self.already_shutdown:
202 logger.debug(f'Shutting down processpool executor {self.title}')
203 self._process_executor.shutdown(wait)
204 print(self.histogram.__repr__(label_formatter='%d'))
205 self.already_shutdown = True
207 def __getstate__(self):
208 state = self.__dict__.copy()
209 state['_process_executor'] = None
213 class RemoteExecutorException(Exception):
214 """Thrown when a bundle cannot be executed despite several retries."""
220 class RemoteWorkerRecord:
227 return hash((self.username, self.machine))
230 return f'{self.username}@{self.machine}'
238 worker: Optional[RemoteWorkerRecord]
239 username: Optional[str]
240 machine: Optional[str]
247 slower_than_local_p95: bool
248 slower_than_global_p95: bool
249 src_bundle: BundleDetails
250 is_cancelled: threading.Event
252 backup_bundles: Optional[List[BundleDetails]]
257 if uuid[-9:-2] == '_backup':
259 suffix = f'{uuid[-6:]}_b{self.uuid[-1:]}'
269 fg('marigold yellow'),
272 fg('cornflower blue'),
273 fg('turquoise blue'),
275 fg('lavender purple'),
278 c = colorz[int(uuid[-2:], 16) % len(colorz)]
279 fname = self.fname if self.fname is not None else 'nofname'
280 machine = self.machine if self.machine is not None else 'nomachine'
281 return f'{c}{suffix}/{fname}/{machine}{reset()}'
284 class RemoteExecutorStatus:
285 def __init__(self, total_worker_count: int) -> None:
286 self.worker_count: int = total_worker_count
287 self.known_workers: Set[RemoteWorkerRecord] = set()
288 self.start_time: float = time.time()
289 self.start_per_bundle: Dict[str, float] = defaultdict(float)
290 self.end_per_bundle: Dict[str, float] = defaultdict(float)
291 self.finished_bundle_timings_per_worker: Dict[
292 RemoteWorkerRecord, List[float]
294 self.in_flight_bundles_by_worker: Dict[RemoteWorkerRecord, Set[str]] = {}
295 self.bundle_details_by_uuid: Dict[str, BundleDetails] = {}
296 self.finished_bundle_timings: List[float] = []
297 self.last_periodic_dump: Optional[float] = None
298 self.total_bundles_submitted: int = 0
300 # Protects reads and modification using self. Also used
301 # as a memory fence for modifications to bundle.
302 self.lock: threading.Lock = threading.Lock()
304 def record_acquire_worker(self, worker: RemoteWorkerRecord, uuid: str) -> None:
306 self.record_acquire_worker_already_locked(worker, uuid)
308 def record_acquire_worker_already_locked(
309 self, worker: RemoteWorkerRecord, uuid: str
311 assert self.lock.locked()
312 self.known_workers.add(worker)
313 self.start_per_bundle[uuid] = None
314 x = self.in_flight_bundles_by_worker.get(worker, set())
316 self.in_flight_bundles_by_worker[worker] = x
318 def record_bundle_details(self, details: BundleDetails) -> None:
320 self.record_bundle_details_already_locked(details)
322 def record_bundle_details_already_locked(self, details: BundleDetails) -> None:
323 assert self.lock.locked()
324 self.bundle_details_by_uuid[details.uuid] = details
326 def record_release_worker(
328 worker: RemoteWorkerRecord,
333 self.record_release_worker_already_locked(worker, uuid, was_cancelled)
335 def record_release_worker_already_locked(
337 worker: RemoteWorkerRecord,
341 assert self.lock.locked()
343 self.end_per_bundle[uuid] = ts
344 self.in_flight_bundles_by_worker[worker].remove(uuid)
345 if not was_cancelled:
346 bundle_latency = ts - self.start_per_bundle[uuid]
347 x = self.finished_bundle_timings_per_worker.get(worker, list())
348 x.append(bundle_latency)
349 self.finished_bundle_timings_per_worker[worker] = x
350 self.finished_bundle_timings.append(bundle_latency)
352 def record_processing_began(self, uuid: str):
354 self.start_per_bundle[uuid] = time.time()
356 def total_in_flight(self) -> int:
357 assert self.lock.locked()
359 for worker in self.known_workers:
360 total_in_flight += len(self.in_flight_bundles_by_worker[worker])
361 return total_in_flight
363 def total_idle(self) -> int:
364 assert self.lock.locked()
365 return self.worker_count - self.total_in_flight()
368 assert self.lock.locked()
370 total_finished = len(self.finished_bundle_timings)
371 total_in_flight = self.total_in_flight()
372 ret = f'\n\n{underline()}Remote Executor Pool Status{reset()}: '
374 if len(self.finished_bundle_timings) > 1:
375 qall = numpy.quantile(self.finished_bundle_timings, [0.5, 0.95])
377 f'⏱=∀p50:{qall[0]:.1f}s, ∀p95:{qall[1]:.1f}s, total={ts-self.start_time:.1f}s, '
378 f'✅={total_finished}/{self.total_bundles_submitted}, '
379 f'💻n={total_in_flight}/{self.worker_count}\n'
383 f'⏱={ts-self.start_time:.1f}s, '
384 f'✅={total_finished}/{self.total_bundles_submitted}, '
385 f'💻n={total_in_flight}/{self.worker_count}\n'
388 for worker in self.known_workers:
389 ret += f' {fg("lightning yellow")}{worker.machine}{reset()}: '
390 timings = self.finished_bundle_timings_per_worker.get(worker, [])
394 qworker = numpy.quantile(timings, [0.5, 0.95])
395 ret += f' 💻p50: {qworker[0]:.1f}s, 💻p95: {qworker[1]:.1f}s\n'
399 ret += f' ...finished {count} total bundle(s) so far\n'
400 in_flight = len(self.in_flight_bundles_by_worker[worker])
402 ret += f' ...{in_flight} bundles currently in flight:\n'
403 for bundle_uuid in self.in_flight_bundles_by_worker[worker]:
404 details = self.bundle_details_by_uuid.get(bundle_uuid, None)
405 pid = str(details.pid) if (details and details.pid != 0) else "TBD"
406 if self.start_per_bundle[bundle_uuid] is not None:
407 sec = ts - self.start_per_bundle[bundle_uuid]
408 ret += f' (pid={pid}): {details} for {sec:.1f}s so far '
410 ret += f' {details} setting up / copying data...'
413 if qworker is not None:
415 ret += f'{bg("red")}>💻p95{reset()} '
416 if details is not None:
417 details.slower_than_local_p95 = True
419 if details is not None:
420 details.slower_than_local_p95 = False
424 ret += f'{bg("red")}>∀p95{reset()} '
425 if details is not None:
426 details.slower_than_global_p95 = True
428 details.slower_than_global_p95 = False
432 def periodic_dump(self, total_bundles_submitted: int) -> None:
433 assert self.lock.locked()
434 self.total_bundles_submitted = total_bundles_submitted
436 if self.last_periodic_dump is None or ts - self.last_periodic_dump > 5.0:
438 self.last_periodic_dump = ts
441 class RemoteWorkerSelectionPolicy(ABC):
442 def register_worker_pool(self, workers):
443 self.workers = workers
446 def is_worker_available(self) -> bool:
450 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
454 class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
456 def is_worker_available(self) -> bool:
457 for worker in self.workers:
463 def acquire_worker(self, machine_to_avoid=None) -> Optional[RemoteWorkerRecord]:
465 for worker in self.workers:
466 if worker.machine != machine_to_avoid:
468 for _ in range(worker.count * worker.weight):
469 grabbag.append(worker)
471 if len(grabbag) == 0:
473 f'There are no available workers that avoid {machine_to_avoid}...'
475 for worker in self.workers:
477 for _ in range(worker.count * worker.weight):
478 grabbag.append(worker)
480 if len(grabbag) == 0:
481 logger.warning('There are no available workers?!')
484 worker = random.sample(grabbag, 1)[0]
485 assert worker.count > 0
487 logger.debug(f'Chose worker {worker}')
491 class RoundRobinRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
492 def __init__(self) -> None:
496 def is_worker_available(self) -> bool:
497 for worker in self.workers:
504 self, machine_to_avoid: str = None
505 ) -> Optional[RemoteWorkerRecord]:
508 worker = self.workers[x]
512 if x >= len(self.workers):
515 logger.debug(f'Selected worker {worker}')
518 if x >= len(self.workers):
521 msg = 'Unexpectedly could not find a worker, retrying...'
526 class RemoteExecutor(BaseExecutor):
529 workers: List[RemoteWorkerRecord],
530 policy: RemoteWorkerSelectionPolicy,
533 self.workers = workers
535 self.worker_count = 0
536 for worker in self.workers:
537 self.worker_count += worker.count
538 if self.worker_count <= 0:
539 msg = f"We need somewhere to schedule work; count was {self.worker_count}"
541 raise RemoteExecutorException(msg)
542 self.policy.register_worker_pool(self.workers)
543 self.cv = threading.Condition()
545 f'Creating {self.worker_count} local threads, one per remote worker.'
547 self._helper_executor = fut.ThreadPoolExecutor(
548 thread_name_prefix="remote_executor_helper",
549 max_workers=self.worker_count,
551 self.status = RemoteExecutorStatus(self.worker_count)
552 self.total_bundles_submitted = 0
553 self.backup_lock = threading.Lock()
554 self.last_backup = None
556 self.heartbeat_thread,
557 self.heartbeat_stop_event,
558 ) = self.run_periodic_heartbeat()
559 self.already_shutdown = False
562 def run_periodic_heartbeat(self, stop_event: threading.Event) -> None:
563 while not stop_event.is_set():
565 logger.debug('Running periodic heartbeat code...')
567 logger.debug('Periodic heartbeat thread shutting down.')
569 def heartbeat(self) -> None:
570 # Note: this is invoked on a background thread, not an
571 # executor thread. Be careful what you do with it b/c it
572 # needs to get back and dump status again periodically.
573 with self.status.lock:
574 self.status.periodic_dump(self.total_bundles_submitted)
576 # Look for bundles to reschedule via executor.submit
577 if config.config['executors_schedule_remote_backups']:
578 self.maybe_schedule_backup_bundles()
580 def maybe_schedule_backup_bundles(self):
581 assert self.status.lock.locked()
582 num_done = len(self.status.finished_bundle_timings)
583 num_idle_workers = self.worker_count - self.task_count
587 and num_idle_workers > 1
588 and (self.last_backup is None or (now - self.last_backup > 9.0))
589 and self.backup_lock.acquire(blocking=False)
592 assert self.backup_lock.locked()
594 bundle_to_backup = None
599 ) in self.status.in_flight_bundles_by_worker.items():
601 # Prefer to schedule backups of bundles running on
604 for record in self.workers:
605 if worker.machine == record.machine:
606 base_score = float(record.weight)
607 base_score = 1.0 / base_score
609 base_score = int(base_score)
612 for uuid in bundle_uuids:
613 bundle = self.status.bundle_details_by_uuid.get(uuid, None)
616 and bundle.src_bundle is None
617 and bundle.backup_bundles is not None
621 # Schedule backups of bundles running
622 # longer; especially those that are
624 start_ts = self.status.start_per_bundle[uuid]
625 if start_ts is not None:
626 runtime = now - start_ts
629 f'score[{bundle}] => {score} # latency boost'
632 if bundle.slower_than_local_p95:
635 f'score[{bundle}] => {score} # >worker p95'
638 if bundle.slower_than_global_p95:
641 f'score[{bundle}] => {score} # >global p95'
644 # Prefer backups of bundles that don't
645 # have backups already.
646 backup_count = len(bundle.backup_bundles)
647 if backup_count == 0:
649 elif backup_count == 1:
651 elif backup_count == 2:
656 f'score[{bundle}] => {score} # {backup_count} dup backup factor'
660 best_score is None or score > best_score
662 bundle_to_backup = bundle
663 assert bundle is not None
664 assert bundle.backup_bundles is not None
665 assert bundle.src_bundle is None
668 # Note: this is all still happening on the heartbeat
669 # runner thread. That's ok because
670 # schedule_backup_for_bundle uses the executor to
671 # submit the bundle again which will cause it to be
672 # picked up by a worker thread and allow this thread
673 # to return to run future heartbeats.
674 if bundle_to_backup is not None:
675 self.last_backup = now
677 f'=====> SCHEDULING BACKUP {bundle_to_backup} (score={best_score:.1f}) <====='
679 self.schedule_backup_for_bundle(bundle_to_backup)
681 self.backup_lock.release()
683 def is_worker_available(self) -> bool:
684 return self.policy.is_worker_available()
687 self, machine_to_avoid: str = None
688 ) -> Optional[RemoteWorkerRecord]:
689 return self.policy.acquire_worker(machine_to_avoid)
691 def find_available_worker_or_block(
692 self, machine_to_avoid: str = None
693 ) -> RemoteWorkerRecord:
695 while not self.is_worker_available():
697 worker = self.acquire_worker(machine_to_avoid)
698 if worker is not None:
700 msg = "We should never reach this point in the code"
704 def release_worker(self, bundle: BundleDetails, *, was_cancelled=True) -> None:
705 worker = bundle.worker
706 assert worker is not None
707 logger.debug(f'Released worker {worker}')
708 self.status.record_release_worker(
716 self.adjust_task_count(-1)
718 def check_if_cancelled(self, bundle: BundleDetails) -> bool:
719 with self.status.lock:
720 if bundle.is_cancelled.wait(timeout=0.0):
721 logger.debug(f'Bundle {bundle.uuid} is cancelled, bail out.')
722 bundle.was_cancelled = True
726 def launch(self, bundle: BundleDetails, override_avoid_machine=None) -> Any:
727 """Find a worker for bundle or block until one is available."""
728 self.adjust_task_count(+1)
730 hostname = bundle.hostname
731 avoid_machine = override_avoid_machine
732 is_original = bundle.src_bundle is None
734 # Try not to schedule a backup on the same host as the original.
735 if avoid_machine is None and bundle.src_bundle is not None:
736 avoid_machine = bundle.src_bundle.machine
738 while worker is None:
739 worker = self.find_available_worker_or_block(avoid_machine)
742 # Ok, found a worker.
743 bundle.worker = worker
744 machine = bundle.machine = worker.machine
745 username = bundle.username = worker.username
746 self.status.record_acquire_worker(worker, uuid)
747 logger.debug(f'{bundle}: Running bundle on {worker}...')
749 # Before we do any work, make sure the bundle is still viable.
750 # It may have been some time between when it was submitted and
751 # now due to lack of worker availability and someone else may
752 # have already finished it.
753 if self.check_if_cancelled(bundle):
755 return self.process_work_result(bundle)
756 except Exception as e:
758 f'{bundle}: bundle says it\'s cancelled upfront but no results?!'
760 self.release_worker(bundle)
762 # Weird. We are the original owner of this
763 # bundle. For it to have been cancelled, a backup
764 # must have already started and completed before
765 # we even for started. Moreover, the backup says
766 # it is done but we can't find the results it
767 # should have copied over. Reschedule the whole
771 f'{bundle}: We are the original owner thread and yet there are '
772 + 'no results for this bundle. This is unexpected and bad.'
774 return self.emergency_retry_nasty_bundle(bundle)
776 # Expected(?). We're a backup and our bundle is
777 # cancelled before we even got started. Something
778 # went bad in process_work_result (I acutually don't
779 # see what?) but probably not worth worrying
780 # about. Let the original thread worry about
781 # either finding the results or complaining about
785 # Send input code / data to worker machine if it's not local.
786 if hostname not in machine:
789 f'{SCP} {bundle.code_file} {username}@{machine}:{bundle.code_file}'
791 start_ts = time.time()
792 logger.info(f"{bundle}: Copying work to {worker} via {cmd}.")
794 xfer_latency = time.time() - start_ts
795 logger.debug(f"{bundle}: Copying to {worker} took {xfer_latency:.1f}s.")
796 except Exception as e:
797 self.release_worker(bundle)
799 # Weird. We tried to copy the code to the worker and it failed...
800 # And we're the original bundle. We have to retry.
803 f"{bundle}: Failed to send instructions to the worker machine?! "
804 + "This is not expected; we\'re the original bundle so this shouldn\'t "
805 + "be a race condition. Attempting an emergency retry..."
807 return self.emergency_retry_nasty_bundle(bundle)
809 # This is actually expected; we're a backup.
810 # There's a race condition where someone else
811 # already finished the work and removed the source
812 # code file before we could copy it. No biggie.
813 msg = f'{bundle}: Failed to send instructions to the worker machine... '
814 msg += 'We\'re a backup and this may be caused by the original (or some '
815 msg += 'other backup) already finishing this work. Ignoring this.'
819 # Kick off the work. Note that if this fails we let
820 # wait_for_process deal with it.
821 self.status.record_processing_began(uuid)
823 f'{SSH} {bundle.username}@{bundle.machine} '
824 f'"source py38-venv/bin/activate &&'
825 f' /home/scott/lib/python_modules/remote_worker.py'
826 f' --code_file {bundle.code_file} --result_file {bundle.result_file}"'
828 logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...')
829 p = cmd_in_background(cmd, silent=True)
832 f'{bundle}: Local ssh process pid={p.pid}; remote worker is {machine}.'
834 return self.wait_for_process(p, bundle, 0)
836 def wait_for_process(
837 self, p: subprocess.Popen, bundle: BundleDetails, depth: int
839 machine = bundle.machine
843 f"I've gotten repeated errors waiting on this bundle; giving up on pid={pid}."
846 self.release_worker(bundle)
847 return self.emergency_retry_nasty_bundle(bundle)
849 # Spin until either the ssh job we scheduled finishes the
850 # bundle or some backup worker signals that they finished it
855 except subprocess.TimeoutExpired:
856 if self.check_if_cancelled(bundle):
858 f'{bundle}: looks like another worker finished bundle...'
862 logger.info(f"{bundle}: pid {pid} ({machine}) is finished!")
866 # If we get here we believe the bundle is done; either the ssh
867 # subprocess finished (hopefully successfully) or we noticed
868 # that some other worker seems to have completed the bundle
869 # and we're bailing out.
871 ret = self.process_work_result(bundle)
872 if ret is not None and p is not None:
876 # Something went wrong; e.g. we could not copy the results
877 # back, cleanup after ourselves on the remote machine, or
878 # unpickle the results we got from the remove machine. If we
879 # still have an active ssh subprocess, keep waiting on it.
880 # Otherwise, time for an emergency reschedule.
881 except Exception as e:
883 logger.error(f'{bundle}: Something unexpected just happened...')
885 msg = f"{bundle}: Failed to wrap up \"done\" bundle, re-waiting on active ssh."
887 return self.wait_for_process(p, bundle, depth + 1)
889 self.release_worker(bundle)
890 return self.emergency_retry_nasty_bundle(bundle)
892 def process_work_result(self, bundle: BundleDetails) -> Any:
893 with self.status.lock:
894 is_original = bundle.src_bundle is None
895 was_cancelled = bundle.was_cancelled
896 username = bundle.username
897 machine = bundle.machine
898 result_file = bundle.result_file
899 code_file = bundle.code_file
901 # Whether original or backup, if we finished first we must
902 # fetch the results if the computation happened on a
904 bundle.end_ts = time.time()
905 if not was_cancelled:
906 assert bundle.machine is not None
907 if bundle.hostname not in bundle.machine:
908 cmd = f'{SCP} {username}@{machine}:{result_file} {result_file} 2>/dev/null'
910 f"{bundle}: Fetching results back from {username}@{machine} via {cmd}"
913 # If either of these throw they are handled in
919 except Exception as e:
927 f'{SSH} {username}@{machine}'
928 f' "/bin/rm -f {code_file} {result_file}"'
931 f'Fetching results back took {time.time() - bundle.end_ts:.1f}s.'
933 dur = bundle.end_ts - bundle.start_ts
934 self.histogram.add_item(dur)
936 # Only the original worker should unpickle the file contents
937 # though since it's the only one whose result matters. The
938 # original is also the only job that may delete result_file
939 # from disk. Note that the original may have been cancelled
940 # if one of the backups finished first; it still must read the
943 logger.debug(f"{bundle}: Unpickling {result_file}.")
945 with open(result_file, 'rb') as rb:
946 serialized = rb.read()
947 result = cloudpickle.loads(serialized)
948 except Exception as e:
950 msg = f'Failed to load {result_file}... this is bad news.'
952 self.release_worker(bundle)
954 # Re-raise the exception; the code in wait_for_process may
955 # decide to emergency_retry_nasty_bundle here.
957 logger.debug(f'Removing local (master) {code_file} and {result_file}.')
958 os.remove(f'{result_file}')
959 os.remove(f'{code_file}')
961 # Notify any backups that the original is done so they
962 # should stop ASAP. Do this whether or not we
963 # finished first since there could be more than one
965 if bundle.backup_bundles is not None:
966 for backup in bundle.backup_bundles:
968 f'{bundle}: Notifying backup {backup.uuid} that it\'s cancelled'
970 backup.is_cancelled.set()
972 # This is a backup job and, by now, we have already fetched
973 # the bundle results.
975 # Backup results don't matter, they just need to leave the
976 # result file in the right place for their originals to
977 # read/unpickle later.
980 # Tell the original to stop if we finished first.
981 if not was_cancelled:
983 f'{bundle}: Notifying original {bundle.src_bundle.uuid} we beat them to it.'
985 bundle.src_bundle.is_cancelled.set()
986 self.release_worker(bundle, was_cancelled=was_cancelled)
989 def create_original_bundle(self, pickle, fname: str):
990 from string_utils import generate_uuid
992 uuid = generate_uuid(omit_dashes=True)
993 code_file = f'/tmp/{uuid}.code.bin'
994 result_file = f'/tmp/{uuid}.result.bin'
996 logger.debug(f'Writing pickled code to {code_file}')
997 with open(f'{code_file}', 'wb') as wb:
1000 bundle = BundleDetails(
1001 pickled_code=pickle,
1007 hostname=platform.node(),
1008 code_file=code_file,
1009 result_file=result_file,
1011 start_ts=time.time(),
1013 slower_than_local_p95=False,
1014 slower_than_global_p95=False,
1016 is_cancelled=threading.Event(),
1017 was_cancelled=False,
1021 self.status.record_bundle_details(bundle)
1022 logger.debug(f'{bundle}: Created an original bundle')
1025 def create_backup_bundle(self, src_bundle: BundleDetails):
1026 assert src_bundle.backup_bundles is not None
1027 n = len(src_bundle.backup_bundles)
1028 uuid = src_bundle.uuid + f'_backup#{n}'
1030 backup_bundle = BundleDetails(
1031 pickled_code=src_bundle.pickled_code,
1033 fname=src_bundle.fname,
1037 hostname=src_bundle.hostname,
1038 code_file=src_bundle.code_file,
1039 result_file=src_bundle.result_file,
1041 start_ts=time.time(),
1043 slower_than_local_p95=False,
1044 slower_than_global_p95=False,
1045 src_bundle=src_bundle,
1046 is_cancelled=threading.Event(),
1047 was_cancelled=False,
1048 backup_bundles=None, # backup backups not allowed
1051 src_bundle.backup_bundles.append(backup_bundle)
1052 self.status.record_bundle_details_already_locked(backup_bundle)
1053 logger.debug(f'{backup_bundle}: Created a backup bundle')
1054 return backup_bundle
1056 def schedule_backup_for_bundle(self, src_bundle: BundleDetails):
1057 assert self.status.lock.locked()
1058 assert src_bundle is not None
1059 backup_bundle = self.create_backup_bundle(src_bundle)
1061 f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...'
1063 self._helper_executor.submit(self.launch, backup_bundle)
1065 # Results from backups don't matter; if they finish first
1066 # they will move the result_file to this machine and let
1067 # the original pick them up and unpickle them.
1069 def emergency_retry_nasty_bundle(self, bundle: BundleDetails) -> fut.Future:
1070 is_original = bundle.src_bundle is None
1071 bundle.worker = None
1072 avoid_last_machine = bundle.machine
1073 bundle.machine = None
1074 bundle.username = None
1075 bundle.failure_count += 1
1081 if bundle.failure_count > retry_limit:
1083 f'{bundle}: Tried this bundle too many times already ({retry_limit}x); giving up.'
1086 raise RemoteExecutorException(
1087 f'{bundle}: This bundle can\'t be completed despite several backups and retries'
1091 f'{bundle}: At least it\'s only a backup; better luck with the others.'
1095 msg = f'>>> Emergency rescheduling {bundle} because of unexected errors (wtf?!) <<<'
1098 return self.launch(bundle, avoid_last_machine)
1101 def submit(self, function: Callable, *args, **kwargs) -> fut.Future:
1102 if self.already_shutdown:
1103 raise Exception('Submitted work after shutdown.')
1104 pickle = make_cloud_pickle(function, *args, **kwargs)
1105 bundle = self.create_original_bundle(pickle, function.__name__)
1106 self.total_bundles_submitted += 1
1107 return self._helper_executor.submit(self.launch, bundle)
1110 def shutdown(self, wait=True) -> None:
1111 if not self.already_shutdown:
1112 logging.debug(f'Shutting down RemoteExecutor {self.title}')
1113 self.heartbeat_stop_event.set()
1114 self.heartbeat_thread.join()
1115 self._helper_executor.shutdown(wait)
1116 print(self.histogram.__repr__(label_formatter='%d'))
1117 self.already_shutdown = True
1121 class DefaultExecutors(object):
1123 self.thread_executor: Optional[ThreadExecutor] = None
1124 self.process_executor: Optional[ProcessExecutor] = None
1125 self.remote_executor: Optional[RemoteExecutor] = None
1127 def ping(self, host) -> bool:
1128 logger.debug(f'RUN> ping -c 1 {host}')
1130 x = cmd_with_timeout(
1131 f'ping -c 1 {host} >/dev/null 2>/dev/null', timeout_seconds=1.0
1137 def thread_pool(self) -> ThreadExecutor:
1138 if self.thread_executor is None:
1139 self.thread_executor = ThreadExecutor()
1140 return self.thread_executor
1142 def process_pool(self) -> ProcessExecutor:
1143 if self.process_executor is None:
1144 self.process_executor = ProcessExecutor()
1145 return self.process_executor
1147 def remote_pool(self) -> RemoteExecutor:
1148 if self.remote_executor is None:
1149 logger.info('Looking for some helper machines...')
1150 pool: List[RemoteWorkerRecord] = []
1151 if self.ping('cheetah.house'):
1152 logger.info('Found cheetah.house')
1156 machine='cheetah.house',
1161 if self.ping('meerkat.cabin'):
1162 logger.info('Found meerkat.cabin')
1166 machine='meerkat.cabin',
1171 if self.ping('wannabe.house'):
1172 logger.info('Found wannabe.house')
1176 machine='wannabe.house',
1181 if self.ping('puma.cabin'):
1182 logger.info('Found puma.cabin')
1186 machine='puma.cabin',
1191 if self.ping('backup.house'):
1192 logger.info('Found backup.house')
1196 machine='backup.house',
1202 # The controller machine has a lot to do; go easy on it.
1204 if record.machine == platform.node() and record.count > 1:
1205 logger.info(f'Reducing workload for {record.machine}.')
1208 policy = WeightedRandomRemoteWorkerSelectionPolicy()
1209 policy.register_worker_pool(pool)
1210 self.remote_executor = RemoteExecutor(pool, policy)
1211 return self.remote_executor
1213 def shutdown(self) -> None:
1214 if self.thread_executor is not None:
1215 self.thread_executor.shutdown()
1216 self.thread_executor = None
1217 if self.process_executor is not None:
1218 self.process_executor.shutdown()
1219 self.process_executor = None
1220 if self.remote_executor is not None:
1221 self.remote_executor.shutdown()
1222 self.remote_executor = None