[
https://issues.apache.org/jira/browse/HDFS-17048?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17732434#comment-17732434
]
Xiaoqiao He commented on HDFS-17048:
------------------------------------
[~liuguanghua] Thanks for your report. Sorry I didn't understand this case
totally, would you mind to give some way to reproduce this case. IIUC, if meet
Active NameNode crash when clean blocks, it will switch to new Active NameNode
(failover) to continue clean, if meet both NameNode are crash (HA mode), it
will be clean when re-full block report from DataNode, right? Thanks.
> FSNamesystem.delete() maybe cause data residue when active namenode crash or
> shutdown
> ---------------------------------------------------------------------------------------
>
> Key: HDFS-17048
> URL: https://issues.apache.org/jira/browse/HDFS-17048
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs
> Environment: hdfs3.3
> Reporter: liuguanghua
> Priority: Major
>
> Consider the following scenario：
> (1) User delete a hdfs dir with many blocks.
> (2) Then ative Namenode is crash or shutdown or failover to standby Namenode
> by administrator
> (3) This may result in residual data
>
> FSNamesystem.delete() will
> (1)delete dir first
> (2)add toRemovedBlocks into markedDeleteQueue.
> (3) MarkedDeleteBlockScrubber Thread will consumer the markedDeleteQueue and
> delete blocks.
> If the active namenode crash, the blocks in markedDeleteQueue will be lost
> and never be deleted. And the block cloud not find via hdfs fsck command. But
> it is alive in datanode disk.
>
> Thus ,
> SummaryA = hdfs dfs -du -s /
> SummaryB =sum( datanode report dfsused)
> SummaryA < SummaryB
>
> This may be unavoidable. But is there any way to find out the blocks that
> should be deleted and clean it ?
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
