[rabbitmq-discuss] Transaction

[Matthias Radestock]

David,

On 06/06/12 10:32, David sUAREZ wrote:
> I have the following problem. In my application I have the following message
> flow.
>
> 1) A publisher sends a message to a queue "A".
> 2) A subcriptor receives the message from queue "A" and send this message to
> the process "P".
> 3) In the process "P" I send a message to the queue "B".
> 4) The subcriptor which received the message sends the ACK to RabbitMQ.
>
> The subcriptor uses a channel "C1" to process messages in the point 2. And
> the point 3 is sent the message to the queue "B" through another channel
> "C2".
>
> I would like the ACK sent by the channel "C1" belonged the same transaction
> that the publication of the message by the channel "C2", that is showed in
> the point 3. Is it possible?. I have only seen  transactions in the same
> channel. Not by connection.
>
> Why do I want this?. May be that I send the publication showed in section 3
> and the machine crash. Then the ack of subcriptor will not be sent. Then the
> "P" process re-processing the previous message and re-send the same message
> showed in the point 3. Duplicating messages.
>
> http://old.nabble.com/file/p33968899/rabbitmq.png

AMQP transactions are channel-bound, so what you want above is not 
possible. Furthermore, AMQP transactions are only guaranteed to be 
atomic when they involve a single queue. See 
http://www.rabbitmq.com/semantics.html#tx for more details.

Generally you'll want to design your system s.t. it can cope with 
duplicate messages. There are many more failure scenarios than the one 
described above, and re-sending of messages, with the resulting 
possibility of duplication, is usually the easiest way to handle them. 
E.g. consider what happens in case of failures in steps 1 or 3; the 
sender has no way of knowing whether the message made it through and 
hence would have to re-send it.

Matthias.