Revision 559819

HADOOP-1516 HClient fails to readjust when ROOT or META redeployed on new region server

Detailed changes:

MiniHBaseCluster
- rewrite abortRegionServer, stopRegionServer - they now remove the
  server from the map of servers.
- rewrite waitOnRegionServer - now removes thread from map of threads

TestCleanRegionServerExit
- reduce Hadoop ipc client timeout and number of retries
- use rewritten stopRegionServer and waitOnRegionServer from MiniHBaseCluster
- add code to verify that failover worked
- moved testRegionServerAbort to separate test file

TestRegionServerAbort
- new test. Uses much the same code as TestCleanRegionServerExit but
  aborts the region server instead of shutting it down
  cleanly. Includes code to verify that failover worked.

hbase-site.xml (in src/contrib/hbase/src/test)
- reduce master lease timeout and time between lease timeout checks so
  that tests will run quicker.

HClient
- Major restructing of code that determines what region server to
  contact for a specific region. The main method findServersForTable
  is now recursive so that it will find the meta and root regions if
  they have not already been located or will re-find them if they have
  been reassigned and the old server can no longer be contacted.
- re-ordered administrative and general purpose methods so they are no
  longer located in seemingly random order.
- re-ordered code in ClientScanner.loadRegions so that if the location
  of the region changes, it will actually try to connect to the new
  server rather than continually trying to use the connection to the
  old server.

HLog
- use HashMap<Text, SequenceFile.Writer> instead of 
  TreeMap<Text, SequenceFile.Writer> because the TreeMap would return
  a value for a key it did not have (it was the value of another
  key). I have observed this before when the key is Text, but could
  not create a simple test case that reproduced the problem.
- added some new DEBUG level logging
- removed call to rollWriter() from closeAndDelete(). We don't need to
  start a new writer if we are closing the log.

HLogKey
- cleaned up per HADOOP-1466 (I initially modified it to add some
  debug logging which was later removed, but when I was making the
  modifications I took the opportunity to clean up the file)
- changed toString() format

HMaster
- better handling of RemoteException
- modified BaseScanner
 - now knows if it is scanning the root or a meta region
 - scanRegion no longer returns a value
 - if scanning the root region, it counts the number of meta regions
   it finds and sets a new AtomicInteger, numberOfMetaRegions when the
   scan is complete.
 - added abstract methods initialScan and maintenanceScan this allowed
   run method to be implemented in the base class.
- boolean rootScanned is now volatile
- modified RootScanner
 - moved actual scan into private method for readability (scanRoot)
 - implementation of abstract methods just call scanRoot
- add constructor for inner static class MetaRegion
- use a BlockingQueue to queue up work for the MetaScanner
- clean up handling of an unexpected region server exit
- PendingOperation.process now returns a boolean so that HMaster.run
  can determine if the operation completed or needs to be retried later
- PendingOperation processing no longer does a wait inside the process
  method since this might cause a deadlock if the current operation is
  waiting for another operation that has yet to be processed

HMsg
- removed MSG_REGIONSERVER_STOP_IN_ARRAY, MSG_NEW_REGION
- added MSG_REPORT_SPLIT

HRegionServer
- changed reportSplit to contain old region and new regions
- use IP from default interface rather than host name
- abort calls HLog.close() instead of HLog.rollWriter()