Affects Version/s: None
Fix Version/s: 0.92.0
The heartbeat in hbase is about to go away. See hbase-1502. This issue is about moving the report of a split region off the heartbeat channel. This work needs to be done before I can finish up hbase-1502.
Here is a provisional 'design' for how report of split to master will now work.
Instead of riding a split message on the regionserver to master heartbeat, the regionserver will put up a znode under (the badly named) 'unassigned' directory; i.e. the regionserver will put the region into 'transition' (RIT). Usually the master does all RIT machinations. Splits will be an exception.
The regionserver will move the znode through two states: SPLITTING and SPLIT. The master will clean up the SPLIT znode on receipt of child changed callback.
There will be no extra data in the SPLITTING znode beyond the SPLITTING state name. The SPLIT znode will contain the name of the split daughters.
Any region in RIT will block the balancer running. This could be an issue if lots of splits going on across a cluster.
Master will need to handle new SPLIT and SPLITTING states and that another may have already written the RIT (unit tests).
As before, regionserver figures region to split and runs CompactSplitThread#split. As before, this invokes SplitTransaction with its prepare, execute and when necessary rollback phase. Into the execute we'll add the setting of SPLITTING znode into RIT. The execution will end with setting the SPLIT znode state. Master will read and clear the SPLIT state when done.