import threading
import time
from typing import Any, Callable, Dict, List, Optional, Set
+import warnings
import cloudpickle # type: ignore
from overrides import overrides
from ansi import bg, fg, underline, reset
import argparse_utils
import config
-from exec_utils import run_silently, cmd_in_background
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
backup_bundles: Optional[List[BundleDetails]]
failure_count: int
+ def __repr__(self):
+ uuid = self.uuid
+ if uuid[-9:-2] == '_backup':
+ uuid = uuid[:-9]
+ suffix = f'{uuid[-6:]}_b{self.uuid[-1:]}'
+ else:
+ suffix = uuid[-6:]
+
+ colorz = [
+ fg('violet red'),
+ fg('red'),
+ fg('orange'),
+ fg('peach orange'),
+ fg('yellow'),
+ fg('marigold yellow'),
+ fg('green yellow'),
+ fg('tea green'),
+ fg('cornflower blue'),
+ fg('turquoise blue'),
+ fg('tropical blue'),
+ fg('lavender purple'),
+ fg('medium purple'),
+ ]
+ c = colorz[int(uuid[-2:], 16) % len(colorz)]
+ fname = self.fname if self.fname is not None else 'nofname'
+ machine = self.machine if self.machine is not None else 'nomachine'
+ return f'{c}{suffix}/{fname}/{machine}{reset()}'
+
class RemoteExecutorStatus:
def __init__(self, total_worker_count: int) -> None:
bundle_uuid,
None
)
- pid = str(details.pid) if details is not None else "TBD"
+ pid = str(details.pid) if (details and details.pid != 0) else "TBD"
if self.start_per_bundle[bundle_uuid] is not None:
sec = ts - self.start_per_bundle[bundle_uuid]
- ret += f' (pid={pid}): {bundle_uuid} for {sec:.1f}s so far '
+ ret += f' (pid={pid}): {details} for {sec:.1f}s so far '
else:
- ret += f' {bundle_uuid} setting up / copying data...'
+ ret += f' {details} setting up / copying data...'
sec = 0.0
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.uuid} 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.uuid} 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
worker.count -= 1
logger.debug(f'Selected worker {worker}')
return worker
- logger.warning("Couldn't find a worker; go fish.")
+ msg = 'Unexpectedly could not find a worker, retrying...'
+ logger.warning(msg)
return 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
if x >= len(self.workers):
x = 0
if x == self.index:
- logger.warning("Couldn't find a worker; go fish.")
+ msg = 'Unexpectedly could not find a worker, retrying...'
+ logger.warning(msg)
return 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.pid}/{bundle.uuid} ***"
- 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.pid}/{bundle.uuid} ***"
- 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:
bundle.worker = worker
machine = bundle.machine = worker.machine
username = bundle.username = worker.username
- fname = bundle.fname
+
self.status.record_acquire_worker(worker, uuid)
- logger.debug(f'{uuid}/{fname}: Running bundle on {worker}...')
+ logger.debug(f'{bundle}: Running bundle on {worker}...')
# Before we do any work, make sure the bundle is still viable.
if self.check_if_cancelled(bundle):
except Exception as e:
logger.exception(e)
logger.error(
- f'{uuid}/{fname}: bundle says it\'s cancelled upfront but no results?!'
+ f'{bundle}: bundle says it\'s cancelled upfront but no results?!'
)
assert bundle.worker is not None
self.status.record_release_worker(
try:
cmd = f'{RSYNC} {bundle.code_file} {username}@{machine}:{bundle.code_file}'
start_ts = time.time()
- logger.info(f"{uuid}/{fname}: Copying work to {worker} via {cmd}.")
+ logger.info(f"{bundle}: Copying work to {worker} via {cmd}.")
run_silently(cmd)
xfer_latency = time.time() - start_ts
- logger.info(f"{uuid}/{fname}: Copying done in {xfer_latency:.1f}s.")
+ logger.info(f"{bundle}: Copying done to {worker} in {xfer_latency:.1f}s.")
except Exception as e:
- logger.exception(e)
- logger.error(
- f'{uuid}/{fname}: 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.
+ msg = f'{bundle}: Failed to send instructions to the worker machine... '
+ msg += 'We\'re a backup and this may be caused by the original (or some '
+ msg += 'other backup) already finishing this work. Ignoring this.'
+ logger.warning(msg)
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'{uuid}/{fname}: Executing {cmd} in the background to kick off work...')
+ logger.debug(f'{bundle}: Executing {cmd} in the background to kick off work...')
p = cmd_in_background(cmd, silent=True)
bundle.pid = pid = p.pid
- logger.debug(f'{uuid}/{fname}: Local ssh process pid={pid}; remote worker is {machine}.')
+ logger.debug(f'{bundle}: Local ssh process pid={pid}; remote worker is {machine}.')
return self.wait_for_process(p, bundle, 0)
def wait_for_process(self, p: subprocess.Popen, bundle: BundleDetails, depth: int) -> Any:
- uuid = bundle.uuid
machine = bundle.machine
- fname = bundle.fname
pid = p.pid
if depth > 3:
logger.error(
self.heartbeat()
if self.check_if_cancelled(bundle):
logger.info(
- f'{uuid}/{fname}: another worker finished bundle, checking it out...'
+ f'{bundle}: another worker finished bundle, checking it out...'
)
break
else:
logger.info(
- f"{uuid}/{fname}: pid {pid} ({machine}) our ssh finished, checking it out..."
+ f"{bundle}: pid {pid} ({machine}) our ssh finished, checking it out..."
)
p = None
break
# Otherwise, time for an emergency reschedule.
except Exception as e:
logger.exception(e)
- logger.error(f'{uuid}/{fname}: Something unexpected just happened...')
+ logger.error(f'{bundle}: Something unexpected just happened...')
if p is not None:
- logger.warning(
- f"{uuid}/{fname}: Failed to wrap up \"done\" bundle, re-waiting on active ssh."
- )
+ msg = f"{bundle}: Failed to wrap up \"done\" bundle, re-waiting on active ssh."
+ logger.warning(msg)
return self.wait_for_process(p, bundle, depth + 1)
else:
self.status.record_release_worker(
machine = bundle.machine
result_file = bundle.result_file
code_file = bundle.code_file
- fname = bundle.fname
- uuid = bundle.uuid
# Whether original or backup, if we finished first we must
# fetch the results if the computation happened on a
if bundle.hostname not in bundle.machine:
cmd = f'{RSYNC} {username}@{machine}:{result_file} {result_file} 2>/dev/null'
logger.info(
- f"{uuid}/{fname}: Fetching results from {username}@{machine} via {cmd}"
+ f"{bundle}: Fetching results from {username}@{machine} via {cmd}"
)
# If either of these throw they are handled in
# if one of the backups finished first; it still must read the
# result from disk.
if is_original:
- logger.debug(f"{uuid}/{fname}: Unpickling {result_file}.")
+ 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:
if bundle.backup_bundles is not None:
for backup in bundle.backup_bundles:
logger.debug(
- f'{uuid}/{fname}: Notifying backup {backup.uuid} that it\'s cancelled'
+ f'{bundle}: Notifying backup {backup.uuid} that it\'s cancelled'
)
backup.is_cancelled.set()
# Tell the original to stop if we finished first.
if not was_cancelled:
logger.debug(
- f'{uuid}/{fname}: Notifying original {bundle.src_bundle.uuid} we beat them to it.'
+ f'{bundle}: Notifying original {bundle.src_bundle.uuid} we beat them to it.'
)
bundle.src_bundle.is_cancelled.set()
def create_original_bundle(self, pickle, fname: str):
from string_utils import generate_uuid
- uuid = generate_uuid(as_hex=True)
+ uuid = generate_uuid(omit_dashes=True)
code_file = f'/tmp/{uuid}.code.bin'
result_file = f'/tmp/{uuid}.result.bin'
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,
failure_count = 0,
)
self.status.record_bundle_details(bundle)
- logger.debug(f'{uuid}/{fname}: Created original bundle')
+ logger.debug(f'{bundle}: Created an original bundle')
return bundle
def create_backup_bundle(self, src_bundle: BundleDetails):
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,
)
src_bundle.backup_bundles.append(backup_bundle)
self.status.record_bundle_details_already_locked(backup_bundle)
- logger.debug(f'{uuid}/{src_bundle.fname}: Created backup bundle')
+ logger.debug(f'{backup_bundle}: Created a backup bundle')
return backup_bundle
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...'
# the original pick them up and unpickle them.
def emergency_retry_nasty_bundle(self, bundle: BundleDetails) -> fut.Future:
- uuid = bundle.uuid
is_original = bundle.src_bundle is None
bundle.worker = None
avoid_last_machine = bundle.machine
if bundle.failure_count > retry_limit:
logger.error(
- f'{uuid}: Tried this bundle too many times already ({retry_limit}x); giving up.'
+ f'{bundle}: Tried this bundle too many times already ({retry_limit}x); giving up.'
)
if is_original:
raise RemoteExecutorException(
- f'{uuid}: This bundle can\'t be completed despite several backups and retries'
+ f'{bundle}: This bundle can\'t be completed despite several backups and retries'
)
else:
- logger.error(f'{uuid}: At least it\'s only a backup; better luck with the others.')
+ logger.error(f'{bundle}: At least it\'s only a backup; better luck with the others.')
return None
else:
- logger.warning(
- f'>>> Emergency rescheduling {uuid} because of unexected errors (wtf?!) <<<'
- )
+ msg = f'>>> Emergency rescheduling {bundle} because of unexected errors (wtf?!) <<<'
+ logger.warning(msg)
+ warnings.warn(msg)
return self.launch(bundle, avoid_last_machine)
@overrides
def ping(self, host) -> bool:
logger.debug(f'RUN> ping -c 1 {host}')
- command = ['ping', '-c', '1', host]
- return subprocess.call(
- command,
- stdout=subprocess.DEVNULL,
- stderr=subprocess.DEVNULL,
- ) == 0
+ try:
+ x = cmd_with_timeout(
+ f'ping -c 1 {host} >/dev/null 2>/dev/null',
+ timeout_seconds=1.0
+ )
+ return x == 0
+ except Exception:
+ return False
def thread_pool(self) -> ThreadExecutor:
if self.thread_executor is None:
RemoteWorkerRecord(
username = 'scott',
machine = 'cheetah.house',
- weight = 12,
- 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