hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jim Kellerman (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-1314) master sees HRS znode expire and splits log while the HRS is still running and accepting edits
Date Fri, 08 May 2009 23:30:45 GMT

    [ https://issues.apache.org/jira/browse/HBASE-1314?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12707556#action_12707556

Jim Kellerman commented on HBASE-1314:

Or increase the timeout interval. In 0.19, the master waits a full 2 minutes before declaring
a region server to be dead (with a comment that the lease timeout interval may have to be
increased on heavily loaded clusters (but this is due to lease renewal piggybacking on regionServerReport)).

I'm not really up-to-date on how the ZK stuff works, but it seems to me that renewing the
lease should happen in a separate thread with a small sleep interval to keep other work from
interfering with lease renewal.

How does the region server renew its lease anyway?

> master sees HRS znode expire and splits log while the HRS is still running and accepting
> ----------------------------------------------------------------------------------------------
>                 Key: HBASE-1314
>                 URL: https://issues.apache.org/jira/browse/HBASE-1314
>             Project: Hadoop HBase
>          Issue Type: Bug
>    Affects Versions: 0.20.0
>            Reporter: Andrew Purtell
>             Fix For: 0.20.0
> ZK session expiration related problem. HRS loses its ephemeral node while it is still
up and running and accepting edits. Master sees it go away and starts splitting its logs while
edits are still being written. After this, all reconstruction logs have to be manually removed
from the region directories or the regions will never deploy (CRC errors). I think on HDFS
edits would be lost, not corrupted. (I am using a HBase root on local file system.) 
> HRS ZK session expires, causing its znode to go away:
> 2009-04-07 03:50:39,953 INFO org.apache.hadoop.hbase.master.ServerManager: localhost.localdomain_1239068648333_60020
znode expired
> 2009-04-07 03:50:40,565 DEBUG org.apache.hadoop.hbase.master.HMaster: Processing todo:
ProcessServerShutdown of localhost.localdomain_1239068648333_60020
> 2009-04-07 03:50:40,637 INFO org.apache.hadoop.hbase.master.RegionServerOperation: process
shutdown of server localhost.localdomain_1239068648333_60020: logSplit: false, rootRescanned:
false, numberOfMetaRegions: 1, onlineMetaRegions.size(): 1
> But here we have the HRS still reporting in, triggering a LeaseStillHeldException:
> 2009-04-07 03:50:40,826 INFO org.apache.hadoop.ipc.HBaseServer: IPC Server handler 2
on 60000, call regionServerReport(address:, startcode: 1239068648333, load:
(requests=14, regions=7, usedHeap=582, maxHeap=888), [Lorg.apache.hadoop.hbase.HMsg;@6da21389,
[Lorg.apache.hadoop.hbase.HRegionInfo;@2bb0bf9a) from error: org.apache.hadoop.hbase.Leases$LeaseStillHeldException
> org.apache.hadoop.hbase.Leases$LeaseStillHeldException
>         at org.apache.hadoop.hbase.master.ServerManager.regionServerReport(ServerManager.java:198)
>         at org.apache.hadoop.hbase.master.HMaster.regionServerReport(HMaster.java:601)
>         at sun.reflect.GeneratedMethodAccessor2.invoke(Unknown Source)
>         at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:25)
>         at java.lang.reflect.Method.invoke(Method.java:597)
>         at org.apache.hadoop.hbase.ipc.HBaseRPC$Server.call(HBaseRPC.java:632)
>         at org.apache.hadoop.hbase.ipc.HBaseServer$Handler.run(HBaseServer.java:909)
> And log splitting starts anyway:
> 2009-04-07 03:50:41,139 INFO org.apache.hadoop.hbase.regionserver.HLog: Splitting 3 log(s)
in file:/data/hbase/log_localhost.localdomain_1239068648333_60020
> 2009-04-07 03:50:41,139 DEBUG org.apache.hadoop.hbase.regionserver.HLog: Splitting 1
of 3: file:/data/hbase/log_localhost.localdomain_1239068648333_60020/hlog.dat.1239075060711
> [...]

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

View raw message