Region would be assigned twice easily with continually killing server and
moving region in testing environment
---------------------------------------------------------------------------------------------------------------
Key: HBASE-4899
URL: https://issues.apache.org/jira/browse/HBASE-4899
Project: HBase
Issue Type: Bug
Reporter: chunhui shen
Before assigning region in ServerShutdownHandler#process, it will check whether
region is in RIT,
however, this checking doesn't work as the excepted in the following case:
1.move region A from server B to server C
2.kill server B
3.start server B immediately
Let's see what happen in the code for the above case
{code}
for step1:
1.1 server B close the region A,
1.2 master setOffline for region
A,(AssignmentManager#setOffline:this.regions.remove(regionInfo))
1.3 server C start to open region A.(Not completed)
for step3:
master ServerShutdownHandler#process() for server B
{
..
splitlog()
...
List<RegionState> regionsInTransition =
this.services.getAssignmentManager()
.processServerShutdown(this.serverName);
...
Skip regions that were in transition unless CLOSING or PENDING_CLOSE
...
assign region
}
In fact, when running
ServerShutdownHandler#process()#this.services.getAssignmentManager().processServerShutdown(this.serverName),
region A is in RIT (step1.3 not completed), but the return List<RegionState>
regionsInTransition doesn't contain it, because region A has removed from
AssignmentManager.regions by AssignmentManager#setOffline in step 1.2
Therefore, region A will be assigned twice.
{code}
Actually, one server killed and started twice will also easily cause region
assigned twice.
Exclude the above reason, another probability :
when execute ServerShutdownHandler#process()#MetaReader.getServerUserRegions
,region is included which is in RIT now.
But after completing MetaReader.getServerUserRegions, the region has been
opened in other server and is not in RIT now.
In our testing environment where balancing,moving and killing are executed
periodly, assigning region twice often happens, and it is hateful because it
will affect other test cases.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
