[rabbitmq-discuss] HA Documentation Error?

[Matthew Sackman]

On Mon, Feb 13, 2012 at 10:15:16AM +0000, Emile Joubert wrote:
> On 10/02/12 19:50, Uday Subbarayan wrote:
> > Note: There is a risk in the above strategy if the master node goes
> > down. If you want to avoid, then publish to all slaves? Here on wards, i
> 
> If the master becomes unavailable then a slave is promoted to take its
> place. Message replication between slaves and the master happens
> internally and automatically. As long as the queue is declared as
> mirrored, the publisher is not required to take any action to republish
> messages to slaves.

Well, that's not necessarily true. If the node that failed was also the
node to which the client was connected, then the client needs to detect
the death of the connection, establish a new connection to a surviving
node and then republish all messages that it never got
publisher-confirms back for - assuming the client even turned
publisher-confirms on. That would be the only way to ensure
at-least-once delivery.

> Active-active mirrored queues are recommended instead of the
> active-passive approach, because it does not require additional software
> and the failover process is quicker.

However, throughput is lower, and it does not have any chance of
integration with other clustering/HA software elsewhere in the stack.

Matthew