100% Free Forever
AI-Powered Learning
Industry Expert Content
Certificates & Badges
Learn At Your Own Pace
Messaging

Work Queues and Load Balancing

Learn how RabbitMQ distributes tasks across a pool of competing consumers, and how prefetch settings enable fair, load-based dispatch.

Messaging PatternsIntermediate9 min readJul 10, 2026
Analogies

Work Queues: Distributing Tasks Across Workers

A work queue lets multiple consumer processes share a single queue so that time-consuming tasks are distributed across a pool of workers rather than processed by one overloaded process. Unlike publish-subscribe, where every subscriber gets its own copy, a work queue delivers each message to exactly one of the competing consumers — this is often called competing consumers. By default RabbitMQ uses round-robin dispatch, handing message 1 to worker A, message 2 to worker B, message 3 to worker A again, and so on, regardless of how long each task actually takes.

🏏

Cricket analogy: It is like a franchise assigning bowlers in rotation from an over-by-over roster — instead of one bowler bowling every over of an IPL innings, deliveries are shared out among the bowling attack so no single player is overworked.

Fair Dispatch with basic_qos (prefetch)

Plain round-robin dispatch can be unfair when tasks vary wildly in duration: a fast worker might sit idle while a slow worker is buried under a backlog of unacknowledged messages it already claimed. Setting channel.basic_qos(prefetch_count=1) tells RabbitMQ not to dispatch a new message to a worker until that worker has acknowledged its current one, which turns naive round-robin into genuine load-based fair dispatch. This is essential for CPU- or I/O-heavy tasks like video transcoding or PDF generation, where task duration is unpredictable.

🏏

Cricket analogy: It is like a captain refusing to send the next batter in until the current one is either out or has completed their innings, rather than blindly rotating the batting order on a fixed schedule regardless of how a partnership is going.

python
import pika

connection = pika.BlockingConnection(pika.ConnectionParameters(host="localhost"))
channel = connection.channel()
channel.queue_declare(queue="pdf_generation", durable=True)

# Only give this worker one unacked message at a time -> fair dispatch
channel.basic_qos(prefetch_count=1)

def callback(ch, method, properties, body):
    print(f"Generating PDF for job: {body}")
    # ... do the CPU-heavy work ...
    ch.basic_ack(delivery_tag=method.delivery_tag)

channel.basic_consume(queue="pdf_generation", on_message_callback=callback)
channel.start_consuming()

Setting prefetch_count too high (or leaving it unlimited, the default) can starve other workers and, combined with slow processing, cause one worker to hoard hundreds of unacked messages that sit idle instead of being redistributed.

Scaling Workers Horizontally

Because work queues use competing consumers on a shared queue, scaling throughput is often as simple as starting more worker processes pointed at the same queue name — no code change is needed in the producer or the queue definition. This horizontal scaling model is why RabbitMQ work queues pair naturally with container orchestration: a Kubernetes Deployment can scale a worker pod from 2 replicas to 20 in response to queue depth metrics, and RabbitMQ will automatically start round-robining (or fair-dispatching, with prefetch set) messages across the new replicas as soon as they connect and start consuming.

🏏

Cricket analogy: It is like a franchise calling up extra net bowlers from the academy during a packed tournament schedule — you add more bowlers to the same practice queue rather than redesigning how the nets are organized.

  • Work queues distribute tasks across a pool of competing consumers sharing one queue, unlike pub-sub's one-copy-per-subscriber model.
  • Default round-robin dispatch ignores task duration, which can leave slow workers overloaded while fast workers idle.
  • basic_qos(prefetch_count=1) enables fair dispatch by withholding new messages from a worker until it acknowledges its current one.
  • Manual acknowledgment (not auto-ack) is required for fair dispatch and for RabbitMQ to know a worker is actually free.
  • Horizontal scaling is as simple as starting more workers on the same queue name, with no producer or queue changes required.
  • Work queues pair naturally with autoscaling systems like Kubernetes HPA driven by queue-depth metrics.

Practice what you learned

Was this page helpful?

Topics covered

#Messaging#RabbitMQStudyNotes#Database#WorkQueuesAndLoadBalancing#Work#Queues#Load#Balancing#DataStructures#StudyNotes#SkillVeris