|| at Jul 2, 2008 at 12:50 pm
Sangmin Lee wrote:
Thank you for the replies.
I meant porting HDFS to Zookeeper to make namenode fault-tolerant and highly
available rather than the simple usage of Zookeeper. Basically by storing
namespace metadata in Zookeeper, we may be able to get highly available
replicated namenodes. So I meant using Zookeeper as frontend of HDFS. Anyway
it seems to me that there has been no actual work going on regarding this,
isn't it? :-)
I think you've just volunteered.
@work we use our own tuple space thing, Anubis:http://smartfrog.svn.sourceforge.net/viewvc/smartfrog/trunk/core/components/anubis/doc/
it lets you do the election/voting stuff to see who is in charge in an
HA cluster, though to deal with a partition event robustly is hard; you
need somebody that really understands the maths behind HA distributed
computing. I am not in that category, so we are lucky there are others
nearby who do, but it implies there is a block of code I am very
reluctant to approach.
In fact, its a bit like security. just like it is easy to do an insecure
system, but very hard to do a secure one, its easy to be
low-availability, really, really hard to be HA.