hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "HBase Review Board (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-2437) Refactor HLog splitLog
Date Wed, 26 May 2010 20:35:46 GMT

    [ https://issues.apache.org/jira/browse/HBASE-2437?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12871931#action_12871931

HBase Review Board commented on HBASE-2437:

Message from: "Todd Lipcon" <todd@cloudera.com>

Oh, it's much much improved! Thanks! I still think a "high level overview" would be good to

why .regionserver.? I'd say just hbase.hlog.split.batch.size or something

if this applies only to hlog splitting, maybe hbase.hlog.split.skip.errors

Which do you mean by "writers" here? I'd support factoring this function out into an HdfsUtil
class somewhere.

I agree if it has no records (I think - do we syncfs after writing the sequencefile header?).
But there's the case where inside SequenceFile we call create, but never actually write any
bytes. This is still worth recovering.

In general I think a corrupt tail means we should drop that record (incomplete written record)
but not shut down. This is only true if it's the tail record, though.

Can't find it now... does my above comment make sense?

This seems really voodoo.. if anything we're probably masking a real bug by doing this. Can
you write a unit test which shows this problem (even if it takes 30 minutes to run, would
be good to have in our arsenal)

The case where this happens is if you crash in the middle of appending a long edit. Consider
the case where a single edit might have 1MB of data (large rows). We can easily crash in the
middle of transferring it, before we call sync on the edit. In this case, the client never
received an "ack" for the write, so we can feel free to throw it away (this isn't data loss,
it's correct operation).

- Todd

This is an automatically generated e-mail. To reply, visit:

On 2010-05-21 12:11:59, stack wrote:

> Refactor HLog splitLog
> ----------------------
>                 Key: HBASE-2437
>                 URL: https://issues.apache.org/jira/browse/HBASE-2437
>             Project: HBase
>          Issue Type: Bug
>          Components: master
>    Affects Versions: 0.21.0
>            Reporter: Cosmin Lehene
>            Assignee: Cosmin Lehene
>             Fix For: 0.21.0
>         Attachments: 2437-v2.txt, 2437-v3.txt, 2437-v4.patch, 2437.txt, HBASE-2437_for_HBase-0.21_with_unit_tests_for_HDFS-0.21.patch
>   Original Estimate: 120h
>  Remaining Estimate: 120h
> the HLog.splitLog got really long and complex and hard to verify for correctness. 
> I started to refactor it and also ported changes from hbase-2337 that deals with premature
deletion of log files in case of errors. Further improvements will be possible, however the
scope of this issue is to clean the code and make it behave correctly (i.e. not lose any edits)
> Added a suite of unit tests that might be ported to 0.20 as well.
> Added a setting (hbase.skip.errors - feel free to suggest a better name) that, when set
to false will make the process less tolerant to failures or corrupted files:  in case a log
file is corrupted or an error stops the process from consistently splitting the log, will
abort the entire operation to avoid losing any edits. When hbase.skip.errors is on any corrupted
files will be partially parsed and then moved to the corrupted logs archive (see hbase-2337).

> Like hbase-2337 the splitLog method will first split all the logs and then proceed to
archive them. If any splitted log file (oldlogfile.log) that is the result of an earlier splitLog
attempt is found in the region directory, it will be deleted - this is safe since we won't
move the original log files until the splitLog process completes.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message