When should I use a message queue instead of a direct API call?

Use a queue when the caller does not need the result to finish its own work: sending emails, generating files, syncing data, notifying other systems. Keep direct calls for anything the user is actively waiting on, like an authorization check during checkout. A useful test is asking whether the operation could happen five minutes late without anyone noticing; if yes, it belongs on a queue.

Why do people say exactly-once delivery is impossible?

Because a consumer can always crash after performing its side effect but before acknowledging the message, and the broker cannot tell that case apart from a crash before the side effect. It must choose between redelivering (possible duplicate) or not (possible loss). The practical answer is at-least-once delivery plus idempotent consumers, which yields exactly-once effects even though delivery itself is not exactly-once.

What metrics should I monitor on a queue?

Queue depth, the age of the oldest message, consumer error rate, and dead letter queue depth. Depth alone can mislead since a fast consumer can handle a deep queue, but a rising oldest-message age means you are falling behind in a way users will eventually feel. DLQ depth should alert at low thresholds, because every message there is promised work that did not happen.

Communicationchapter 2 of 4 · 4 lessons

Message Queues

Async broker that decouples producers from consumers in time, scale, and failure.

← All chapters View in cheatsheet

Your checkout calls the email service to send a receipt, and waits for it to finish before telling the customer the order went through.

One afternoon the email provider slows down. Not fails, slows: 30 seconds a response. Your checkout threads pile up waiting, the pool runs out, and now nobody can buy anything at all because a receipt is slow.

Nothing was wrong with your checkout code. It was tied to the slowest thing it touched.

A queue puts a buffer between the two. Checkout drops a message and moves on in a few milliseconds, and the email service picks it up whenever it can. That one move is why a queue sits in the middle of almost every system that has grown past a single process.

Lessons

4 in this chapter

Why Decouple With a QueueThe producer's job ends at enqueue, so slow or dead consumers stop being the producer's problem.2 min
Delivery GuaranteesAt-most-once loses messages, at-least-once duplicates them, and exactly-once is mostly your code's job.3 min
Retries and Dead Letter QueuesRetry transient failures with backoff, and give permanent failures somewhere to go that is not the front of the queue.3 min
Kafka vs RabbitMQ Style BrokersA log you read at your own pace versus a broker that hands out jobs and deletes them.3 min