[
https://issues.apache.org/jira/browse/HDFS-9657?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15152003#comment-15152003
]
Li Bo commented on HDFS-9657:
-----------------------------
Thanks Zhe’s comment.
90% errors are transient and without permanent data loss, postponing the
recoveries may reduce a lot of unnecessary work. Configuring the task number by
time segments is a solution. I have just created HDFS-9826 which delays the
recoveries for a configured time period. I am thinking about how to combine
them together.
> Schedule EC tasks at proper time to reduce the impact of recovery traffic
> -------------------------------------------------------------------------
>
> Key: HDFS-9657
> URL: https://issues.apache.org/jira/browse/HDFS-9657
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Li Bo
> Assignee: Li Bo
> Attachments: HDFS-9657-001.patch, HDFS-9657-002.patch
>
>
> The EC recover tasks consume a lot of network bandwidth and disk I/O.
> Recovering a corrupt block requires transferring 6 blocks , hence creating a
> 6X overhead in network bandwidth and disk I/O. When a datanode fails , the
> recovery of the whole blocks on this datanode may use up the network
> bandwith. We need to start a recovery task at a proper time in order to give
> less impact to the system.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
