org.apache.hadoop.hbase.regionserver

Class MultiVersionConsistencyControl

java.lang.Object

org.apache.hadoop.hbase.regionserver.MultiVersionConsistencyControl

@InterfaceAudience.Private
public class MultiVersionConsistencyControl
extends Object

Manages the read/write consistency within memstore. This provides an interface for readers to determine what entries to ignore, and a mechanism for writers to obtain new write numbers, then "commit" the new writes for readers to read (thus forming atomic transactions).

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	MultiVersionConsistencyControl.WriteEntry

Field Summary

Fields

Modifier and Type	Field and Description
static long	FIXED_SIZE

Constructor Summary

Constructors

Constructor and Description
MultiVersionConsistencyControl() Default constructor.

Method Summary

Modifier and Type	Method and Description
MultiVersionConsistencyControl.WriteEntry	beginMemstoreInsertWithSeqNum(long curSeqNum) This function starts a MVCC transaction with current region's log change sequence number.
void	cancelMemstoreInsert(MultiVersionConsistencyControl.WriteEntry writeEntry) Cancel a write insert that failed.
void	completeMemstoreInsert(MultiVersionConsistencyControl.WriteEntry e) Complete a MultiVersionConsistencyControl.WriteEntry that was created by beginMemstoreInsert().
void	completeMemstoreInsertWithSeqNum(MultiVersionConsistencyControl.WriteEntry e, SequenceId seqId) Complete a MultiVersionConsistencyControl.WriteEntry that was created by beginMemstoreInsertWithSeqNum(long).
static long	getPreAssignedWriteNumber(AtomicLong sequenceId) Get a mvcc write number before an actual one(its log sequence Id) being assigned
void	initialize(long startPoint) Initializes the memstoreRead/Write points appropriately.
long	memstoreReadPoint()
void	waitForPreviousTransactionsComplete() Wait for all previous MVCC transactions complete
void	waitForPreviousTransactionsComplete(MultiVersionConsistencyControl.WriteEntry waitedEntry)

Methods inherited from class java.lang.Object

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

Field Detail

FIXED_SIZE

public static final long FIXED_SIZE

Constructor Detail

MultiVersionConsistencyControl

public MultiVersionConsistencyControl()

Default constructor. Initializes the memstoreRead/Write points to 0.

Method Detail

initialize

public void initialize(long startPoint)

Initializes the memstoreRead/Write points appropriately.

Parameters:

startPoint -

getPreAssignedWriteNumber

public static long getPreAssignedWriteNumber(AtomicLong sequenceId)

Get a mvcc write number before an actual one(its log sequence Id) being assigned

Parameters:

sequenceId -

Returns:

long a faked write number which is bigger enough not to be seen by others before a real one is assigned

beginMemstoreInsertWithSeqNum

public MultiVersionConsistencyControl.WriteEntry beginMemstoreInsertWithSeqNum(long curSeqNum)

This function starts a MVCC transaction with current region's log change sequence number. Since we set change sequence number when flushing current change to WAL(late binding), the flush order may differ from the order to start a MVCC transaction. For example, a change begins a MVCC firstly may complete later than a change which starts MVCC at a later time. Therefore, we add a safe bumper to the passed in sequence number to start a MVCC so that no other concurrent transactions will reuse the number till current MVCC completes(success or fail). The "faked" big number is safe because we only need it to prevent current change being seen and the number will be reset to real sequence number(set in log sync) right before we complete a MVCC in order for MVCC to align with flush sequence.

Parameters:

curSeqNum -

Returns:

WriteEntry a WriteEntry instance with the passed in curSeqNum

completeMemstoreInsertWithSeqNum

public void completeMemstoreInsertWithSeqNum(MultiVersionConsistencyControl.WriteEntry e,
                                             SequenceId seqId)
                                      throws IOException

Complete a MultiVersionConsistencyControl.WriteEntry that was created by beginMemstoreInsertWithSeqNum(long). At the end of this call, the global read point is at least as large as the write point of the passed in WriteEntry. Thus, the write is visible to MVCC readers.

Throws:

IOException

cancelMemstoreInsert

public void cancelMemstoreInsert(MultiVersionConsistencyControl.WriteEntry writeEntry)

Cancel a write insert that failed. Removes the write entry without advancing read point or without interfering with write entries queued behind us. It is like #advanceMemstore(WriteEntry) only this method will move the read point to the sequence id that is in WriteEntry even if it ridiculous (see the trick in HRegion where we call getPreAssignedWriteNumber(AtomicLong) just to mark it as for special handling).

Parameters:

writeEntry - Failed attempt at write. Does cleanup.

completeMemstoreInsert

public void completeMemstoreInsert(MultiVersionConsistencyControl.WriteEntry e)

Complete a MultiVersionConsistencyControl.WriteEntry that was created by beginMemstoreInsert(). At the end of this call, the global read point is at least as large as the write point of the passed in WriteEntry. Thus, the write is visible to MVCC readers.

waitForPreviousTransactionsComplete

public void waitForPreviousTransactionsComplete()

Wait for all previous MVCC transactions complete

waitForPreviousTransactionsComplete

public void waitForPreviousTransactionsComplete(MultiVersionConsistencyControl.WriteEntry waitedEntry)

memstoreReadPoint

public long memstoreReadPoint()