This article mainly introduces the method of using Beanstalkd for asynchronous task processing in Python. Now I share it with you and give it as a reference. Let’s take a look together
Use Beanstalkd as the message queue service, and then combine it with Python’s decorator syntax to implement a simple asynchronous task processing tool.
Final effect
Define task:
from xxxxx.job_queue import JobQueue queue = JobQueue() @queue.task('task_tube_one') def task_one(arg1, arg2, arg3): # do task
Submit task:
task_one.put(arg1="a", arg2="b", arg3="c")
Then These tasks can be performed by the background work thread.
Implementation process
1. Understand Beanstalk Server
Beanstalk is a simple, fast work queue. https://github.com /kr/beanstalkd
Beanstalk is a message queue service implemented in C language. It provides a common interface and was originally designed to reduce page latency in large-scale web applications by running time-consuming tasks asynchronously. There are different Beanstalkd Client implementations for different languages. There are beanstalkc and so on in Python. I use beanstalkc as a tool to communicate with beanstalkd server.
2. Implementation principle of asynchronous task execution
beanstalkd can only schedule task strings. In order for the program to support submitting functions and parameters, the function is then executed by the woker and the parameters are carried. A middle layer is needed to register functions with passed parameters.
The implementation mainly includes 3 parts:
Subscriber: Responsible for registering the function on a tube of beanstalk. The implementation is very simple, registering the corresponding relationship between the function name and the function itself. (This means that the same function name cannot exist in the same group (tube)). Data is stored in class variables.
class Subscriber(object): FUN_MAP = defaultdict(dict) def __init__(self, func, tube): logger.info('register func:{} to tube:{}.'.format(func.__name__, tube)) Subscriber.FUN_MAP[tube][func.__name__] = func
JobQueue: Conveniently converts an ordinary function into a decorator with Putter capability
class JobQueue(object): @classmethod def task(cls, tube): def wrapper(func): Subscriber(func, tube) return Putter(func, tube) return wrapper
Putter: Combine the function name, function parameters, and specified grouping into an object, then serialize json into a string, and finally push it to the beanstalkd queue through beanstalkc.
class Putter(object): def __init__(self, func, tube): self.func = func self.tube = tube # 直接调用返回 def __call__(self, *args, **kwargs): return self.func(*args, **kwargs) # 推给离线队列 def put(self, **kwargs): args = { 'func_name': self.func.__name__, 'tube': self.tube, 'kwargs': kwargs } logger.info('put job:{} to queue'.format(args)) beanstalk = beanstalkc.Connection(host=BEANSTALK_CONFIG['host'], port=BEANSTALK_CONFIG['port']) try: beanstalk.use(self.tube) job_id = beanstalk.put(json.dumps(args)) return job_id finally: beanstalk.close()
Worker: Take the string from the beanstalkd queue, and then deserialize it into an object through json.loads to obtain the function name, parameters and tube . Finally, the function code corresponding to the function name is obtained from the Subscriber, and then the parameters are passed to execute the function.
class Worker(object): worker_id = 0 def __init__(self, tubes): self.beanstalk = beanstalkc.Connection(host=BEANSTALK_CONFIG['host'], port=BEANSTALK_CONFIG['port']) self.tubes = tubes self.reserve_timeout = 20 self.timeout_limit = 1000 self.kick_period = 600 self.signal_shutdown = False self.release_delay = 0 self.age = 0 self.signal_shutdown = False signal.signal(signal.SIGTERM, lambda signum, frame: self.graceful_shutdown()) Worker.worker_id += 1 import_module_by_str('pear.web.controllers.controller_crawler') def subscribe(self): if isinstance(self.tubes, list): for tube in self.tubes: if tube not in Subscriber.FUN_MAP.keys(): logger.error('tube:{} not register!'.format(tube)) continue self.beanstalk.watch(tube) else: if self.tubes not in Subscriber.FUN_MAP.keys(): logger.error('tube:{} not register!'.format(self.tubes)) return self.beanstalk.watch(self.tubes) def run(self): self.subscribe() while True: if self.signal_shutdown: break if self.signal_shutdown: logger.info("graceful shutdown") break job = self.beanstalk.reserve(timeout=self.reserve_timeout) # 阻塞获取任务,最长等待 timeout if not job: continue try: self.on_job(job) self.delete_job(job) except beanstalkc.CommandFailed as e: logger.warning(e, exc_info=1) except Exception as e: logger.error(e) kicks = job.stats()['kicks'] if kicks < 3: self.bury_job(job) else: message = json.loads(job.body) logger.error("Kicks reach max. Delete the job", extra={'body': message}) self.delete_job(job) @classmethod def on_job(cls, job): start = time.time() msg = json.loads(job.body) logger.info(msg) tube = msg.get('tube') func_name = msg.get('func_name') try: func = Subscriber.FUN_MAP[tube][func_name] kwargs = msg.get('kwargs') func(**kwargs) logger.info(u'{}-{}'.format(func, kwargs)) except Exception as e: logger.error(e.message, exc_info=True) cost = time.time() - start logger.info('{} cost {}s'.format(func_name, cost)) @classmethod def delete_job(cls, job): try: job.delete() except beanstalkc.CommandFailed as e: logger.warning(e, exc_info=1) @classmethod def bury_job(cls, job): try: job.bury() except beanstalkc.CommandFailed as e: logger.warning(e, exc_info=1) def graceful_shutdown(self): self.signal_shutdown = True
When writing the above code, I found a problem:
Register the function name and function through Subscriber The corresponding relationship is that it runs in a Python interpreter, that is, in one process, and the Worker runs asynchronously in another process. How can the Worker get the same Subscriber as Putter? Finally, I found that this problem can be solved through Python's decorator mechanism.
This sentence solves the Subscriber problem
import_module_by_str('pear.web.controllers.controller_crawler')
# import_module_by_str 的实现 def import_module_by_str(module_name): if isinstance(module_name, unicode): module_name = str(module_name) __import__(module_name)
When executing import_module_by_str, __import__ will be called to dynamically load classes and functions. After loading the module containing the function using JobQueue into memory. When running Woker, the Python interpreter will first execute the @-decorated decorator code and load the corresponding relationship in Subscriber into memory.
For actual use, please see https://github.com/jiyangg/Pear/blob/master/pear/jobs/job_queue.py
Related recommendations:
php-beanstalkd message queue class instance detailed explanation
The above is the detailed content of How to use Beanstalkd in Python for asynchronous task processing. For more information, please follow other related articles on the PHP Chinese website!