[
https://issues.apache.org/jira/browse/HBASE-10210?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13857109#comment-13857109
]
Enis Soztutar commented on HBASE-10210:
---------------------------------------
It seems that we still have two different mechanisms to track online region
servers. RegionServerTracker via zk, and ServerManager via regionServerStartup
+ regionServerReport.
What if we do not accept a region server report unless we processed the zk
notification. Would that work? We can ignore the region server report until we
get the zk notification.
We can stop registering the RS on region server startup or region server
reports. Only RSTracker can register new region servers. ServerManager can use
the same list from RST?
> during master startup, RS can be you-are-dead-ed by master in error
> -------------------------------------------------------------------
>
> Key: HBASE-10210
> URL: https://issues.apache.org/jira/browse/HBASE-10210
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.98.0, 0.96.1, 0.99.0, 0.96.1.1
> Reporter: Sergey Shelukhin
> Assignee: Sergey Shelukhin
> Attachments: HBASE-10210.patch
>
>
> Not sure of the root cause yet, I am at "how did this ever work" stage.
> We see this problem in 0.96.1, but didn't in 0.96.0 + some patches.
> It looks like RS information arriving from 2 sources - ZK and server itself,
> can conflict. Master doesn't handle such cases (timestamp match), and anyway
> technically timestamps can collide for two separate servers.
> So, master YouAreDead-s the already-recorded reporting RS, and adds it too.
> Then it discovers that the new server has died with fatal error!
> Note the threads.
> Addition is called from master initialization and from RPC.
> {noformat}
> 2013-12-19 11:16:45,290 INFO
> [master:h2-ubuntu12-sec-1387431063-hbase-10:60000] master.ServerManager:
> Finished waiting for region servers count to settle; checked in 2, slept for
> 18262 ms, expecting minimum of 1, maximum of 2147483647, master is running.
> 2013-12-19 11:16:45,290 INFO
> [master:h2-ubuntu12-sec-1387431063-hbase-10:60000] master.ServerManager:
> Registering
> server=h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800
> 2013-12-19 11:16:45,290 INFO
> [master:h2-ubuntu12-sec-1387431063-hbase-10:60000] master.HMaster: Registered
> server found up in zk but who has not yet reported in:
> h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800
> 2013-12-19 11:16:45,380 INFO [RpcServer.handler=4,port=60000]
> master.ServerManager: Triggering server recovery; existingServer
> h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800
> looks stale, new
> server:h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800
> 2013-12-19 11:16:45,380 INFO [RpcServer.handler=4,port=60000]
> master.ServerManager: Master doesn't enable ServerShutdownHandler during
> initialization, delay expiring server
> h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800
> ...
> 2013-12-19 11:16:46,925 ERROR [RpcServer.handler=7,port=60000]
> master.HMaster: Region server
> h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800
> reported a fatal error:
> ABORTING region server
> h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800:
> org.apache.hadoop.hbase.YouAreDeadException: Server REPORT rejected;
> currently processing
> h2-ubuntu12-sec-1387431063-hbase-8.cs1cloud.internal,60020,1387451803800 as
> dead server
> {noformat}
> Presumably some of the recent ZK listener related changes b
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
