[
https://issues.apache.org/jira/browse/HDFS-8669?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jing Zhao updated HDFS-8669:
----------------------------
Attachment: HDFS-8669.000.patch
Initial patch to fix. The main change is to add capability in StripeReader so
that it can check missing block locations or failed block readers and
initialize parity chunk reading before submitting the read requests.
Still need to add more tests and fix bugs.
> Erasure Coding: handle missing internal block locations in
> DFSStripedInputStream
> --------------------------------------------------------------------------------
>
> Key: HDFS-8669
> URL: https://issues.apache.org/jira/browse/HDFS-8669
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Jing Zhao
> Assignee: Jing Zhao
> Attachments: HDFS-8669.000.patch
>
>
> Currently DFSStripedInputStream assumes we always have complete internal
> block location information, i.e., we can always get all the DataNodes for a
> striped block group. In a lot of scenarios the client cannot get complete
> block location info, e.g., some internal blocks are missing and the NameNode
> has not finished the recovery yet. We should add functionality to handle
> missing block locations in DFSStripedInputStream.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
