You’d use Kafka more as an unbounded buffer and build different paradigms on top of it. It not unusual to ingest 100s if mbits of data into kafka, potentially saturating the network while also reading that much out. Amqp is better for large number of queues where each queue has less messages in. Think mqtt, websockets - many, many consumers.

It would be reasonable to use both next to each other.

But I’d never go for rabbitmq. I’d go for azure servicebus or artemis with qpid.

I generally think of messaging systems falling into 4 distinct categores: PubSub, Streaming, Queues and Enterprise Messaging Systems.

PubSub sytems are focused on non-durable (usually), low latency messaging generally without acknowledgements and generally at-most-once. i.e things like Redis PUBSUB, NATS, etc

Queues are generally focused on fanout to multiple consumers with at-least-once processing of durable messages with acknowledgements. i.e Celery/Sidekiq, Que, AWS SQS.

Streaming systems are designed for throughput and usually are based on some form of a distributed log concept. Generally offload offset management to consumers. i.e Kafka, Kinesis

Enterpise Messaging Systems favor flexibility above all else and usually have some mechanism of encoding the flow of data separately from the applications themselves. i.e exchange routing topologies in AMQP as an example. They can generally implement pubsub, queues and direct messaging paradigms. Tradeoffs being poorer availability, complexity and poor performance vs specialised systems. i.e RabbitMQ, HornetMQ, etc.

So you end up using Kafka when it's limitations aren't a problem and you need the throughput. It works best when each and every message in a stream is homogenous as such failure to process a message is unlikely to be independent of failure to process a following message. This alleviates the main drawback of streaming systems which is head of line blocking.

Some cases where it works very well is event streams, data replication/CDC, etc.

If order of message processing matters, then Kafka is better suited then AMQP. For example, In a distributed application for money transfers, if AMQP used, message order will be lost and some problems will occur in the following scenario:

User A with an accound of $1000 makes order for two transfers T1 ($600) and T2 ($500)

  - Rabbit delivers T1 to server1, before processing message, server1 enters a full GC.
  - Rabbit delivers T2 to server2 and server2 processes message immediately, now User A's account have $500  
  - Server1 resumes its life after the end of GC, but fails to process T1 since account's balance is less than required amount.

In Kafka, when user account identifier is used for partitioning key, all User A's messages will be processed by same consumer (i.e server1), so even if server1 enters a full GC, that is OK, since T2 will be processed after T1.

It makes the brokers as dumb as possible to optimize for performance and the logic sits in the client. You can ask to read back from the log at any point or at the current tail. Acknowledging messages is just writing a bookmark to another topic saying where you last read up to, or you can keep track of it yourself somewhere else.

You can always build more complex logic on top which Confluent has done with things like ksqldb.