事件队列与消息代理
使用 Redis 队列、RabbitMQ 和 Kafka 实现智能体的解耦通信
事件队列与消息代理 是 CoddyKit 上的免费 AI Agents 课时。 这是第 2 节课,共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 AI Agents 学习路径的一部分,你的进度在网页和 CoddyKit 应用中同步。 AI Agents 课程共包含 4 节课。
为什么智能体要使用消息队列
消息队列将触发器(事件生产者)与智能体(事件消费者)解耦。生产者可以以任意速率触发事件,而智能体会按照自己的处理速度处理事件。这样可以防止过载,并支持重试。
使用 Redis 列表作为队列
Redis 列表可以作为简单队列使用:LPUSH 将元素添加到队首(入队),BRPOP 从队尾移除元素,并在队列为空时阻塞(出队)。这是一种可靠的先进先出队列。
import redis
r = redis.Redis(host='localhost', port=6379, decode_responses=True)
QUEUE_NAME = 'agent:tasks'
# Producer: add a task to the queue
def enqueue_task(task: dict):
import json
r.lpush(QUEUE_NAME, json.dumps(task))
print(f'Enqueued task: {task["id"]}')
# Consumer: blocking pop (waits up to 5 seconds for a message)
def dequeue_task(timeout: int = 5):
import json
result = r.brpop(QUEUE_NAME, timeout=timeout)
if result:
queue_name, raw_data = result
return json.loads(raw_data)
return None # Timed out - no tasks
# Producer side
enqueue_task({'id': 'task-1', 'type': 'email', 'email_id': 'email-abc'})
enqueue_task({'id': 'task-2', 'type': 'email', 'email_id': 'email-def'})
# Check queue length
print(f'Queue length: {r.llen(QUEUE_NAME)}')使用 Redis 的智能体工作循环
智能体工作器会持续轮询队列以获取新任务,处理每个任务,然后继续循环。如果在超时时间内没有任务到达,它会再次循环,而不会持续消耗资源。
import redis
import json
import time
import logging
logger = logging.getLogger('worker')
r = redis.Redis(host='localhost', port=6379, decode_responses=True)
def process_task(task: dict) -> bool:
task_type = task.get('type')
if task_type == 'email':
logger.info(f'Processing email: {task.get("email_id")}')
# Call email agent here
return True
elif task_type == 'file':
logger.info(f'Processing file: {task.get("filepath")}')
return True
else:
logger.warning(f'Unknown task type: {task_type}')
return False
def run_worker(queue_name: str = 'agent:tasks'):
print(f'Worker started, watching queue: {queue_name}')
while True:
try:
task = dequeue_task(timeout=5)
if task:
success = process_task(task)
if not success:
# Re-queue failed tasks for retry
r.lpush('agent:failed', json.dumps(task))
else:
logger.debug('No tasks, waiting...')
except KeyboardInterrupt:
print('Worker stopped')
break
except Exception as e:
logger.error(f'Worker error: {e}')
time.sleep(1) # Brief pause on unexpected error
print('Worker function defined')使用 Redis Pub/Sub 处理事件
Redis Pub/Sub 与列表不同:它会将消息广播给所有当前订阅者。对于需要让多个智能体响应同一事件的扇出通知,请使用 Pub/Sub。
import redis
import json
import threading
r = redis.Redis(host='localhost', port=6379, decode_responses=True)
# Publisher: send event to all subscribers
def publish_event(channel: str, event: dict):
r.publish(channel, json.dumps(event))
print(f'Published to {channel}: {event}')
# Subscriber: listen for events in a background thread
def subscribe_and_handle(channel: str, handler_fn):
pubsub = r.pubsub()
pubsub.subscribe(channel)
def listener():
for message in pubsub.listen():
if message['type'] == 'message':
event = json.loads(message['data'])
handler_fn(event)
thread = threading.Thread(target=listener, daemon=True)
thread.start()
return thread
def handle_agent_event(event):
print(f'Agent received event: {event}')
# Start subscriber
subscribe_and_handle('agent:events', handle_agent_event)
# Publish an event
publish_event('agent:events', {'type': 'user_action', 'action': 'login', 'user_id': 42})
import time
time.sleep(0.1) # Give subscriber time to receive使用 pika 的 RabbitMQ 基础
RabbitMQ 是功能完整的消息代理,支持路由、确认和死信队列。pika 库用于连接 Python 与 RabbitMQ。
import pika
import json
# Connect to RabbitMQ
connection = pika.BlockingConnection(
pika.ConnectionParameters(
host='localhost',
port=5672,
credentials=pika.PlainCredentials('guest', 'guest')
)
)
channel = connection.channel()
# Declare a durable queue (survives broker restart)
channel.queue_declare(
queue='agent_tasks',
durable=True # Queue survives RabbitMQ restart
)
# Publish a message
def publish_task(task: dict):
channel.basic_publish(
exchange='',
routing_key='agent_tasks',
body=json.dumps(task),
properties=pika.BasicProperties(
delivery_mode=2, # Make message persistent
content_type='application/json'
)
)
print(f'Published task: {task["id"]}')
publish_task({'id': 'task-1', 'type': 'process_email', 'email_id': 'email-xyz'})
connection.close()带确认机制的 RabbitMQ 消费者
消费者必须在处理消息后确认消息。如果智能体在确认之前崩溃,RabbitMQ 会将消息重新放回队列,供其他消费者处理。这样可以确保消息不会丢失。
import pika
import json
def create_consumer():
connection = pika.BlockingConnection(
pika.ConnectionParameters(host='localhost')
)
channel = connection.channel()
channel.queue_declare(queue='agent_tasks', durable=True)
# Process one message at a time (fair dispatch)
channel.basic_qos(prefetch_count=1)
def process_message(ch, method, properties, body):
task = json.loads(body)
print(f'Processing: {task["id"]}')
try:
# Do agent work here
print(f'Completed task: {task["id"]}')
# Acknowledge: message is removed from queue
ch.basic_ack(delivery_tag=method.delivery_tag)
except Exception as e:
print(f'Failed task {task["id"]}: {e}')
# Negative acknowledge + requeue=True: puts message back
ch.basic_nack(delivery_tag=method.delivery_tag, requeue=True)
channel.basic_consume(queue='agent_tasks', on_message_callback=process_message)
print('Consumer ready. Waiting for messages...')
channel.start_consuming()
print('RabbitMQ consumer defined')死信队列
处理失败次数过多的消息应进入死信队列(DLQ),以便人工检查,而不是无限重试。请配置 RabbitMQ,自动将失败消息路由到 DLQ。
import pika
connection = pika.BlockingConnection(pika.ConnectionParameters('localhost'))
channel = connection.channel()
# Create the dead letter queue first
channel.queue_declare(queue='agent_tasks_dlq', durable=True)
# Create main queue with dead-letter exchange config
channel.queue_declare(
queue='agent_tasks',
durable=True,
arguments={
'x-dead-letter-exchange': '',
'x-dead-letter-routing-key': 'agent_tasks_dlq',
'x-message-ttl': 3600000, # Messages expire after 1 hour
'x-max-delivery-count': 3 # Max 3 delivery attempts
}
)
# Consumer: nack without requeue sends to DLQ
def careful_consumer(ch, method, properties, body):
import json
task = json.loads(body)
try:
# Process task
ch.basic_ack(delivery_tag=method.delivery_tag)
except Exception:
# Do NOT requeue - send to DLQ
ch.basic_nack(delivery_tag=method.delivery_tag, requeue=False)
print('Dead letter queue configured')
connection.close()任务队列模式
任务队列模式将触发与执行解耦:轻量级生产者将任务描述推送到队列,工作器再拉取并执行这些任务。多个工作器可以并行处理任务。
import redis
import json
from datetime import datetime
r = redis.Redis(host='localhost', port=6379, decode_responses=True)
# Task descriptor - what needs to be done
def create_agent_task(task_type: str, params: dict, priority: int = 5) -> dict:
return {
'id': f'{task_type}-{int(datetime.utcnow().timestamp() * 1000)}',
'type': task_type,
'params': params,
'priority': priority,
'created_at': datetime.utcnow().isoformat(),
'retry_count': 0,
'max_retries': 3
}
# Enqueue with priority (multiple queues by priority)
def enqueue_with_priority(task: dict):
priority = task.get('priority', 5)
queue = f'agent:tasks:p{priority}'
r.lpush(queue, json.dumps(task))
print(f'Enqueued {task["id"]} to priority-{priority} queue')
# Worker dequeues from high-priority queue first
def dequeue_priority(timeout: int = 5):
for p in [1, 2, 3, 4, 5]: # Check priority 1 first
result = r.brpop(f'agent:tasks:p{p}', timeout=0.1)
if result:
return json.loads(result[1])
return None
enqueue_with_priority(create_agent_task('email_analysis', {'email_id': 'abc'}, priority=2))
enqueue_with_priority(create_agent_task('file_process', {'path': '/tmp/file.txt'}, priority=5))Celery:高级任务队列
Celery 是最受欢迎的 Python 任务队列。它支持将 Redis 和 RabbitMQ 作为代理,提供自动重试和任务路由,并可通过 Flower 进行监控。请将它用于生产环境中的智能体工作负载。
from celery import Celery
import os
# Create Celery app with Redis broker
app = Celery(
'agent_tasks',
broker=os.environ.get('REDIS_URL', 'redis://localhost:6379/0'),
backend=os.environ.get('REDIS_URL', 'redis://localhost:6379/0')
)
# Configure retry behavior
app.conf.update(
task_acks_late=True,
task_reject_on_worker_lost=True,
task_serializer='json',
result_expires=3600
)
@app.task(bind=True, max_retries=3, default_retry_delay=60)
def run_email_agent(self, email_id: str):
try:
# Agent logic here
print(f'Processing email: {email_id}')
return {'status': 'success', 'email_id': email_id}
except Exception as exc:
raise self.retry(exc=exc)
# Trigger a task (from any Python code)
# run_email_agent.delay('email-abc') # Fire and forget
# result = run_email_agent.apply_async(args=['email-abc'], countdown=60) # Delayed
print('Celery task defined')监控队列健康状况
监控队列深度,以便检测积压。如果队列深度不断增加,您需要增加工作器数量,或者说明智能体处理速度过慢。当深度超过阈值时,请发出警报。
import redis
r = redis.Redis(host='localhost', port=6379, decode_responses=True)
def check_queue_health(queue_name: str, max_depth: int = 100) -> dict:
depth = r.llen(queue_name)
oldest_raw = r.lindex(queue_name, -1) # Get last item (oldest)
oldest_age_seconds = None
if oldest_raw:
import json
from datetime import datetime
oldest = json.loads(oldest_raw)
if 'created_at' in oldest:
created = datetime.fromisoformat(oldest['created_at'])
oldest_age_seconds = (datetime.utcnow() - created).total_seconds()
health = {
'queue': queue_name,
'depth': depth,
'max_depth': max_depth,
'overloaded': depth > max_depth,
'oldest_item_age_seconds': oldest_age_seconds
}
if health['overloaded']:
print(f'ALERT: Queue {queue_name} depth={depth} exceeds max={max_depth}')
return health
print('Queue health monitor defined')
print('Usage: check_queue_health("agent:tasks")')选择 Redis 还是 RabbitMQ
在需要简单性、已经使用 Redis,或任务生命周期较短且可以接受崩溃时丢失少量消息的情况下,请使用 Redis。在需要保证消息送达、复杂路由、死信队列,或需要具有不同订阅的多个消费者时,请使用 RabbitMQ。
# Decision guide as code comments
# Use Redis lists when:
# - Simple FIFO queue is enough
# - Redis already in stack
# - Can tolerate rare message loss on Redis crash
# - Low operational overhead matters
# Use RabbitMQ when:
# - Message acknowledgment is critical (no data loss)
# - Need complex routing (topic exchanges, fanout)
# - Need dead-letter queues for failed messages
# - Multiple consumer types need different message subsets
# - Need message TTL and expiry
# Use Celery (on top of either) when:
# - Need task scheduling and delays
# - Need automatic retries with exponential backoff
# - Need task result storage
# - Need monitoring dashboard (Flower)
print('Redis: simple, fast, ok with rare loss')
print('RabbitMQ: guaranteed delivery, complex routing')
print('Celery: high-level abstraction over both')知识检查:消息队列
请测试您对智能体消息队列和消息代理的理解。
消息队列总结
消息队列将智能体触发与执行解耦:Redis 列表通过 LPUSH/BRPOP 提供简单的 FIFO 队列;Redis Pub/Sub 支持广播事件;RabbitMQ 通过确认机制和死信队列提供可靠的消息送达;Celery 增加高级调度、重试和监控功能。监控队列深度可以在积压演变为服务中断之前发出警报。
常见问题解答
「事件队列与消息代理」课时是免费的吗?
是的 — 「事件队列与消息代理」的完整文本可在网页上免费阅读。要进行交互式练习(内置代码编辑器和全天候 AI 导师)并解锁 AI Agents 课程的其余内容,请升级到 CoddyKit PRO。 AI Agents 课程共包含 4 节课。
「事件队列与消息代理」这节课中我会学到什么?
使用 Redis 队列、RabbitMQ 和 Kafka 实现智能体的解耦通信 你通过在浏览器中直接运行的动手代码来练习 AI Agents,全天候 AI 导师会在你学习这节课的过程中回答你的问题。
学习 AI Agents 需要有经验吗?
无需任何先前经验。CoddyKit 上的 AI Agents 课程适合初学者到高级学习者,你可以从这里开始或从头开始,按照自己的节奏学习。 这是第 2 节课,共 4 节。
「事件队列与消息代理」课时需要多长时间?
大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动,所以你能稳步进步,并在网页和应用中从离开的地方继续。
我能在这节 AI Agents 课中编写并运行代码吗?
能。每节 AI Agents 课都包含内置代码编辑器,你可以在浏览器中直接编写并运行真实代码,并获得即时 AI 反馈 — 无需本地设置。