load balancing

[Vadim Zaliva]

Hi!

I am trying to use ActiveMQ and Stomp protocol to implement scenario  
where are multiple consumers are processing messages from single  
queue. I want to make sure that only one consumer will get the  
message, and he must acknowledge it, so the message will get back to  
queue if it has not acknowledged for some time.

One of problems I see is that there is no NAK command in stomp. So,  
once I have received the message I could not send NAK to reject it  
back to the queue, so it will be delivered to another consumer.

Another problem, or rather a question, is what happens to not ACK-ed  
messages. Are they delivered to other consumers and under what  
conditions?

Perhaps I can use stomp transactions, but I am not sure how they apply  
to receiving messages.

Sincerely,
Vadim

---------------------------------------------------------------------
To unsubscribe from this list, please visit:

    http://xircles.codehaus.org/manage_email

[andrewk]

Hello,

I too have complained of the missing NAK feature in these various forums - it comes up every few months now I would say.

In client ack mode, if you don't ack, then the message is retained on the server in some limbo, but not sent again (yet).

Once the connection is lost to the client that received the message, but did not ACK, the message is returned to the queue, and will then be sent again to a different connection/subscription.

Some have suggested before to use disconnect/reconnect/resubscribe as a poor man's NAK - as this has the effect of returning the un-ack'd message back to use.  This does have the downside that the same message could keep getting resent again and again, although perhaps the message timeout settings in activemq would help.  It has has the downside that unless you are using durable topics, it can result in missing messages.

Stomp transactions do no not help either, as you correctly point out, they only work with sending, not receiving, messages.

I wrote a workaround (that not everyone likes) for use in ActiveMessaging.

Rather than use the hack of severing the connection to simulate a NAK, I wrote my own logic to create a clone of the message (retaining a count of how many times it has been cloned, original message id and send time, etc. in custom headers), and then in a transaction ack the original message and put this cloned message on the queue. So in effect, it is like returning the original message, except because it is a clone, it loses its place in the queue, and some meta data like message id will differ.

Not a perfect solution, but I like it better than killing the connection and resubscribing for the reasons above, though I'll probably end up implementing that option too (there was a user offering to do it and send me a patch, but I don't think that ended up happening).

-Andrew Kuklewicz


 

Vadim Zaliva wrote:

Hi!

I am trying to use ActiveMQ and Stomp protocol to implement scenario  
where are multiple consumers are processing messages from single  
queue. I want to make sure that only one consumer will get the  
message, and he must acknowledge it, so the message will get back to  
queue if it has not acknowledged for some time.

One of problems I see is that there is no NAK command in stomp. So,  
once I have received the message I could not send NAK to reject it  
back to the queue, so it will be delivered to another consumer.

Another problem, or rather a question, is what happens to not ACK-ed  
messages. Are they delivered to other consumers and under what  
conditions?

Perhaps I can use stomp transactions, but I am not sure how they apply  
to receiving messages.

Sincerely,
Vadim

---------------------------------------------------------------------
To unsubscribe from this list, please visit:

    http://xircles.codehaus.org/manage_email

[Dexter Tad-y]

Hi,

ActiveMQ supports a Stomp SUBSCRIBE header called 'activemq.prefetchSize'. You might want to set this to 1 so that there is only one message pending for the consumer. This minimizes the number of unack'd messages at a time.

A solution I can think of to the NACK requirement is to wrap your ACK frame and SEND (for re-enqueue) under one transaction. This way, the message goes back to the queue. The drawback to this approach is that the message might get sent back to the same consumer.

To solve the problem of the above solution, you can extend it by implementing a routing mechanism:
  1. all consumer subscribes from a main queue
  2. each consumer is assigned to listen to its own queue
  3. if a consumer decides to NACK a message (either from the main queue or its own), it should send the message to the next consumer.

This solution solves the problem where the same message goes to the same consumer who NACK'd the message in the first place. But it adds complexity to the solution where in each consumer is required to know the queue(s) of the other consumer(s).

Cheers!

Dex


--- On Sat, 11/15/08, Vadim Zaliva <krokodil@...> wrote:

> From: Vadim Zaliva <krokodil@...>
> Subject: [stomp-user] load balancing
> To: user@...
> Date: Saturday, November 15, 2008, 12:37 PM
> Hi!
>
> I am trying to use ActiveMQ and Stomp protocol to implement
> scenario where are multiple consumers are processing
> messages from single queue. I want to make sure that only
> one consumer will get the message, and he must acknowledge
> it, so the message will get back to queue if it has not
> acknowledged for some time.
>
> One of problems I see is that there is no NAK command in
> stomp. So, once I have received the message I could not send
> NAK to reject it back to the queue, so it will be delivered
> to another consumer.
>
> Another problem, or rather a question, is what happens to
> not ACK-ed messages. Are they delivered to other consumers
> and under what conditions?
>
> Perhaps I can use stomp transactions, but I am not sure how
> they apply to receiving messages.
>
> Sincerely,
> Vadim
>
> ---------------------------------------------------------------------
> To unsubscribe from this list, please visit:
>
>    http://xircles.codehaus.org/manage_email

     

---------------------------------------------------------------------
To unsubscribe from this list, please visit:

    http://xircles.codehaus.org/manage_email

[Vadim Zaliva]

On Nov 14, 2008, at 21:10, andrewk wrote:

> I too have complained of the missing NAK feature in these various  
> forums -
> it comes up every few months now I would say.

I would like to thank everybody for their responses. Looks like I will
be switching to RabbitMQ from ActiveMQ and to AMQP from STOMP. AMQP
have NAK feature which will allow me to implement load-balancing  
semantics
in more natural manner.

Sincerely,
Vadim

---------------------------------------------------------------------
To unsubscribe from this list, please visit:

    http://xircles.codehaus.org/manage_email