[
https://issues.apache.org/jira/browse/HDFS-5723?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13880721#comment-13880721
]
Vinay commented on HDFS-5723:
-----------------------------
bq. Hi Vinay, one question about the patch: so this inconsistent generation
stamp can also be caused by a delayed block-received report? I.e., after the
first close(), the DN's report gets delayed and is received by NN when the
append starts. In that case, will we have any issue by wrongly putting the
(block, DN) into the corruptBlockMap?
It can happen. Earlier block with prev genstamp will be marked corrupt, if
append pipeline setup with genstamp update happened before previous report
comes.
But append pipeline creation will update the genstamp in that datanode also, so
one more block-received report is expectedn with correct genstamp. This time it
will remove the old block from corrupt replica map.
> Append failed FINALIZED replica should not be accepted as valid when that
> block is underconstruction
> ----------------------------------------------------------------------------------------------------
>
> Key: HDFS-5723
> URL: https://issues.apache.org/jira/browse/HDFS-5723
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: namenode
> Affects Versions: 2.2.0
> Reporter: Vinay
> Assignee: Vinay
> Attachments: HDFS-5723.patch
>
>
> Scenario:
> 1. 3 node cluster with
> dfs.client.block.write.replace-datanode-on-failure.enable set to false.
> 2. One file is written with 3 replicas, blk_id_gs1
> 3. One of the datanode DN1 is down.
> 4. File was opened with append and some more data is added to the file and
> synced. (to only 2 live nodes DN2 and DN3)-- blk_id_gs2
> 5. Now DN1 restarted
> 6. In this block report, DN1 reported FINALIZED block blk_id_gs1, this should
> be marked corrupted.
> but since NN having appended block state as UnderConstruction, at this time
> its not detecting this block as corrupt and adding to valid block locations.
> As long as the namenode is alive, this datanode also will be considered as
> valid replica and read/append will fail in that datanode.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
