Currently, we override RABBITMQ_SERVER_ERL_ARGS in tripleo in order to set a blanket TCP timeout of 15 seconds on all connecting and listening erlang sockets.
This is leftover from long ago when oslo.messaging did not have proper heartbeat support. Without AMQP heartbeats, it was not possible to detect quickly when a client lost connection, so the TCP timeout was implemented.
Heartbeat support has been available in oslo.messaging for many releases now, so this is no longer required to detect dead clients and can be removed.
However, as a somewhat-intended side effect, setting the TCP timeouts also affects dead peer detection on the connections between RabbitMQ nodes in a clustered environment. Normally these timeouts are configured via the erlang net_ticktime mechanism, see https://www.rabbitmq.com/nettick.html for more info.
Using net_ticktime is preferrable to TCP timeouts. It is the standard practice, it is simpler, and also in testing seems to detect failures faster (see
https://bugzilla.redhat.com/show_bug.cgi?id=1485989).
Fix proposed to branch: master
Review: https:/