Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
Description
When cluster replication enabled and one RegionServer crashed, its WALs will be move from WALs dir to oldWALs dir and its replication queue will moved to other RegionServer's replication queue.
HDFS layout (WAL Storage)
/hbase/WALs/RS1/1.log
/hbase/WALs/RS1/2.log
/hbase/WALs/RS1/3.log
ZooKeeper layout (Replication queue storage)
/hbase/replication/rs/RS1/peerId/1.log
/hbase/replication/rs/RS1/peerId/2.log
/hbase/replication/rs/RS1/peerId/3.log
Failover finished:
HDFS layout (WAL Storage)
/hbase/.oldWALs/1.log
/hbase/.oldWALs/2.log
/hbase/.oldWALs/3.log
ZooKeeper layout (Replication queue storage)
/hbase/replication/rs/RS2/peerId-RS1/1.log
/hbase/replication/rs/RS2/peerId-RS1/2.log
/hbase/replication/rs/RS2/peerId-RS1/3.log
And if enabled hbase.separate.oldlogdir.by.regionserver, the HDFS layout may be:
HDFS layout (WAL Storage)
/hbase/.oldWALs/RS1/1.log
/hbase/.oldWALs/RS1/2.log
/hbase/.oldWALs/RS1/3.log
Then if RS2 crashed, the HDFS layout will not change but ZooKeeper layout may changed.
ZooKeeper layout (Replication queue storage)
/hbase/replication/rs/RS3/peerId-RS1-RS2/1.log
/hbase/replication/rs/RS3/peerId-RS1-RS2/2.log
/hbase/replication/rs/RS3/peerId-RS1-RS2/3.log
So even the replication queue was transfered many times, the HDFS layout never change.
Another case is master-cluster disaster, the failover work not finished. Then ReplicationSyncUp tool can start replication source to replicate the WAL data. The HDFS layout need to consider two more case:
/hbase/WALs/RS1/1.log
/hbase/WALs/RS1/2.log
/hbase/WALs/RS1/3.log
or
/hbase/WALs/RS1-splitting/1.log
/hbase/WALs/RS1-splitting/2.log
/hbase/WALs/RS1-splitting/3.log