kimchy
commented
14 years ago The elasticsearch script currently set the preferIPv4Stack to true because of problems I had with using the default in Java (which defaults to the machine). This menas that you need to either change the flag on the script, or set the flag on the embedded client side.
I need to revisit this decision to set the ipv4 flag to true and use the JVM default, but it needs to work well out of the box, without needing to set it explicitly. See this thread for why I default to ipv4: http://old.nabble.com/Suspect-warning-in-jgroups-2.8.0-CR6-to26560364.html#a26560364
When I run a cluster made up of embedded and standalone nodes, former ones seems not to be able to connect to latter ones, and I get the following log messages on embedded nodes:
OOB-1,elasticsearch,caffeine.local-16363 - jgroups.pbcast.NAKACK - caffeine.local-16363: dropped message from caffeine.local-17690 (not in xmit_table), keys are [caffeine.local-16363], view=[caffeine.local-16363|0] [caffeine.local-16363]
The problem seems to be related to JGroups which, on IPv6-enabled machines, prefers IPv6 addresses over IPv4 ones, but embedded instances aren't able to guess a proper IPv6 address.
That can be fixed by: Specifying java.net.preferIPv4Stack=true on the embedded node, forcing the use of IPv4. or Specifying java.net.preferIPv4Stack=false and java.net.preferIPv6Stack=false on the standalone node, apparently forcing JGroups to take the proper default.