Affects Version/s: None
Fix Version/s: 0.99.2
Hadoop Flags:Incompatible change, Reviewed
Release Note:This fix includes our enabling useMulti flag as default. multi is a zk method only available in later versions of zookeeper. This change means HBase 1.0 requires a zookeeper that is at least version 3.4+. See
When a regionserver crash, another regionserver will try to take over the replication hlogs queue and help the the the dead regionserver to finish the replcation.See NodeFailoverWorker in ReplicationSourceManager
Currently hbase.zookeeper.useMulti is false in default configuration. The operation of taking over replication queue is not atomic. The ReplicationSourceManager firstly lock the replication node of dead regionserver and then copy the replication queue, and delete replication node of dead regionserver at last. The operation of the lockOtherRS just creates a persistent zk node named "lock" which prevent other regionserver taking over the replication queue.
But if a regionserver crashed after creating this "lock" zk node and before coping the replication queue to its replication queue, the "lock" zk node will be left forever and
no other regionserver can take over the replication queue.
In out production cluster, we encounter this problem. We found the replication queue was there and no regionserver took over it and a "lock" zk node left there.
hbase.32561.log:2014-09-24,14:09:28,790 INFO org.apache.hadoop.hbase.replication.ReplicationZookeeper: Won't transfer the queue, another RS took care of it because of: KeeperErrorCode = NoNode for /hbase/hhsrv-micloud/replication/rs/hh-hadoop-srv-st09.bj,12610,1410937824255/lock
hbase.32561.log:2014-09-24,14:14:45,148 INFO org.apache.hadoop.hbase.replication.ReplicationZookeeper: Won't transfer the queue, another RS took care of it because of: KeeperErrorCode = NoNode for /hbase/hhsrv-micloud/replication/rs/hh-hadoop-srv-st10.bj,12600,1410937795685/lock
A quick solution is that the lock operation just create an ephemeral "lock" zookeeper node and when the lock node is deleted, other regionserver will be notified to check if there are replication queue left.
Suggestions are welcomed! Thanks.