Non ws browsers behind clustered rmq

[redsquare]

Hi, I have a three node cluster of rabbitmq load balanced by nginx (tried haproxy also). When I use sock.js/stomp.js to connect on either ie9/android stock (basically and non ws client) then I cannot connect and hold an open connection. I believe this is due to the cowboy plugin/webserver doing things in memory so when the next request comes into a different node there is no knowledge of the address. This happens with the /stomp/xhr_send routes. Is there no way to ha this across the cluster or do we need to deploy 'sticky/ip_hash' sessions?

[redsquare]

I undertstand the issue better, its down to webstomp creating exclusive queues which do not get mirrored meaning any non websocket browser client must end up polling to the same node.

[michaelklishin]

Yes, this is an interesting consequence of using exclusive queues. This depends on rabbitmq/rabbitmq-stomp#24, then.

[michaelklishin]

@essen now that rabbitmq/rabbitmq-stomp#24 is fixed, we can look into propagating endpoint/queue attributes for Web STOMP clients, too:

• durable
• auto-delete
• exclusive

via headers or similar. What do you think?

[essen]

I think I need to read more about it before I can answer that. :-)

I'll get back to you.

[essen]

From what I understand, we shouldn't have to do anything.

The Web STOMP plugin just receives STOMP data as part of Websocket frames or other. There's no filtering going on that I can tell, we call rabbit_stomp_processor directly. This means that if you set the STOMP headers they will be propagated to rabbitmq-stomp and rabbitmq-stomp will handle them as expected.

I can experiment to confirm. How can I know if a queue is either durable/auto-delete/exclusive, though?

[michaelklishin]

@essen you can take a look at what flags (D, AD, E) it has in the management UI or list queues with rabbitmqctl. I'll assume this is closed, feel free to give it a try and re-open if things don't work as intended in master.