I vaguely recall that this kind of stack trace can happen with certain combinations of IPv6 enabled and the host OS doing unexpected things in the entry for the loopback address in /etc/hosts. (Java parses /etc/hosts as part of looking at the network interface information.) Sorry, but I don't remember the specifics.
Hadoop runs its tests with -Djava.net.preferIPv4Stack=true. Maybe we should try the same? Alternatively, maybe as a first step we can try adding more logging to that test to get more details on what is different in good runs vs. failed runs. --Chris Nauroth On 3/19/16, 11:44 PM, "Patrick Hunt" <[email protected]> wrote: >This is very odd. It failed 2 of the last three times it was run on H9 >with the following: > >2016-03-20 06:06:18,480 [myid:] - INFO >[main:JUnit4ZKTestRunner$LoggedInvokeMethod@74] - TEST METHOD FAILED >testBindByAddress >java.net.SocketException: No such device >at java.net.NetworkInterface.isLoopback0(Native Method) >at java.net.NetworkInterface.isLoopback(NetworkInterface.java:339) >at >org.apache.zookeeper.test.ClientPortBindTest.testBindByAddress(ClientPortB >indTest.java:61) > >https://builds.apache.org/job/ZooKeeper_branch34/buildTimeTrend > >Why would it pass one of the times though if there is no loopback >device on the host? That seems very odd! > >Patrick >
