Re: Load balancing and round robin

Round-robin occurs on queues not exchanges. This is because queues are stateful and exchanges are (usually) stateless.

Queues manage state, buffering it on behalf of both broker and consumers. In this equation, they treat messages as resources to be ‘managed’ and queues themselves are shared resources by default. So when you have *multiple* consumers on one queue, then they round robin i.e. each message is delivered to at most one consumer, and no consumer can see another consuer’s messages. This makes queues ideal for patterns like master/worker. If you only have one consumer on one queue, then of course it will get all the messages form that queue.

On the other hand: all the standard Exchanges can be thought of as stateless - like routing tables. They treat messages as values to be copied to any queue that is subscribed with the right pattern or “key”. You can do pubsub with any exchange. Fanout is “one to all”, basically broadcast, and Direct exchanges give you multicast. Direct can be thought of as a restriction on fanout’s broadcast model. So for example if I have two queues bound to one direct exchange with the same key “ibm” then a message tagged “ibm” will be sent to both queues.

The same is true for Topic exchanges: for example if I have two queues bound to one topic exchange with the pattern “system.event” then a message tagged “system.event” will be sent to both queues. This would also happen if one queue were bound with “system.event” and the other with a more general routing pattern such as “system.*”.

You *could* imagine ’stateful’ exchanges eg for last value caching. And, it would be good to have other kinds of exchanges such as anycast, and you could potentially define a ’round robin exchange’, but both of these would be complex. It is usually easier to push all the state management into the queues.

Cheers

alexis

alexis