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-2312) Possible data loss when RS goes into GC pause while rolling HLog
Date Wed, 28 Jul 2010 21:22:18 GMT

    [ https://issues.apache.org/jira/browse/HBASE-2312?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12893371#action_12893371
] 

HBase Review Board commented on HBASE-2312:
-------------------------------------------

Message from: "Nicolas" <nspiegelberg@facebook.com>


bq.  On 2010-07-28 10:17:14, Todd Lipcon wrote:
bq.  > This looks good, except that it requires the HDFS patches to be in place to even
compile. Although I think it's OK to not fix the bug in the case that we're on stock HDFS,
I think we still need to be able to run, buggily.
bq.  > 
bq.  > Did we determine that all of the other solutions were flawed/too complicated?

HBASE-2312 had some lengthy discussions that ultimately led to this route.  For 0.21 + 0.22,
only HADOOP-6840 is needed.  That is a very minor change consisting of only small API additions
(so, no regression needed) that Dhruba & I think the lack of risk will make it easy to
approve.  0.20-append needs a couple more JIRAs, but that should be even less flack.  As long
as we ship with 0.20-append HDFS or newer, we'll be fine.

I understand that this diff is early and we will have to wait until we ship with the 0.20-append
JAR before application.  Basically, I also wanted to show that the API change indeed fixes
our issue painlessly.  Any +1s on HADOOP-6840 would be appreciated :)  Your comment is that
we should also backward-support 0.20.3?


bq.  On 2010-07-28 10:17:14, Todd Lipcon wrote:
bq.  > trunk/src/main/java/org/apache/hadoop/hbase/master/HMaster.java, line 650
bq.  > <http://review.cloudera.org/r/396/diff/1/?file=3523#file3523line650>
bq.  >
bq.  >     Check return value of rename (this is one of those stupid APIs that returns
false instead of throwing)

will do


bq.  On 2010-07-28 10:17:14, Todd Lipcon wrote:
bq.  > trunk/src/main/java/org/apache/hadoop/hbase/master/HMaster.java, line 651
bq.  > <http://review.cloudera.org/r/396/diff/1/?file=3523#file3523line651>
bq.  >
bq.  >     rather than reassigning, just pass splitDir below?

there are 2 cases:
#1 : the master hasn't already tried to split, hits this code
#2 : master tries to split, dies while splitting, backup master appears

In case #2, you don't need to rename the directory because an region cannot run on a directory
of that name.  Therefore, the 'logDir' is correct from the get-go and doesn't need a rename.
 So, you need to either have duplicate splitLog code or update 'logDir'


bq.  On 2010-07-28 10:17:14, Todd Lipcon wrote:
bq.  > trunk/src/main/java/org/apache/hadoop/hbase/master/ProcessServerShutdown.java, line
297
bq.  > <http://review.cloudera.org/r/396/diff/1/?file=3524#file3524line297>
bq.  >
bq.  >     extract "-splitting" out to a constant?
bq.  >     Or can we move some of this common code into HLog?

sure.  will do


bq.  On 2010-07-28 10:17:14, Todd Lipcon wrote:
bq.  > trunk/src/main/java/org/apache/hadoop/hbase/master/ProcessServerShutdown.java, line
300
bq.  > <http://review.cloudera.org/r/396/diff/1/?file=3524#file3524line300>
bq.  >
bq.  >     maybe:
bq.  >     assert !(logDirExists && splitDirExists) : "Both " + rsLogDir + " and
" + rsSplitDir + " exist";
bq.  >     so if the assertion failure happens it's more understandable

will do


- Nicolas


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
http://review.cloudera.org/r/396/#review513
-----------------------------------------------------------





> Possible data loss when RS goes into GC pause while rolling HLog
> ----------------------------------------------------------------
>
>                 Key: HBASE-2312
>                 URL: https://issues.apache.org/jira/browse/HBASE-2312
>             Project: HBase
>          Issue Type: Bug
>          Components: master, regionserver
>    Affects Versions: 0.20.3
>            Reporter: Karthik Ranganathan
>            Assignee: Nicolas Spiegelberg
>
> There is a very corner case when bad things could happen(ie data loss):
> 1)	RS #1 is going to roll its HLog - not yet created the new one, old one will get no
more writes
> 2)	RS #1 enters GC Pause of Death
> 3)	Master lists HLog files of RS#1 that is has to split as RS#1 is dead, starts splitting
> 4)	RS #1 wakes up, created the new HLog (previous one was rolled) and appends an edit
- which is lost
> The following seems like a possible solution:
> 1)	Master detects RS#1 is dead
> 2)	The master renames the /hbase/.logs/<regionserver name>  directory to something
else (say /hbase/.logs/<regionserver name>-dead)
> 3)	Add mkdir support (as opposed to mkdirs) to HDFS - so that a file create fails if
the directory doesn't exist. Dhruba tells me this is very doable.
> 4)	RS#1 comes back up and is not able create the new hlog. It restarts itself.

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


Mime
View raw message