# Message queues vs event logs

> RabbitMQ/SQS-style queues delete what they deliver; Kafka-style logs keep it for replay.

Concept · Async work & messaging · https://levelcode.org/concepts/queue-vs-log

Two quite different systems are both called "message brokers".

## Message queues

RabbitMQ, Amazon SQS, Google Cloud Tasks, ActiveMQ.

- A message is delivered to **one** consumer, which processes it and acknowledges it; then it's deleted.
- If the consumer doesn't acknowledge in time (SQS's visibility timeout), the message is redelivered to another consumer.
- Add consumers to process faster ("competing consumers").
- Usually no ordering guarantee, or ordering per group in FIFO queues.

Best for **jobs**: send this email, resize this image, call this webhook.

## Event logs

Apache Kafka, Amazon Kinesis, Redpanda, Apache Pulsar.

- Messages are appended to a **partitioned, ordered log** and retained for a set time (or forever), regardless of whether they've been read.
- Each **consumer group** tracks its own offset per partition. Many groups read the same events independently: billing, analytics and search indexing each get every event.
- Order is guaranteed **within a partition**; choose a key (such as order ID) so related events share a partition.
- Within a consumer group, each partition is read by at most one consumer, so partition count caps parallelism. (Kafka 4.2's *share groups* let several consumers share a partition queue-style, at the cost of ordering.)
- Consumers can **replay** history by resetting offsets — to rebuild a derived store or fix a bug.

Best for **events**: facts that several systems react to, or that you want to keep as a history.

## Choosing

| Need | Choose |
|---|---|
| Distribute independent tasks to workers | Queue |
| Several independent consumers of the same events | Log (or a queue per consumer behind pub/sub fan-out, such as SNS → SQS) |
| Per-key ordering at high throughput | Log |
| Replay or rebuild from history | Log |
| Per-message delays, priorities, simple retries | Queue |

For modest volumes, a database table used as a queue (`SELECT … FOR UPDATE SKIP LOCKED`) is often enough ([when not to add infrastructure](https://levelcode.org/concepts/when-not-to-add-infrastructure)). Whatever you choose, design for redelivery: see [delivery guarantees](https://levelcode.org/concepts/delivery-semantics).