# This file is a part of Dramatiq.
#
# Copyright (C) 2017,2018,2019,2020 CLEARTYPE SRL <[email protected]>
#
# Dramatiq is free software; you can redistribute it and/or modify it
# under the terms of the GNU Lesser General Public License as published by
# the Free Software Foundation, either version 3 of the License, or (at
# your option) any later version.
#
# Dramatiq is distributed in the hope that it will be useful, but WITHOUT
# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
# FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
# License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with this program. If not, see <http://www.gnu.org/licenses/>.
import os
import time
from collections import defaultdict
from functools import lru_cache
from itertools import chain
from queue import Empty, PriorityQueue
from threading import Event, Thread
from .common import current_millis, iter_queue, join_all, q_name
from .errors import ActorNotFound, ConnectionError, RateLimitExceeded, Retry
from .logging import get_logger
from .middleware import Middleware, SkipMessage
from .results.middleware import Results
#: The number of milliseconds to wait before restarting consumers
#: after a connection error.
CONSUMER_RESTART_DELAY = int(os.getenv("dramatiq_restart_delay", 3000))
CONSUMER_RESTART_DELAY_SECS = CONSUMER_RESTART_DELAY / 1000
#: The number of seconds to wait before retrying post_process_message
#: calls after a connection error.
POST_PROCESS_MESSAGE_RETRY_DELAY_SECS = 5
#: The number of messages to prefetch from the queue for each worker.
#: In-progress messages are included in the count. When set to "1",
#: a new message is only fetched once a previous one is done. If no
#: provided, two times the number of configured worker threads are
#: prefetched.
QUEUE_PREFETCH = int(os.getenv("dramatiq_queue_prefetch", 0))
#: The number of messages to prefetch from the delay queue for each worker.
#: When set to "1", a new message is only fetched once a previous one is done. If no
#: provided, a thousand times the number of configured worker threads are
#: prefetched.
DELAY_QUEUE_PREFETCH = int(os.getenv("dramatiq_delay_queue_prefetch", 0))
[docs]
class Worker:
"""Workers consume messages off of all declared queues and
distribute those messages to individual worker threads for
processing. Workers don't block the current thread so it's
up to the caller to keep it alive.
Don't run more than one Worker per process.
Parameters:
broker(Broker)
queues(Set[str]): An optional subset of queues to listen on. By
default, if this is not provided, the worker will listen on
all declared queues.
worker_timeout(int): The number of milliseconds workers should
wake up after if the queue is idle.
worker_threads(int): The number of worker threads to spawn.
"""
def __init__(self, broker, *, queues=None, worker_timeout=1000, worker_threads=8):
self.logger = get_logger(__name__, type(self))
self.broker = broker
self.consumers = {}
self.consumer_whitelist = queues and set(queues)
# Load a small factor more messages than there are workers to
# avoid waiting on network IO as much as possible. The factor
# must be small so we don't starve other workers out.
self.queue_prefetch = QUEUE_PREFETCH or min(worker_threads * 2, 65535)
# Load a large factor more delay messages than there are
# workers as those messages could have far-future etas.
self.delay_prefetch = DELAY_QUEUE_PREFETCH or min(worker_threads * 1000, 65535)
self.workers = []
self.work_queue = PriorityQueue()
self.worker_timeout = worker_timeout
self.worker_threads = worker_threads
[docs]
def start(self):
"""Initialize the worker boot sequence and start up all the
worker threads.
"""
self.broker.emit_before("worker_boot", self)
worker_middleware = _WorkerMiddleware(self)
self.broker.add_middleware(worker_middleware)
for _ in range(self.worker_threads):
self._add_worker()
self.broker.emit_after("worker_boot", self)
[docs]
def pause(self):
"""Pauses all the worker threads.
"""
for child in chain(self.consumers.values(), self.workers):
child.pause()
for child in chain(self.consumers.values(), self.workers):
child.paused_event.wait()
[docs]
def resume(self):
"""Resumes all the worker threads.
"""
for child in chain(self.consumers.values(), self.workers):
child.resume()
[docs]
def stop(self, timeout=600000):
"""Gracefully stop the Worker and all of its consumers and
workers.
Parameters:
timeout(int): The number of milliseconds to wait for
everything to shut down.
"""
self.broker.emit_before("worker_shutdown", self)
self.logger.info("Shutting down...")
# Stop workers before consumers. The consumers are kept alive
# during this process so that heartbeats keep being sent to
# the broker while workers finish their current tasks.
self.logger.debug("Stopping workers...")
for thread in self.workers:
thread.stop()
join_all(self.workers, timeout)
self.logger.debug("Workers stopped.")
self.logger.debug("Stopping consumers...")
for thread in self.consumers.values():
thread.stop()
join_all(self.consumers.values(), timeout)
self.logger.debug("Consumers stopped.")
self.logger.debug("Requeueing in-memory messages...")
messages_by_queue = defaultdict(list)
for _, message in iter_queue(self.work_queue):
messages_by_queue[message.queue_name].append(message)
for queue_name, messages in messages_by_queue.items():
try:
self.consumers[queue_name].requeue_messages(messages)
except ConnectionError:
self.logger.warning("Failed to requeue messages on queue %r.", queue_name, exc_info=True)
self.logger.debug("Done requeueing in-progress messages.")
self.logger.debug("Closing consumers...")
for consumer in self.consumers.values():
consumer.close()
self.logger.debug("Consumers closed.")
self.broker.emit_after("worker_shutdown", self)
self.logger.info("Worker has been shut down.")
[docs]
def join(self):
"""Wait for this worker to complete its work in progress.
This method is useful when testing code.
"""
while True:
for consumer in self.consumers.values():
consumer.delay_queue.join()
self.work_queue.join()
# If nothing got put on the delay queues while we were
# joining on the work queue then it should be safe to exit.
# This could still miss stuff but the chances are slim.
for consumer in self.consumers.values():
if consumer.delay_queue.unfinished_tasks:
break
else:
if self.work_queue.unfinished_tasks:
continue
return
def _add_consumer(self, queue_name, *, delay=False):
if queue_name in self.consumers:
self.logger.debug("A consumer for queue %r is already running.", queue_name)
return
canonical_name = q_name(queue_name)
if self.consumer_whitelist and canonical_name not in self.consumer_whitelist:
self.logger.debug("Dropping consumer for queue %r: not whitelisted.", queue_name)
return
consumer = self.consumers[queue_name] = _ConsumerThread(
broker=self.broker,
queue_name=queue_name,
prefetch=self.delay_prefetch if delay else self.queue_prefetch,
work_queue=self.work_queue,
worker_timeout=self.worker_timeout,
)
consumer.start()
def _add_worker(self):
worker = _WorkerThread(
broker=self.broker,
consumers=self.consumers,
work_queue=self.work_queue,
worker_timeout=self.worker_timeout
)
worker.start()
self.workers.append(worker)
class _WorkerMiddleware(Middleware):
def __init__(self, worker):
self.logger = get_logger(__name__, type(self))
self.worker = worker
def after_declare_queue(self, broker, queue_name):
self.logger.debug("Adding consumer for queue %r.", queue_name)
self.worker._add_consumer(queue_name)
def after_declare_delay_queue(self, broker, queue_name):
self.logger.debug("Adding consumer for delay queue %r.", queue_name)
self.worker._add_consumer(queue_name, delay=True)
class _ConsumerThread(Thread):
def __init__(self, *, broker, queue_name, prefetch, work_queue, worker_timeout):
super().__init__(daemon=True)
self.logger = get_logger(__name__, "ConsumerThread(%s)" % queue_name)
self.running = False
self.paused = False
self.paused_event = Event()
self.consumer = None
self.broker = broker
self.prefetch = prefetch
self.queue_name = queue_name
self.work_queue = work_queue
self.worker_timeout = worker_timeout
self.delay_queue = PriorityQueue()
def run(self):
self.logger.debug("Running consumer thread...")
self.running = True
self.broker.emit_after("consumer_thread_boot", self)
while self.running:
if self.paused:
self.logger.debug("Consumer is paused. Sleeping for %.02fms...", self.worker_timeout)
self.paused_event.set()
time.sleep(self.worker_timeout / 1000)
continue
try:
self.consumer = self.broker.consume(
queue_name=self.queue_name,
prefetch=self.prefetch,
timeout=self.worker_timeout,
)
for message in self.consumer:
if message is not None:
self.handle_message(message)
elif self.paused:
break
self.handle_delayed_messages()
if not self.running:
break
except ConnectionError as e:
self.logger.critical("Consumer encountered a connection error: %s", e)
self.delay_queue = PriorityQueue()
except Exception:
self.logger.critical("Consumer encountered an unexpected error.", exc_info=True)
# Avoid leaving any open file descriptors around when
# an exception occurs.
self.close()
# While the consumer is running (i.e. hasn't been shut down),
# try to restart it once a second.
if self.running:
self.logger.info("Restarting consumer in %0.2f seconds.", CONSUMER_RESTART_DELAY_SECS)
self.close()
time.sleep(CONSUMER_RESTART_DELAY_SECS)
# If it's no longer running, then shut it down gracefully.
self.broker.emit_before("consumer_thread_shutdown", self)
self.logger.debug("Consumer thread stopped.")
def handle_delayed_messages(self):
"""Enqueue any delayed messages whose eta has passed.
"""
for eta, message in iter_queue(self.delay_queue):
if eta > current_millis():
self.delay_queue.put((eta, message))
self.delay_queue.task_done()
break
queue_name = q_name(message.queue_name)
new_message = message.copy(queue_name=queue_name)
del new_message.options["eta"]
self.broker.enqueue(new_message)
self.post_process_message(message)
self.delay_queue.task_done()
def handle_message(self, message):
"""Handle a message received off of the underlying consumer.
If the message has an eta, delay it. Otherwise, put it on the
work queue.
"""
try:
if "eta" in message.options:
self.logger.debug("Pushing message %r onto delay queue.", message.message_id)
self.broker.emit_before("delay_message", message)
self.delay_queue.put((message.options.get("eta", 0), message))
else:
actor = self.broker.get_actor(message.actor_name)
self.logger.debug("Pushing message %r onto work queue.", message.message_id)
self.work_queue.put((actor.priority, message))
except ActorNotFound:
self.logger.error(
"Received message for undefined actor %r. Moving it to the DLQ.",
message.actor_name, exc_info=True,
)
message.fail()
self.post_process_message(message)
def post_process_message(self, message):
"""Called by worker threads whenever they're done processing
individual messages, signaling that each message is ready to
be acked or rejected.
"""
while True:
try:
if message.failed:
self.logger.debug("Rejecting message %r.", message.message_id)
self.broker.emit_before("nack", message)
self.consumer.nack(message)
self.broker.emit_after("nack", message)
else:
self.logger.debug("Acknowledging message %r.", message.message_id)
self.broker.emit_before("ack", message)
self.consumer.ack(message)
self.broker.emit_after("ack", message)
return
# This applies to the Redis broker. The alternative to
# constantly retrying would be to give up here and let the
# message be re-processed after the worker is eventually
# stopped or restarted, but we'd be doing the same work
# twice in that case and the behaviour would surprise
# users who don't deploy frequently.
except ConnectionError as e:
self.logger.warning(
"Failed to post_process_message(%s) due to a connection error: %s\n"
"The operation will be retried in %s seconds until the connection recovers.\n"
"If you restart this worker before this operation succeeds, the message will be re-processed later.",
message, e, POST_PROCESS_MESSAGE_RETRY_DELAY_SECS
)
time.sleep(POST_PROCESS_MESSAGE_RETRY_DELAY_SECS)
continue
# Not much point retrying here so we bail. Most likely,
# the message will be re-run after the worker is stopped
# or restarted (because its ack lease will have expired).
except Exception: # pragma: no cover
self.logger.exception(
"Unhandled error during post_process_message(%s). You've found a bug in Dramatiq. Please report it!\n"
"Although your message has been processed, it will be processed again once this worker is restarted.",
message,
)
return
def requeue_messages(self, messages):
"""Called on worker shutdown and whenever there is a
connection error to move unacked messages back to their
respective queues asap.
"""
self.consumer.requeue(messages)
def pause(self):
"""Pause this consumer.
"""
self.paused = True
self.paused_event.clear()
def resume(self):
"""Resume this consumer.
"""
self.paused = False
self.paused_event.clear()
def stop(self):
"""Initiate the ConsumerThread shutdown sequence.
Code calling this method should then join on the thread and
wait for it to finish shutting down.
"""
self.logger.debug("Stopping consumer thread...")
self.running = False
def close(self):
"""Close this consumer thread and its underlying connection.
"""
try:
if self.consumer:
self.requeue_messages(m for _, m in iter_queue(self.delay_queue))
self.consumer.close()
except ConnectionError:
pass
class _WorkerThread(Thread):
"""WorkerThreads process incoming messages off of the work queue
on a loop. By themselves, they don't do any sort of network IO.
Parameters:
broker(Broker)
consumers(dict[str, _ConsumerThread])
work_queue(Queue)
worker_timeout(int)
"""
def __init__(self, *, broker, consumers, work_queue, worker_timeout):
super().__init__(daemon=True)
self.logger = get_logger(__name__, "WorkerThread")
self.running = False
self.paused = False
self.paused_event = Event()
self.broker = broker
self.consumers = consumers
self.work_queue = work_queue
self.timeout = worker_timeout / 1000
def run(self):
self.logger.debug("Running worker thread...")
self.running = True
self.broker.emit_after("worker_thread_boot", self)
while self.running:
if self.paused:
self.logger.debug("Worker is paused. Sleeping for %.02f...", self.timeout)
self.paused_event.set()
time.sleep(self.timeout)
continue
try:
_, message = self.work_queue.get(timeout=self.timeout)
self.process_message(message)
except Empty:
continue
self.broker.emit_before("worker_thread_shutdown", self)
self.logger.debug("Worker thread stopped.")
def process_message(self, message):
"""Process a message pulled off of the work queue then push it
back to its associated consumer for post processing. Stuff any SkipMessage
exception or BaseException into the message [proxy] so that it may be used
by the stub broker to provide a nicer testing experience. Also used by the
results middleware to pass exceptions into results.
Parameters:
message(MessageProxy)
"""
actor = None
try:
self.logger.debug("Received message %s with id %r.", message, message.message_id)
self.broker.emit_before("process_message", message)
res = None
if not message.failed:
actor = self.broker.get_actor(message.actor_name)
res = actor(*message.args, **message.kwargs)
if res is not None \
and message.options.get("pipe_target") is None \
and not has_results_middleware(self.broker):
self.logger.warning(
"Actor '%s' returned a value that is not None, and you haven't added the "
"Results middleware to the broker, so the value has been discarded. "
"Consider adding the Results middleware to your broker or piping the "
"result into another actor." % actor.actor_name
)
self.broker.emit_after("process_message", message, result=res)
except SkipMessage as e:
if message.failed:
message.stuff_exception(e)
self.logger.warning("Message %s was skipped.", message)
self.broker.emit_after("skip_message", message)
except BaseException as e:
message.stuff_exception(e)
throws = message.options.get("throws") or (actor and actor.options.get("throws"))
if isinstance(e, RateLimitExceeded):
self.logger.debug("Rate limit exceeded in message %s: %s.", message, e)
elif throws and isinstance(e, throws):
self.logger.info("Failed to process message %s with expected exception %s.", message, type(e).__name__)
elif not isinstance(e, Retry):
self.logger.error("Failed to process message %s with unhandled exception.", message, exc_info=True)
self.broker.emit_after("process_message", message, exception=e)
finally:
# NOTE: There is no race here as any message that was
# processed must have come off of a consumer. Therefore,
# there has to be a consumer for that message's queue so
# this is safe. Probably.
self.consumers[message.queue_name].post_process_message(message)
self.work_queue.task_done()
# See discussion #351. Keeping a reference to the
# exception can lead to memory bloat because it may be a
# while before a GC triggers.
message.clear_exception()
def pause(self):
"""Pause this worker.
"""
self.paused = True
self.paused_event.clear()
def resume(self):
"""Resume this worker.
"""
self.paused = False
self.paused_event.clear()
def stop(self):
"""Initiate the WorkerThread shutdown process.
Code calling this method should then join on the thread and
wait for it to finish shutting down.
"""
self.logger.debug("Stopping worker thread...")
self.running = False
@lru_cache(maxsize=128)
def has_results_middleware(broker):
return any(type(m) is Results for m in broker.middleware)
Franz — Desktop Client for Apache Kafka