org.apache.hadoop.hbase.regionserver

Class SplitTransaction

java.lang.Object

org.apache.hadoop.hbase.regionserver.SplitTransaction

@InterfaceAudience.Private
public class SplitTransaction
extends Object

Executes region split as a "transaction". Call prepare() to setup the transaction, execute(Server, RegionServerServices) to run the transaction and rollback(Server, RegionServerServices) to cleanup if execute fails.

Here is an example of how you would use this class:

  SplitTransaction st = new SplitTransaction(this.conf, parent, midKey)
  if (!st.prepare()) return;
  try {
    st.execute(server, services);
  } catch (IOException ioe) {
    try {
      st.rollback(server, services);
      return;
    } catch (RuntimeException e) {
      myAbortable.abort("Failed split, abort");
    }
  }
 

This class is not thread safe. Caller needs ensure split is run by one thread only.

Constructor Summary

Constructors

Constructor and Description
SplitTransaction(HRegion r, byte[] splitrow) Constructor

Method Summary

Methods

Modifier and Type	Method and Description
Put	addLocation(Put p, ServerName sn, long openSeqNum)
static void	createNodeSplitting(ZooKeeperWatcher zkw, HRegionInfo region, ServerName serverName, HRegionInfo a, HRegionInfo b) Creates a new ephemeral node in the PENDING_SPLIT state for the specified region.
PairOfSameType<HRegion>	execute(Server server, RegionServerServices services) Run the transaction.
boolean	prepare() Does checks on split inputs.
boolean	rollback(Server server, RegionServerServices services)
PairOfSameType<HRegion>	stepsAfterPONR(Server server, RegionServerServices services, PairOfSameType<HRegion> regions)
PairOfSameType<HRegion>	stepsBeforePONR(Server server, RegionServerServices services, boolean testing)
static int	transitionSplittingNode(ZooKeeperWatcher zkw, HRegionInfo parent, HRegionInfo a, HRegionInfo b, ServerName serverName, int znodeVersion, EventType beginState, EventType endState) Transitions an existing ephemeral node for the specified region which is currently in the begin state to be in the end state.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

SplitTransaction

public SplitTransaction(HRegion r,
                byte[] splitrow)

Constructor

Parameters:

r - Region to split

splitrow - Row to split around

Method Detail

prepare

public boolean prepare()

Does checks on split inputs.

Returns:

true if the region is splittable else false if it is not (e.g. its already closed, etc.).

stepsBeforePONR

public PairOfSameType<HRegion> stepsBeforePONR(Server server,
                                      RegionServerServices services,
                                      boolean testing)
                                        throws IOException

Throws:

IOException

execute

public PairOfSameType<HRegion> execute(Server server,
                              RegionServerServices services)
                                throws IOException

Run the transaction.

Parameters:

server - Hosting server instance. Can be null when testing (won't try and update in zk if a null server)

services - Used to online/offline regions.

Returns:

Regions created

Throws:

IOException - If thrown, transaction failed. Call rollback(Server, RegionServerServices)

IOException

See Also:

rollback(Server, RegionServerServices)

stepsAfterPONR

public PairOfSameType<HRegion> stepsAfterPONR(Server server,
                                     RegionServerServices services,
                                     PairOfSameType<HRegion> regions)
                                       throws IOException

Throws:

IOException

addLocation

public Put addLocation(Put p,
              ServerName sn,
              long openSeqNum)

rollback

public boolean rollback(Server server,
               RegionServerServices services)
                 throws IOException

Parameters:

server - Hosting server instance (May be null when testing).

services -

Returns:

True if we successfully rolled back, false if we got to the point of no return and so now need to abort the server to minimize damage.

Throws:

IOException - If thrown, rollback failed. Take drastic action.

createNodeSplitting

public static void createNodeSplitting(ZooKeeperWatcher zkw,
                       HRegionInfo region,
                       ServerName serverName,
                       HRegionInfo a,
                       HRegionInfo b)
                                throws org.apache.zookeeper.KeeperException,
                                       IOException

Creates a new ephemeral node in the PENDING_SPLIT state for the specified region. Create it ephemeral in case regionserver dies mid-split.

Does not transition nodes from other states. If a node already exists for this region, a KeeperException.NodeExistsException will be thrown.

Parameters:

zkw - zk reference

region - region to be created as offline

serverName - server event originates from

Throws:

org.apache.zookeeper.KeeperException

IOException

transitionSplittingNode

public static int transitionSplittingNode(ZooKeeperWatcher zkw,
                          HRegionInfo parent,
                          HRegionInfo a,
                          HRegionInfo b,
                          ServerName serverName,
                          int znodeVersion,
                          EventType beginState,
                          EventType endState)
                                   throws org.apache.zookeeper.KeeperException,
                                          IOException

Transitions an existing ephemeral node for the specified region which is currently in the begin state to be in the end state. Master cleans up the final SPLIT znode when it reads it (or if we crash, zk will clean it up).

Does not transition nodes from other states. If for some reason the node could not be transitioned, the method returns -1. If the transition is successful, the version of the node after transition is returned.

This method can fail and return false for three different reasons:

• Node for this region does not exist
• Node for this region is not in the begin state
• After verifying the begin state, update fails because of wrong version (this should never actually happen since an RS only does this transition following a transition to the begin state. If two RS are conflicting, one would fail the original transition to the begin state and not this transition)

Does not set any watches.

This method should only be used by a RegionServer when splitting a region.

Parameters:

zkw - zk reference

parent - region to be transitioned to opened

a - Daughter a of split

b - Daughter b of split

serverName - server event originates from

znodeVersion - expected version of data before modification

beginState - the expected current state the znode should be

endState - the state to be transition to

Returns:

version of node after transition, -1 if unsuccessful transition

Throws:

org.apache.zookeeper.KeeperException - if unexpected zookeeper exception

IOException