One discussion point...
In previous systems we've built, we've tended to have a master node
configured with a list of all of its slaves. Dead or inactive slaves
are then listed as such in the config as well. This let's the master
know what nodes to expect. It seems to me that since the HDFS relies
on a stable (roughly) set of nodes to provide persistent storage,
moving a node list to the master will allow simpler startup code and
general error checking, without loosing any real freedom.
I'd not suggest this for the task tracker at this point, since it can
actually use ad hoc nodes, but HDFS gains nothing but instability
from ah hoc nodes showing up and dropping out. Much better to have a
configured list of servers IMO.
Thoughts? Should we file a bug to make this so?
On Mar 20, 2006, at 1:33 PM, Yoram Arnon wrote:
While playing around with a hadoop dfs cluster, we've observed data
loss.
This may be related to our having stopped and restarted the DFS a
couple of
times, possibly with nodes not all going online and offline at just
the
right timing, but many of our files, ranging in size from less than
1GB to
multi GB each, have at least one block missing. Blocks are missing
from
relatively new files, generated within the last two weeks, the file
system
was never more than 25% full, and there's no outstanding reason why
this
loss should have happened.
In the upcoming days/weeks we'll be looking into the reasons, and
for ways
of making the DFS more robust against this kind of loss.
Yoram
