Back to the real issue, it seems like a concurrency bug, like surefire isn't doing the right then when parallelizing the tests. There is the fork factor of 4 (perThread and then threads set to 4). Originally, this was done in
HBASE-5064; I'm surprised that we didn't break things there. Conceptually, the parallelization shouldn't work unless we get consistently lucky with test order - the minicluster will have problems if it can't find the right port, which it shouldn't if another jvm is also running a mini-cluster.
The this wouldn't be a big issue except that the master goes down if the http server port is already taken:
java.net.BindException: Address already in use
at sun.nio.ch.Net.bind(Native Method)
Should this be the case? I feel like hbasetestutility's minicluster should have the webui off by default, unless otherwise desired (I tend to turn it off in the tests I write, just to speed things up). Also, if the webui is taken, it should just bump the port number to the next available.