[
https://issues.apache.org/jira/browse/HADOOP-2956?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12575964#action_12575964
]
eric baldeschwieler commented on HADOOP-2956:
---------------------------------------------
I think we should try for some simple heuristics and see if they can add value.
One would be to just watch for nodes with significantly less read / write
throughput than other nodes. Since off rack writes are randomly assigned,
these should be relatively evenly distributed. If one node consistently under
performs, that should be a huge red flag.
Another thing to watch is corrupted blocks.
> HDFS should blacklist datanodes that are not performing well
> ------------------------------------------------------------
>
> Key: HADOOP-2956
> URL: https://issues.apache.org/jira/browse/HADOOP-2956
> Project: Hadoop Core
> Issue Type: Improvement
> Components: dfs
> Reporter: dhruba borthakur
>
> On a large cluster, a few datanodes could be under-performing. There were
> cases when the network connectivity of a few of these bad datanodes were
> degraded, resulting in long long times (in the order of two hours) to
> transfer blocks to and from these datanodes.
> A similar issue arises when disks a single disk on a datanode fail or change
> to read-only mode: in this case the entire datanode shuts down.
> HDFS should detect and handle network and disk performance degradation more
> gracefully. One option would be to blacklist these datanodes, de-prioritise
> their use and alert the administrator.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
