[ 
https://issues.apache.org/jira/browse/HDFS-6545?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14032962#comment-14032962
 ] 

Kihwal Lee edited comment on HDFS-6545 at 6/16/14 9:25 PM:
-----------------------------------------------------------

I think we can skip {{saveNamespace()}}. Presence of "finalize" op in a 
consumable edit segment should not cause any trouble, since any subsequent 
rolling upgrade will create a rollback image with txid past the op.  The only 
potential inconvenience is that if ANN or SBN needs to be restarted before the 
new version saving a checkpoint for some reason, it may need "-rollingUpgrade 
started" start-up option if there was a layout version change.  But this can 
also happen on SBN with the current code.    

In summary, removing {{saveNamespace()}} from {{finalizeRollingUpgrade()}} 
seems reasonable.



was (Author: kihwal):
I think we can skip {{saveNamespace()}}. Presence of "finalize" op in a 
consumable edit segment should not cause any trouble, since any subsequent 
rolling upgrade will create a rollback image with txid past the op.  The only 
potential inconvenience is that if ANN or SBN needs to be restarted before the 
new version saving a checkpoint for some reason, it may need "-rollingUpgrade 
started" start-up option if there was a layout version change.  But this can 
also happen on SBN with the current code.  If SBN was not up when ANN is 
finalizing an upgrade, SBN may need to be started with "-rollingUpgrade 
started" in order for it to read the rollback image saved by the old version. 
Or it will need to be bootstrapped fresh in order to download the new fsimage 
created only on ANN through {{saveNamespace()}}.  

In summary, removing {{saveNamespace()}} from {{finalizeRollingUpgrade()}} 
seems reasonable.


> Finalizing rolling upgrade can make NN unavailable for a long duration
> ----------------------------------------------------------------------
>
>                 Key: HDFS-6545
>                 URL: https://issues.apache.org/jira/browse/HDFS-6545
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>    Affects Versions: 2.4.0
>            Reporter: Kihwal Lee
>            Priority: Critical
>
> In {{FSNamesystem#finalizeRollingUpgrade()}}, {{saveNamespace()}} is directly 
> called.  For name nodes with a big name space, it can take minutes to save a 
> new fsimage.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to