[javagroups-users] JGroups 2.2.9.1: problems with UNICAST after network failure

[Matthias W. <mat...@ba...>]

Hi,

I have 3 processes running, two of them sitting on machine A (I'll call them A1 
and A2), one sitting on machine B (I'll call it B1). A1 is the coordinator. In 
case of temporary network failure the group is splitted into two subgroups: 
group G1 with A1 as coordinator and A2 as groupmember and group G2 with B1 as 
coordinator.

I get into trouble after the network failure is resolved: the two subgroups are 
merged into one group again with A1 as coordinator. The new view ist 
distributed among members, I can send multicast messages to the group without 
problems. I can also send unicast messages between A1 and B1. But there is no 
chance to send unicast messages between A2 and B1: every message is discarded 
by UNICAST with "discarding message to XXX as this member left the group" as 
log entry.

After a look into the sources, I guess the problem lies here:

UNICAST keeps a list of previous group members (UNICAST.previous_members). In 
case of a view change with use_gms set to "true" (default) the leaving members 
are added to this list. If UNICAST.down() detects that the message receiver is 
on this list, the message is discarded. The member is only removed from the 
list if a unicast message is received from it. So we have a problem: B1 can't 
send a unicast message to A2 until ist gets one from A2 and vice versa. A1 and 
A2 can exchange unicast messages because coordinators send 
Event.ENABLE_UNICASTS_TO down the stack after a merge request.

It would be nice, if someone could confirm this.

My protocol stack:

<?xml version="1.0" encoding="ISO-8859-1"?>
<config>
<UDP 
	mcast_addr="239.0.0.1"
	mcast_port="10001"
	ip_ttl="1"    
	mcast_send_buf_size="32000"
	mcast_recv_buf_size="64000"
	ucast_send_buf_size="32000"
	ucast_recv_buf_size="64000"
	loopback="true"
	use_incoming_packet_handler="true"
	use_outgoing_packet_handler="false"
/>
<PING 
	timeout="3000"
	num_initial_members="1"
	down_thread="false"
	up_thread="false"
/>
<MERGE2 
	max_interval="10000"
	min_interval="5000"
	down_thread="false"
	up_thread="false"		
/>
<FD_SOCK 
	down_thread="false"
	up_thread="false"	
/>
<FD 
	timeout="10000"
	max_tries="2"
	shun="true"
	down_thread="false"
	up_thread="false"		
/>
<VERIFY_SUSPECT 
	timeout="1500" 
	down_thread="false"
	up_thread="false"		
/>
<pbcast.NAKACK 
	max_xmit_size="8192"
	gc_lag="50"
	retransmit_timeout="600,1200,2400,4000"
	down_thread="false"
	up_thread="false"		
/>
<UNICAST 
	timeout="1000,1500,2000"
	down_thread="false"
	up_thread="false"
/>	
<pbcast.STABLE 
	stability_delay="1000"
	desired_avg_gossip="20000"
	max_bytes="0"
	down_thread="false"
	up_thread="false"		
/>
<FRAG2 
	frag_size="8192"
	down_thread="false"
	up_thread="false"
/>
<VIEW_SYNC 
	avg_send_interval="60000"
	down_thread="false"
	up_thread="false"
/>	
<pbcast.GMS 
	print_local_addr="true"
	join_timeout="3000"
	join_retry_timeout="2000"
	shun="true"
	down_thread="false"
	up_thread="false"		
/>
</config>

Regards,
Matthias Weber