hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andrew Purtell (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (HBASE-10829) Flush is skipped after log replay if the last recovered edits file is skipped
Date Wed, 26 Mar 2014 18:31:17 GMT

     [ https://issues.apache.org/jira/browse/HBASE-10829?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Andrew Purtell updated HBASE-10829:
-----------------------------------

    Fix Version/s:     (was: 0.98.2)
                   0.98.1

> Flush is skipped after log replay if the last recovered edits file is skipped
> -----------------------------------------------------------------------------
>
>                 Key: HBASE-10829
>                 URL: https://issues.apache.org/jira/browse/HBASE-10829
>             Project: HBase
>          Issue Type: Bug
>            Reporter: Enis Soztutar
>            Assignee: Enis Soztutar
>            Priority: Critical
>             Fix For: 0.98.1, 0.99.0, 0.96.3
>
>         Attachments: hbase-10829_v1.patch, hbase-10829_v2.patch, hbase-10829_v3.patch
>
>
> We caught this in an extended test run where IntegrationTestBigLinkedList failed with
some missing keys. 
> The problem is that HRegion.replayRecoveredEdits() would return -1 if all the edits in
the log file is skipped, which is true for example if the log file only contains a single
compaction record (HBASE-2231) or somehow the edits cannot be applied (column family deleted,
etc). 
> The callee, HRegion.replayRecoveredEditsIfAny() only looks for the last returned seqId
to decide whether a flush is necessary or not before opening the region, and discarding replayed
recovered edits files. 
> Therefore, if the last recovered edits file is skipped but some edits from earlier recovered
edits files are applied, the mandatory flush before opening the region is skipped. If the
region server dies after this point before a flush, the edits are lost. 
> This is important to fix, though the sequence of events are super rare for a production
cluster. 



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

Mime
View raw message