hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jean-Daniel Cryans (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-1439) race between master and regionserver after missed heartbeat
Date Tue, 07 Jul 2009 14:24:15 GMT

    [ https://issues.apache.org/jira/browse/HBASE-1439?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12728121#action_12728121

Jean-Daniel Cryans commented on HBASE-1439:

Since both the RSs and the Master (with HBASE-1575) will handle session expiration, the possibility
that a region server deletes its hlogs after sleeping for too long is greatly lessened as
it will just abort and restart. The other possibility, as I described in the previous post,
is that the region server cannot reach any ZK server for any reason and gets detected as dead
by the Master which can still reach a ZK server. When I first tested that, it seemed like
a very serious issue but now that I gave it some thought I think that it if happened HDFS
would probably be wedged too by such a network partition. 

Anyways, the fix would be to have a timer in a generic Watch specialization that catches all
disconnections and that after ~tickTime*1.5 would send a fabricated session expiration. IMO
this is a nice to have for 0.20.0. Should we punt this?

> race between master and regionserver after missed heartbeat
> -----------------------------------------------------------
>                 Key: HBASE-1439
>                 URL: https://issues.apache.org/jira/browse/HBASE-1439
>             Project: Hadoop HBase
>          Issue Type: Bug
>    Affects Versions: 0.19.1
>         Environment: CentOS 5.2 x86_64, HBase 0.19.1, Hadoop 0.19.1
>            Reporter: Andrew Purtell
>            Assignee: Jean-Daniel Cryans
>            Priority: Blocker
>             Fix For: 0.20.0
> Seen on one of our 0.19.1 clusters:
> {code}
> java.io.FileNotFoundException: File does not exist: hdfs://jdc2-atr-dc-2.atr.trendmicro.com:50000
> /data/hbase/log_10.3.134.207_1242286427894_60020/hlog.dat.1242528291898
>  at org.apache.hadoop.hdfs.DistributedFileSystem.getFileStatus(DistributedFileSystem.java:415)
>  at org.apache.hadoop.fs.FileSystem.getLength(FileSystem.java:679)
>  at org.apache.hadoop.hbase.io.SequenceFile$Reader.<init>(SequenceFile.java:1431)
>  at org.apache.hadoop.hbase.io.SequenceFile$Reader.<init>(SequenceFile.java:1426)
>  at org.apache.hadoop.hbase.regionserver.HLog.splitLog(HLog.java:753)
>  at org.apache.hadoop.hbase.regionserver.HLog.splitLog(HLog.java:716)
>  at org.apache.hadoop.hbase.master.ProcessServerShutdown.process(ProcessServerShutdown.java:249)
>  at org.apache.hadoop.hbase.master.HMaster.processToDoQueue(HMaster.java:442)
>  at org.apache.hadoop.hbase.master.HMaster.run(HMaster.java:377)
> 2009-05-17 04:05:55,481 INFO org.apache.hadoop.hbase.master.RegionServerOperation: process
> shutdown of server logSplit: false, rootRescanned: false, numberOfMetaRegions:
> onlineMetaRegions.size(): 1
> {code}
> I do not have the region server log yet, but here is my conjecture:
> Here, the write ahead log for is missing in DFS: java.io.FileNotFoundException:
when the master tries to split it after declaring the region server crashed. There have been
recent trouble reports on this cluster that indicate severe memory stress, e.g. kernel panics
due to OOM. Based on that I think it is likely that the region server here missed a heartbeat
so the master declared it crashed and began to split the log. But, the log was then deleted
out from underneath the master's split thread. I think the region server was actually still
running but partially swapped out or the node was otherwise overloaded so it missed its heartbeat.
Then, when the region server came back after being swapped in, it realized it missed its heartbeat
and shut down, deleting the log as is normally done. 
> Even if the above is not the actual cause in this case, I think the scenario is plausible.
What do you think?

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

View raw message