hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Bryan Pendleton (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-91) snapshot a map-reduce to DFS ... and restore
Date Mon, 20 Mar 2006 18:53:58 GMT
    [ http://issues.apache.org/jira/browse/HADOOP-91?page=comments#action_12371127 ] 

Bryan Pendleton commented on HADOOP-91:

This would be very useful, although, it should be noted that snapshotting a job to DFS means
that it will take as much extra space to store as the replication (well, replication+1) level.
If you're running jobs that produce large intermediate results, then attempting to checkpoint
with, say, the default 3x replication, requires 4 times as much space as the job would, otherwise.
For no-side-effect jobs, perhaps the default should be to checkpoint but with replication
of 1 (assuming per-file replication gets added to DFS), and just let lost blocks turn into
lost tasks that just get re-run. Hadoop should minimize space usage wherever possible, if
it's really going to scale up to huge workloads.

> snapshot a map-reduce to DFS ... and restore
> --------------------------------------------
>          Key: HADOOP-91
>          URL: http://issues.apache.org/jira/browse/HADOOP-91
>      Project: Hadoop
>         Type: New Feature
>   Components: mapred
>     Reporter: eric baldeschwieler
>     Priority: Minor

> The idea is to be able to issue a command to the job tracker that
> will halt a map-reduce and archive it to a directory in such a way
> that it can later be restarted.
> We could also set a mode that would cause this to happen to a job
> when it fails.  This would allow one to debug and restart a failing
> job reasonably, which might be important, for long running jobs.  It
> has certainly been important in similar systems I've seen before.  One 
> could restart with a new jar or work bench a single failing map or reduce.

This message is automatically generated by JIRA.
If you think it was sent incorrectly contact one of the administrators:
For more information on JIRA, see:

View raw message