[jboss-jira] [JBoss JIRA] (JGRP-2260) UNICAST3 doesn't remove dead nodes from its tables

[Rich DiCroce (JIRA)]

Rich DiCroce created JGRP-2260:
----------------------------------

             Summary: UNICAST3 doesn't remove dead nodes from its tables
                 Key: JGRP-2260
                 URL: https://issues.jboss.org/browse/JGRP-2260
             Project: JGroups
          Issue Type: Bug
    Affects Versions: 4.0.10
         Environment: WildFly 12.0.0.Final
            Reporter: Rich DiCroce
            Assignee: Bela Ban


Scenario: 2 WildFly instances clustered together. A ForkChannel is defined, with a MessageDispatcher on top. I start both nodes, then stop the second one. 6-7 minutes after stopping the second node, I start getting log spam on the first node:

{quote}
12:47:04,519 WARN  [org.jgroups.protocols.UDP] (TQ-Bundler-4,ee,RCD_GP (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null)) JGRP000032: RCD_GP (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null): no physical address for RCD_NMS (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null), dropping message
12:47:06,522 WARN  [org.jgroups.protocols.UDP] (TQ-Bundler-4,ee,RCD_GP (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null)) JGRP000032: RCD_GP (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null): no physical address for RCD_NMS (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null), dropping message
12:47:08,524 WARN  [org.jgroups.protocols.UDP] (TQ-Bundler-4,ee,RCD_GP (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null)) JGRP000032: RCD_GP (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null): no physical address for RCD_NMS (flags=0), site-id=DEFAULT, rack-id=null, machine-id=null), dropping message
{quote}

After some debugging, I discovered that the reason is because UNICAST3 is still trying to retransmit to the dead node. Its send_table still contains an entry for the dead node with state OPEN.

After looking at the source code for UNICAST3, I have a theory about what's happening.
* When a node leaves the cluster, down(Event) gets invoked with a view change, which calls closeConnection(Address) for each node that left. That sets the connection state to CLOSING.
* Suppose that immediately after the view change is handled, a message with the dead node as its destination gets passed to down(Message). That invokes getSenderEntry(Address), which finds the connection... and sets the state back to OPEN.

Consequently, the connection is never closed or removed from the table, so retransmit attempts continue forever even though they will never succeed.

This issue is easily reproducible for me, although unfortunately I can't give you the application in question. But if you have fixes you want to try, I'm happy to drop in a patched JAR and see if the issue still happens.

This is my JGroups subsystem configuration:
{code:xml}
        <subsystem xmlns="urn:jboss:domain:jgroups:6.0">
            <channels default="ee">
                <channel name="ee" stack="main">
                    <fork name="shared-dispatcher"/>
                    <fork name="group-topology"/>
                </channel>
            </channels>
            <stacks>
                <stack name="main">
                    <transport type="UDP" socket-binding="jgroups" site="${gp.site:DEFAULT}"/>
                    <protocol type="PING"/>
                    <protocol type="MERGE3">
                        <property name="min_interval">
                            1000
                        </property>
                        <property name="max_interval">
                            5000
                        </property>
                    </protocol>
                    <protocol type="FD_SOCK"/>
                    <protocol type="FD_ALL2">
                        <property name="interval">
                            3000
                        </property>
                        <property name="timeout">
                            8000
                        </property>
                    </protocol>
                    <protocol type="VERIFY_SUSPECT"/>
                    <protocol type="pbcast.NAKACK2"/>
                    <protocol type="UNICAST3"/>
                    <protocol type="pbcast.STABLE"/>
                    <protocol type="pbcast.GMS">
                        <property name="join_timeout">
                            100
                        </property>
                    </protocol>
                    <protocol type="UFC"/>
                    <protocol type="MFC"/>
                    <protocol type="FRAG3"/>
                </stack>
            </stacks>
        </subsystem>
{code}



--
This message was sent by Atlassian JIRA
(v7.5.0#75005)