[jira] [Comment Edited] (HDFS-13135) Lease not deleted when deleting INodeReference

2018-02-12 Thread Kihwal Lee (JIRA)

[ 
https://issues.apache.org/jira/browse/HDFS-13135?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16361517#comment-16361517
 ] 

Kihwal Lee edited comment on HDFS-13135 at 2/12/18 10:06 PM:
-

>From what I remember, a change (bug fix?) intentionally made a lease to be 
>removed from LeaseManager, but left it in the INode in the snapshot.  I argue 
>that it is not a correct design.  Some blocks can be left in 
>under-construction state forever without any block recovery. This can cause 
>data loss, since re-replication won't happen for those blocks.  When namenode 
>is restarted, all leases will be restored based on the files 
>under-construction section and also on INodeUnderConstructionFeature. 

I am not a snapshot expert, so I can't say what is the correct fix.  I've seen 
conflicting requirements on under-construction blocks in a snapshot. IMO, if an 
under-construction block ends up only in a snapshot (current view deleted while 
being written), a block recovery should somehow be done.


was (Author: kihwal):
>From what I remember, a change (bug fix?) intentionally made a lease to be 
>removed from LeaseManager, but left it in the INode in the snapshot.  I argue 
>that it is not a correct design.  Some blocks can be left in 
>under-construction state forever without any block recovery. This can cause 
>data loss, since re-replication won't happen for those blocks.  When namenode 
>is restarted, all leases will be restored based on the files 
>under-construction section and also on INodeUnderConstructionFeature. 

I am not a snapshot expert, so I can't say what is the correct fix.  I've seen 
conflicting requirements on under-construction blocks in a snapshot. IMO, an 
under-construction block ends up only in a snapshot (current view deleted while 
being written), a block recovery should somehow be done.

> Lease not deleted when deleting INodeReference
> --
>
> Key: HDFS-13135
> URL: https://issues.apache.org/jira/browse/HDFS-13135
> Project: Hadoop HDFS
>  Issue Type: Bug
>Reporter: Sean Mackrory
>Assignee: Sean Mackrory
>Priority: Major
> Attachments: HDFS-13135.001.patch
>
>
> In troubleshooting an occurrence of HDFS-13115, it seemed that there was 
> another underlying root cause that should also be addressed. There was an 
> INodeReference that was deleted and the lease on it was not subsequently 
> deleted because it was never added to the reclaim context.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

-
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org



[jira] [Comment Edited] (HDFS-13135) Lease not deleted when deleting INodeReference

2018-02-12 Thread Kihwal Lee (JIRA)

[ 
https://issues.apache.org/jira/browse/HDFS-13135?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16361517#comment-16361517
 ] 

Kihwal Lee edited comment on HDFS-13135 at 2/12/18 10:06 PM:
-

>From what I remember, a change (bug fix?) intentionally made a lease to be 
>removed from LeaseManager, but left it in the INode in the snapshot.  I argue 
>that it is not a correct design.  Some blocks can be left in 
>under-construction state forever without any block recovery. This can cause 
>data loss, since re-replication won't happen for those blocks.  When namenode 
>is restarted, all leases will be restored based on the files 
>under-construction section and also on INodeUnderConstructionFeature. 

I am not a snapshot expert, so I can't say what is the correct fix.  I've seen 
conflicting requirements on under-construction blocks in a snapshot. IMO, an 
under-construction block ends up only in a snapshot (current view deleted while 
being written), a block recovery should somehow be done.


was (Author: kihwal):
>From what I remember, a change (bug fix?) intentionally made a lease to be 
>removed from LeaseManager, but left it in the INode in the snapshot.  I argue 
>that it is not a correct design.  Some blocks can be left in 
>under-construction state forever without any block recovery. This can cause 
>data loss, since re-replication won't happen for those blocks.  When namenode 
>is restarted, all leases will be restored based on the files 
>under-construction section and also on INodeUnderConstructionFeature. 

> Lease not deleted when deleting INodeReference
> --
>
> Key: HDFS-13135
> URL: https://issues.apache.org/jira/browse/HDFS-13135
> Project: Hadoop HDFS
>  Issue Type: Bug
>Reporter: Sean Mackrory
>Assignee: Sean Mackrory
>Priority: Major
> Attachments: HDFS-13135.001.patch
>
>
> In troubleshooting an occurrence of HDFS-13115, it seemed that there was 
> another underlying root cause that should also be addressed. There was an 
> INodeReference that was deleted and the lease on it was not subsequently 
> deleted because it was never added to the reclaim context.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

-
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org