[
https://issues.apache.org/jira/browse/HDFS-16985?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17713467#comment-17713467
]
ASF GitHub Bot commented on HDFS-16985:
---------------------------------------
Hexiaoqiao commented on PR #5564:
URL: https://github.com/apache/hadoop/pull/5564#issuecomment-1512646630
Agree that we need to protect data for this case. But the current
improvement will leave another issue.
Considering that DataNode notify NameNode without delete block file here,
meta at NameNode will be inconsistent at next blockreport round because
`ReplicaMap` is not updated, and it will be involved at next blockreport,
right? The result is that NameNode believe this replica is health, but actually
it had lost.
> delete local block file when FileNotFoundException occurred may lead to
> missing block.
> --------------------------------------------------------------------------------------
>
> Key: HDFS-16985
> URL: https://issues.apache.org/jira/browse/HDFS-16985
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: datanode
> Reporter: Chengwei Wang
> Assignee: Chengwei Wang
> Priority: Major
> Labels: pull-request-available
>
> We encounterd several missing-block problem in our production cluster which
> hdfs running on AWS EC2 + EBS.
> The root cause：
> # the block remains only 1 replication left and hasn't been reconstruction
> # DN checks block file existing when BlockSender construction
> # the EBS checking failed and throw FileNotFoundException (EBS may be in
> fault condition)
> # DN invalidateBlock and schedule block async deletion
> # EBS already back to normal when DN do delete block
> # the block file be delete permanently and can't be recovered
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
