from ansi import bg, fg, underline, reset
import argparse_utils
import config
-from exec_utils import run_silently, cmd_in_background, cmd_with_timeout
from decorator_utils import singleton
+from exec_utils import run_silently, cmd_in_background, cmd_with_timeout
import histogram as hist
logger = logging.getLogger(__name__)
pid: int
start_ts: float
end_ts: float
- too_slow: bool
- super_slow: bool
+ slower_than_local_p95: bool
+ slower_than_global_p95: bool
src_bundle: BundleDetails
is_cancelled: threading.Event
was_cancelled: bool
if qworker is not None:
if sec > qworker[1]:
ret += f'{bg("red")}>💻p95{reset()} '
- elif sec > qworker[0]:
- ret += f'{fg("red")}>💻p50{reset()} '
- if qall is not None:
- if sec > qall[1] * 1.5:
- ret += f'{bg("red")}!!!{reset()}'
if details is not None:
- logger.debug(f'Flagging {details} for another backup')
- details.super_slow = True
- elif sec > qall[1]:
+ details.slower_than_local_p95 = True
+ else:
+ if details is not None:
+ details.slower_than_local_p95 = False
+
+ if qall is not None:
+ if sec > qall[1]:
ret += f'{bg("red")}>∀p95{reset()} '
if details is not None:
- logger.debug(f'Flagging {details} for a backup')
- details.too_slow = True
- elif sec > qall[0]:
- ret += f'{fg("red")}>∀p50{reset()}'
+ details.slower_than_global_p95 = True
+ else:
+ details.slower_than_global_p95 = False
ret += '\n'
return ret
class RemoteWorkerSelectionPolicy(ABC):
def register_worker_pool(self, workers):
- random.seed()
self.workers = workers
@abstractmethod
class WeightedRandomRemoteWorkerSelectionPolicy(RemoteWorkerSelectionPolicy):
+ @overrides
def is_worker_available(self) -> bool:
for worker in self.workers:
if worker.count > 0:
return True
return False
+ @overrides
def acquire_worker(
self,
machine_to_avoid = None
def __init__(self) -> None:
self.index = 0
+ @overrides
def is_worker_available(self) -> bool:
for worker in self.workers:
if worker.count > 0:
return True
return False
+ @overrides
def acquire_worker(
self,
machine_to_avoid: str = None
)
self.status = RemoteExecutorStatus(self.worker_count)
self.total_bundles_submitted = 0
+ self.backup_lock = threading.Lock()
+ self.last_backup = None
def is_worker_available(self) -> bool:
return self.policy.is_worker_available()
# Look for bundles to reschedule.
num_done = len(self.status.finished_bundle_timings)
- if num_done > 7 or (num_done > 5 and self.is_worker_available()):
- for worker, bundle_uuids in self.status.in_flight_bundles_by_worker.items():
- for uuid in bundle_uuids:
- bundle = self.status.bundle_details_by_uuid.get(uuid, None)
- if (
- bundle is not None and
- bundle.too_slow and
- bundle.src_bundle is None and
- config.config['executors_schedule_remote_backups']
- ):
- self.consider_backup_for_bundle(bundle)
-
- def consider_backup_for_bundle(self, bundle: BundleDetails) -> None:
- assert self.status.lock.locked()
- if (
- bundle.too_slow
- and len(bundle.backup_bundles) == 0 # one backup per
- ):
- msg = f"*** Rescheduling {bundle} (first backup) ***"
- logger.debug(msg)
- self.schedule_backup_for_bundle(bundle)
- return
- elif (
- bundle.super_slow
- and len(bundle.backup_bundles) < 2 # two backups in dire situations
- and self.status.total_idle() > 4
- ):
- msg = f"*** Rescheduling {bundle} (second backup) ***"
- logger.debug(msg)
- self.schedule_backup_for_bundle(bundle)
- return
+ num_idle_workers = self.worker_count - self.task_count
+ now = time.time()
+ if (
+ config.config['executors_schedule_remote_backups']
+ and num_done > 2
+ and num_idle_workers > 1
+ and (self.last_backup is None or (now - self.last_backup > 1.0))
+ and self.backup_lock.acquire(blocking=False)
+ ):
+ try:
+ assert self.backup_lock.locked()
+
+ bundle_to_backup = None
+ best_score = None
+ for worker, bundle_uuids in self.status.in_flight_bundles_by_worker.items():
+ # Prefer to schedule backups of bundles on slower machines.
+ base_score = 0
+ for record in self.workers:
+ if worker.machine == record.machine:
+ base_score = float(record.weight)
+ base_score = 1.0 / base_score
+ base_score *= 200.0
+ base_score = int(base_score)
+ break
+
+ for uuid in bundle_uuids:
+ bundle = self.status.bundle_details_by_uuid.get(uuid, None)
+ if (
+ bundle is not None
+ and bundle.src_bundle is None
+ and bundle.backup_bundles is not None
+ ):
+ score = base_score
+
+ # Schedule backups of bundles running longer; especially those
+ # that are unexpectedly slow.
+ start_ts = self.status.start_per_bundle[uuid]
+ if start_ts is not None:
+ runtime = now - start_ts
+ score += runtime
+ logger.debug(f'score[{bundle}] => {score} # latency boost')
+
+ if bundle.slower_than_local_p95:
+ score += runtime / 2
+ logger.debug(f'score[{bundle}] => {score} # >worker p95')
+
+ if bundle.slower_than_global_p95:
+ score += runtime / 2
+ logger.debug(f'score[{bundle}] => {score} # >global p95')
+
+ # Prefer backups of bundles that don't have backups already.
+ backup_count = len(bundle.backup_bundles)
+ if backup_count == 0:
+ score *= 2
+ elif backup_count == 1:
+ score /= 2
+ elif backup_count == 2:
+ score /= 8
+ else:
+ score = 0
+ logger.debug(f'score[{bundle}] => {score} # {backup_count} dup backup factor')
+
+ if (
+ score != 0
+ and (best_score is None or score > best_score)
+ ):
+ bundle_to_backup = bundle
+ assert bundle is not None
+ assert bundle.backup_bundles is not None
+ assert bundle.src_bundle is None
+ best_score = score
+
+ if bundle_to_backup is not None:
+ self.last_backup = now
+ logger.info(f'=====> SCHEDULING BACKUP {bundle_to_backup} (score={best_score:.1f}) <=====')
+ self.schedule_backup_for_bundle(bundle_to_backup)
+ finally:
+ self.backup_lock.release()
def check_if_cancelled(self, bundle: BundleDetails) -> bool:
with self.status.lock:
xfer_latency = time.time() - start_ts
logger.info(f"{bundle}: Copying done to {worker} in {xfer_latency:.1f}s.")
except Exception as e:
- logger.exception(e)
- logger.error(
- f'{bundle}: failed to send instructions to worker machine?!?'
- )
assert bundle.worker is not None
self.status.record_release_worker(
bundle.worker,
if is_original:
# Weird. We tried to copy the code to the worker and it failed...
# And we're the original bundle. We have to retry.
+ logger.exception(e)
+ logger.error(
+ f'{bundle}: Failed to send instructions to the worker machine?! ' +
+ 'This is not expected; we\'re the original bundle so this shouldn\'t ' +
+ 'be a race condition. Attempting an emergency retry...'
+ )
return self.emergency_retry_nasty_bundle(bundle)
else:
# This is actually expected; we're a backup.
# There's a race condition where someone else
# already finished the work and removed the source
# code file before we could copy it. No biggie.
+ logger.warning(
+ f'{bundle}: Failed to send instructions to the worker machine... ' +
+ 'We\'re a backup and this may be caused by the original (or some ' +
+ 'other backup) already finishing this work. Ignoring this.'
+ )
return None
# Kick off the work. Note that if this fails we let
# wait_for_process deal with it.
self.status.record_processing_began(uuid)
cmd = (f'{SSH} {bundle.username}@{bundle.machine} '
- f'"source py39-venv/bin/activate &&'
+ f'"source py38-venv/bin/activate &&'
f' /home/scott/lib/python_modules/remote_worker.py'
f' --code_file {bundle.code_file} --result_file {bundle.result_file}"')
logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...')
if is_original:
logger.debug(f"{bundle}: Unpickling {result_file}.")
try:
- with open(f'{result_file}', 'rb') as rb:
+ with open(result_file, 'rb') as rb:
serialized = rb.read()
result = cloudpickle.loads(serialized)
except Exception as e:
pid = 0,
start_ts = time.time(),
end_ts = 0.0,
- too_slow = False,
- super_slow = False,
+ slower_than_local_p95 = False,
+ slower_than_global_p95 = False,
src_bundle = None,
is_cancelled = threading.Event(),
was_cancelled = False,
pid = 0,
start_ts = time.time(),
end_ts = 0.0,
- too_slow = False,
- super_slow = False,
+ slower_than_local_p95 = False,
+ slower_than_global_p95 = False,
src_bundle = src_bundle,
is_cancelled = threading.Event(),
was_cancelled = False,
def schedule_backup_for_bundle(self,
src_bundle: BundleDetails):
assert self.status.lock.locked()
+ assert src_bundle is not None
backup_bundle = self.create_backup_bundle(src_bundle)
logger.debug(
f'{backup_bundle.uuid}/{backup_bundle.fname}: Scheduling backup for execution...'
RemoteWorkerRecord(
username = 'scott',
machine = 'cheetah.house',
- weight = 14,
- count = 4,
- ),
- )
- if self.ping('video.house'):
- logger.info('Found video.house')
- pool.append(
- RemoteWorkerRecord(
- username = 'scott',
- machine = 'video.house',
- weight = 1,
- count = 4,
- ),
- )
- if self.ping('wannabe.house'):
- logger.info('Found wannabe.house')
- pool.append(
- RemoteWorkerRecord(
- username = 'scott',
- machine = 'wannabe.house',
- weight = 2,
- count = 4,
+ weight = 25,
+ count = 6,
),
)
if self.ping('meerkat.cabin'):
count = 2,
),
)
- if self.ping('backup.house'):
- logger.info('Found backup.house')
+ # if self.ping('kiosk.house'):
+ # logger.info('Found kiosk.house')
+ # pool.append(
+ # RemoteWorkerRecord(
+ # username = 'pi',
+ # machine = 'kiosk.house',
+ # weight = 1,
+ # count = 2,
+ # ),
+ # )
+ if self.ping('hero.house'):
+ logger.info('Found hero.house')
pool.append(
RemoteWorkerRecord(
username = 'scott',
- machine = 'backup.house',
- weight = 1,
- count = 4,
+ machine = 'hero.house',
+ weight = 30,
+ count = 10,
),
)
- if self.ping('kiosk.house'):
- logger.info('Found kiosk.house')
+ if self.ping('puma.cabin'):
+ logger.info('Found puma.cabin')
pool.append(
RemoteWorkerRecord(
- username = 'pi',
- machine = 'kiosk.house',
- weight = 1,
- count = 2,
+ username = 'scott',
+ machine = 'puma.cabin',
+ weight = 25,
+ count = 6,
),
)
- if self.ping('puma.cabin'):
- logger.info('Found puma.cabin')
+ if self.ping('backup.house'):
+ logger.info('Found backup.house')
pool.append(
RemoteWorkerRecord(
username = 'scott',
- machine = 'puma.cabin',
- weight = 12,
- count = 4,
+ machine = 'backup.house',
+ weight = 3,
+ count = 2,
),
)
projects / python_utils.git / blobdiff