[
https://issues.apache.org/jira/browse/HBASE-4168?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anirudh Todi updated HBASE-4168:
--------------------------------
Attachment: HBASE-4168(2).patch
@Ted - the patch I submitted is from the open-source trunk which I checked out
here - https://svn.apache.org/repos/asf/hbase/trunk/
I see your source of confusion now. In trunk's CatalogTracker, line 469 is:
{noformat}
} else if (cause != null && cause.getMessage() != null
{noformat}
the internal branch had:
{noformat}
} else if (cause.getMessage() != null)
{noformat}
and when I conducted Experiment-4 using the internal branch, cause turned out
to be null - and I received a NullPointerException at that line
However, would it still be better to return false and retry connecting to the
META instead of throwing an exception there?
I have uploaded a new patch in which I am handling the IOException unwrapped
from RemoteException in a similar manner.
> A client continues to try and connect to a powered down regionserver
> --------------------------------------------------------------------
>
> Key: HBASE-4168
> URL: https://issues.apache.org/jira/browse/HBASE-4168
> Project: HBase
> Issue Type: Bug
> Reporter: Anirudh Todi
> Assignee: Anirudh Todi
> Priority: Minor
> Attachments: HBASE-4168(2).patch, HBASE-4168-revised.patch,
> HBASE-4168.patch, hbase-hadoop-master-msgstore232.snc4.facebook.com.log
>
>
> Experiment-1
> Started a dev cluster - META is on the same regionserver as my key-value. I
> kill the regionserver process but donot power down the machine.
> The META is able to migrate to a new regionserver and the regions are also
> able to reopen elsewhere.
> The client is able to talk to the META and find the new kv location and get
> it.
> Experiment-2
> Started a dev cluster - META is on a different regionserver as my key-value.
> I kill the regionserver process but donot power down the machine.
> The META remains where it is and the regions are also able to reopen
> elsewhere.
> The client is able to talk to the META and find the new kv location and get
> it.
> Experiment-3
> Started a dev cluster - META is on a different regionserver as my key-value.
> I power down the machine hosting this regionserver.
> The META remains where it is and the regions are also able to reopen
> elsewhere.
> The client is able to talk to the META and find the new kv location and get
> it.
> Experiment-4 (This is the problematic one)
> Started a dev cluster - META is on the same regionserver as my key-value. I
> power down the machine hosting this regionserver.
> The META is able to migrate to a new regionserver - however - it takes a
> really long time (~30 minutes)
> The regions on that regionserver DONOT reopen (I waited for 1 hour)
> The client is able to find the new location of the META, however, the META
> keeps redirecting the client to powered down
> regionserver as the location of the key-value it is trying to get. Thus the
> client's get is unsuccessful.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
