New in 0.95, when you do a process listing, there are two 'hbase'
processes. There is the old java process itself and then there is the
launcher script that is just hanging around waiting on the java process to
exit so it can clear the process's znode. Clearing the znode will mean the
master will know sooner of the missing regionserver. Otherwise it would
have to wait wait on the melting of the ephemeral znode (tens of seconds).
This 'watcher' was added by:
HBASE-5844 Delete the region servers znode after a regions server crash
What do folks think about how this facility has been implemented. I am not
questioning the merit of fast-clear of the znode. This is a good idea. I
just want to talk about how it is done. I am not a fan . Our little
wrapper script verges into the process supervision space . Rather than
do up our own, instead should we ship templates for a few common
supervisors and undo our custom one?