Microservice Patterns
Request/Reply over Messaging
Get an answer back through a message broker: send a request with a reply address and a correlation id, match replies as they arrive, and decide what a missing reply means.
What you will learn
- Implement request/reply with replyTo addresses and correlation ids.
- Handle timeouts and late replies safely.
- Avoid shared reply queues between requester instances.
- Decide when request/reply over messaging is better than plain HTTP.
Before you start
Overview
When you post a letter that needs an answer, you write your return address on the envelope and a reference number inside ("re: your invoice 42"). The reply comes to your mailbox, and the reference tells you which letter it answers. Request/reply over messaging works exactly like that.
The requester sends a message with two headers: replyTo (the queue where it reads replies) and a correlation id. A worker processes the request and publishes the result to the replyTo queue, copying the correlation id. The requester matches incoming replies to its pending requests by that id.
Run it
A shows the happy path with two workers answering in any order. In B workers sometimes crash after the work but before replying: the requester times out and the redelivered request later produces a reply nobody is waiting for. In C two requester instances share one reply queue — switch on per-instance queues to fix it.
Rules that keep it correct
- One reply queue per requester instance (or a shared queue plus routing by instance id). Replies must reach the instance that holds the pending request.
- Always set a timeout and treat it as "unknown outcome", not "failed": the work may still complete.
- Ignore or log late replies by correlation id; never apply them to a different request.
- Make the work idempotent: redelivery after a crash runs it again (see Idempotent Consumer).
Implementation with Spring AMQP
// Requester: RabbitTemplate creates a reply queue, sets replyTo + correlation id and waits.
@Service
class PriceClient {
private final RabbitTemplate rabbit;
PriceClient(RabbitTemplate rabbit) {
this.rabbit = rabbit;
this.rabbit.setReplyTimeout(2_000);
}
Price quote(QuoteRequest request) {
Price price = (Price) rabbit.convertSendAndReceive("pricing", "price-requests", request);
if (price == null) throw new PricingTimeoutException(request.id()); // timeout: outcome unknown
return price;
}
}
// Worker: the return value is published to the request's replyTo address automatically.
@Component
class PricingWorker {
@RabbitListener(queues = "price-requests", concurrency = "2")
Price handle(QuoteRequest request) {
return engine.price(request);
}
}Advanced With Kafka, ReplyingKafkaTemplate does the same using a reply topic and the KafkaHeaders.CORRELATION_ID header; give each requester instance its own reply partition or topic. For long-running work prefer the HTTP flavour of the pattern: return 202 Accepted with a status URL the client polls.
Trade-offs
| Benefits | Costs | |
|---|---|---|
| Decoupling | Requesters do not need to know where workers run; the broker buffers bursts. | More moving parts than an HTTP call: queues, correlation, timeouts. |
| Load | Workers pull work at their own pace. | The requester still waits — latency includes broker hops and queueing. |
| Failures | Requests survive worker crashes (redelivery). | Late and duplicate replies must be handled explicitly. |
Production considerations
- Use exclusive, auto-delete reply queues (or direct reply-to in RabbitMQ) so stale queues do not pile up.
- Propagate the correlation id into logs and traces — it is the only link between request and reply.
- Alert on unmatched replies: they reveal timeouts that are too short or routing mistakes.