[
https://issues.apache.org/jira/browse/HDFS-2121?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13059156#comment-13059156
]
Allen Wittenauer commented on HDFS-2121:
----------------------------------------
Doesn't this assume that the process reading the block is long-lasting and/or
slow to read the entire block? That doesn't seem to jive with the vast
majority of HDFS use cases. Moving the block would need to be worth the
network overhead incurred.
This goes against the "move the code, not the data" mantra that Hadoop is based
off.
Hotspots are likely better taken care of by increasing replication rather than
just moving them.
> Add ability to create block replicas during off-switch reads
> ------------------------------------------------------------
>
> Key: HDFS-2121
> URL: https://issues.apache.org/jira/browse/HDFS-2121
> Project: Hadoop HDFS
> Issue Type: New Feature
> Reporter: Arun C Murthy
>
> Idea: when a DFSClient reads an off-switch replica, we could tee an extra,
> temporal replica to the local DN if the NN thinks it makes sense.
> Then we could delete the ones which were part of the original 2 on the rack
> to rebalance or just rely on the balancer... all this would ensure better
> spread across the cluster. This would also help with hotspots on datasets.
> Thoughts?
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
