Harden stack to prevent loss of intra-stack events -------------------------------------------------- Key: JGRP-180 URL: https://jira.jboss.org/jira/browse/JGRP-180 Project: JGroups Issue Type: Task Affects Versions: 2.2.8, 2.2.9, 2.2.9.1 Reporter: Bela Ban Assignee: Bela Ban Fix For: 2.x When an *intra-stack* event is lost, this can be serious (*inter-stack* message are simply retransmitted, so that is not critical). Losses are mainly caused by runtime exceptions, e.g. OutOfMemory exception (OOM), or resource problems, and to a lesser extent by program bugs. For example, when a user sends a message using Channel.send(), and there is an exception, then the user will simply send the message again, possibly after fixing the cause of the exception. However, for events such as VIEW_CHANGE that are multicast by the GMS protocol, a loss can be serious: in this case, the view would never be received ! The same applies to the up direction: when NAKACK has successfully delivered a message, if that message is lost travelling between NAKACK and the Channel, then is serious (essentially loss of that message). So while these error situations don't occur very often, if they do occur, they have serious consequences. SOLUTION: - Do nothing for user messages: Channel.send() throws an exception, user has to resend message. Note that in http://jira.jboss.com/jira/browse/JGRP-179, we made retransmission handling atomic, e.g. if there is an exception, there will *not* be a gap in the seqnos for NAKACK and UNICAST - Provide either a pass{Up/Down}Reliably() method or an Event with a RELIABLE field, such that this event needs to be acked. The sender (e.g. GMS on a VIEW_CHANGE) sends down the message and waits until it gets an ACK, which could be sent by the NAKACK or UNICAST protocols, or as last resort by the transport (TP). If the ACK is not received within M ms, the event is resent.