accumulo-notifications mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "John Vines (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (ACCUMULO-1759) Empty walogs block recovery after power outage.
Date Wed, 09 Oct 2013 20:16:42 GMT

    [ https://issues.apache.org/jira/browse/ACCUMULO-1759?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13790804#comment-13790804
] 

John Vines commented on ACCUMULO-1759:
--------------------------------------

We catted several of the files and they were empty.

Unfortunately the circumstances around this are hard to reproduce, so we haven't been able
to recreate it sans patch. [~supermallen] believes that his patch shouldn't interfere with
when/how the walogs are created though.

> Empty walogs block recovery after power outage.
> -----------------------------------------------
>
>                 Key: ACCUMULO-1759
>                 URL: https://issues.apache.org/jira/browse/ACCUMULO-1759
>             Project: Accumulo
>          Issue Type: Bug
>    Affects Versions: 1.5.0
>         Environment: * HDP 1.3
> ** {{dfs.durable.sync=true}}
> ** {{dfs.datanode.synconclose=true}}
> * encrytion patch from ACCUMULO-998
>            Reporter: Luke Brassard
>
> Power was abruptly cut to the cluster. Upon restart of HDFS, there was a single Rfile
that was missing a block. 
> Here are some details about the cluster:
> * HDP 1.3
> ** {{dfs.durable.sync=true}}
> ** {{dfs.datanode.synconclose=true}}
> * encrytion patch from ACCUMULO-998
> After restarting Accumulo, the Master was complaining with a series of these:
> {code}
> 2013-10-09 18:15:16,649 [recovery.HadoopLogCloser] INFO : Waiting for file to be closed
/accumulo/wal/10.10.0.1+9997/d52ab315-5ac1-4a5c-9085-67ae29b98b88
> 2013-10-09 18:15:16,663 [recovery.HadoopLogCloser] INFO : Waiting for file to be closed
/accumulo/wal/10.10.0.2+9997/d0192739-74e2-43a0-985f-3ed668259995
> 2013-10-09 18:15:16,742 [recovery.HadoopLogCloser] INFO : Waiting for file to be closed
/accumulo/wal/10.10.0.3+9997/de54e6dc-964a-4b33-b4fb-052e81749913
> 2013-10-09 18:15:16,833 [recovery.HadoopLogCloser] INFO : Waiting for file to be closed
/accumulo/wal/10.10.0.4+9997/cda5daec-25f3-443b-818a-990d3eddd56f
> {code}
> Inspection of the files above showed that they were all empty, but referenced in the
{{!METADATA}} table. The solution was to move or remove the files from HDFS and delete the
references from the metadata. The instance was then able to stabilize and assign the rest
of the tablets.
> It is unclear why these empty walogs existed in the first place. Is it possible that
there should have been data in these walogs? Or should the files have been disregarded since
they were empty?
> Regarding the Rfile that was missing the block, it was not being referenced by the {{!METADATA}}
table, so removing it had no negative effect on the instance. Should there have been some
reference to this file?



--
This message was sent by Atlassian JIRA
(v6.1#6144)

Mime
View raw message