[
https://issues.apache.org/jira/browse/HBASE-3380?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12973945#action_12973945
]
Jonathan Gray commented on HBASE-3380:
--------------------------------------
bq. I don't mind the changes as it helps on cold starts, but I think that in
the failover case calling regionServerTracker.getOnlineServers().size() would
be simple and reliable.
I don't think it's necessarily simple and I would prefer not to do halfway
fixes where we are still actually relying on heartbeats.
The simplest implementation of this fix would create a new bug where the RS has
actually died but znode is not gone yet. So master will wait until it reaches
N regionservers but will only ever get heartbeats from N-1. So then we have a
new maxTimeout parameter to deal with this case (and now you're basically where
we are with my patch but without added benefits of the new configs).
I think these timeout changes are all that is necessary to fix both cases until
we remove heartbeats in 0.92.
> Master failover can split logs of live servers
> ----------------------------------------------
>
> Key: HBASE-3380
> URL: https://issues.apache.org/jira/browse/HBASE-3380
> Project: HBase
> Issue Type: Bug
> Reporter: Jean-Daniel Cryans
> Priority: Blocker
> Fix For: 0.90.0
>
> Attachments: HBASE-3380-v1.patch
>
>
> The reason why TestMasterFailover fails is that when it does the master
> failover, the new master doesn't wait long enough for all region servers to
> checkin so it goes ahead and split logs... which doesn't work because of the
> way lease timeouts work:
> {noformat}
> 2010-12-21 07:30:36,977 DEBUG [Master:0;vesta.apache.org:33170]
> wal.HLogSplitter(256): Splitting hlog 1 of 1:
>
> hdfs://localhost:49187/user/hudson/.logs/vesta.apache.org,38743,1292916616340/vesta.apache.org%3A38743.1292916617204,
> length=0
> 2010-12-21 07:30:36,977 DEBUG [WriterThread-1]
> wal.HLogSplitter$WriterThread(619): Writer thread
> Thread[WriterThread-1,5,main]: starting
> 2010-12-21 07:30:36,977 DEBUG [WriterThread-2]
> wal.HLogSplitter$WriterThread(619): Writer thread
> Thread[WriterThread-2,5,main]: starting
> 2010-12-21 07:30:36,977 INFO [Master:0;vesta.apache.org:33170]
> util.FSUtils(625): Recovering file
>
> hdfs://localhost:49187/user/hudson/.logs/vesta.apache.org,38743,1292916616340/vesta.apache.org%3A38743.1292916617204
> 2010-12-21 07:30:36,979 WARN [IPC Server handler 8 on 49187]
> namenode.FSNamesystem(1122): DIR* NameSystem.startFile:
> failed to create file
> /user/hudson/.logs/vesta.apache.org,38743,1292916616340/vesta.apache.org%3A38743.1292916617204
> for
> DFSClient_hb_m_vesta.apache.org:33170_1292916630791 on client 127.0.0.1,
> because this file is already being created by
> DFSClient_hb_rs_vesta.apache.org,38743,1292916616340_1292916617166 on
> 127.0.0.1
> ...
> 2010-12-21 07:33:44,332 WARN [Master:0;vesta.apache.org:33170]
> util.FSUtils(644): Waited 187354ms for lease recovery on
>
> hdfs://localhost:49187/user/hudson/.logs/vesta.apache.org,38743,1292916616340/vesta.apache.org%3A38743.1292916617204:
> org.apache.hadoop.hdfs.protocol.AlreadyBeingCreatedException: failed to
> create file
>
> /user/hudson/.logs/vesta.apache.org,38743,1292916616340/vesta.apache.org%3A38743.1292916617204
> for DFSClient_hb_m_vesta.apache.org:33170_1292916630791 on client 127.0.0.1,
> because this file is already
> being created by
> DFSClient_hb_rs_vesta.apache.org,38743,1292916616340_1292916617166 on
> 127.0.0.1
> {noformat}
> I think that we should always check in ZK the number of live region servers
> before waiting for them to check in, this way we know how many we should
> expect during failover. There's also a case where we still want to timeout,
> since RS can die during that time, but we should wait a bit longer.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
