I copied the metric.h from 2.4.1 to a recent 2.5.0 checkout and compiled it
for Linux. I'm trying it on my Linux data source now, and although it
doesn't seem to be susceptible to the connection-dropping crash, I *can*
still get the multicast listen thread to hang if I hammer it with enough
connections.
The last peep I get out of the mcast_listen_thread() is it attempting to
hash_insert_data() on a value it's received. And then that's it, not
another peep from any of the mcast threads until or unless I restart.
During this time, the unicast listen thread is still active, and it *does*
perform a data dump.
Odd.
Anyway, I think this still could possibly be related to why gmond
occasionally crashes on that box, but I'm *still* getting erroneous dead
cluster messages from gmetad...
