@stack: the patch compiles if you run, e.g...
CLASSPATH="../hadoop-hdfs/build/hadoop-hdfs-0.21.0-dev.jar:../hadoop-hdfs/build/hadoop-hdfs-test-0.21.0-dev.jar" ant bin-package
You're correct that MiniDFSCluster is not a part of hadoop-common but rather hadoop-hdfs.
In the future, I will run something like the following to test patches since I think it closely mimics Hudson, e.g.
ant test-patch -Dpatch.file=HADOOP-4675-v9.patch -Dfindbugs.home=... -Djava5.home=... -Dforrest.home=...
where none of the non-hadoop-common jars are in the classpath.
I read through your test and agree: it is creative. It's a shame that MiniDFSCluster isn't available in common. Maybe the test should be commented out until Mini clusters are available for common tests? It's not ideal but having this patch linger on without resolution isn't ideal either. The code in the patch shouldn't be prone to bugs since it's simply sending specially crafted UDP messages. I trust that Java sockets have been thoroughly tested and the xdr_ functions are very straight-forward (e.g. xdr_int is basically htonl()).
Btw, your approach to sending the ganglia meta data each time is practical. It can be optimized in the future but that might add unwelcome complexity.