[
https://issues.apache.org/jira/browse/HDFS-15170?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246935#comment-17246935
]
Wei-Chiu Chuang commented on HDFS-15170:
----------------------------------------
{code:java}
+ // If the block is an EC block, the whole block group is marked
+ // corrupted, so if this block is getting deleted, remove the block
+ // group from corrupt replica map explicitly, since removal of the
+ // block from corrupt replicas may be delayed if the blocks are on
+ // stale storage due to failover or any other reason.
+ corruptReplicas.removeFromCorruptReplicasMap(b.getStored(), node);
^^^^^^^^^^^^^
{code}
this line is not needed if
dfs.namenode.corrupt.block.delete.immediately.enabled is true? it will be
removed later by invalidateBlock().
Or do we want to call {{corruptReplicas.removeFromCorruptReplicasMap(corrupt,
node)}} instead? b.getStored() is the internal block whereas corrupt is the EC
block group id. The code doesn't seem to match the comment.
> EC: Block gets marked as CORRUPT in case of failover and pipeline recovery
> --------------------------------------------------------------------------
>
> Key: HDFS-15170
> URL: https://issues.apache.org/jira/browse/HDFS-15170
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: erasure-coding
> Reporter: Ayush Saxena
> Assignee: Ayush Saxena
> Priority: Critical
> Attachments: HDFS-15170-01.patch, HDFS-15170-02.patch,
> HDFS-15170-03.patch
>
>
> Steps to Repro :
> 1. Start writing a EC file.
> 2. After more than one stripe has been written, stop one datanode.
> 3. Post pipeline recovery, keep on writing the data.
> 4.Close the file.
> 5. transition the namenode to standby and back to active.
> 6. Turn on the shutdown datanode in step 2
> The BR from datanode 2 will make the block corrupt and during invalidate
> block won't remove it, since post failover the blocks would be on stale
> storage.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
