Atomikos-XA事务恢复

说事务恢复流程之前,我们来讨论下,会啥会出现事务恢复?XA二阶段提交协议不是强一致性的吗?要解答这个问题,我们就要来看看XA二阶段协议有什么问题?

问题一 :单点故障

由于协调者的重要性,一旦协调者TM发生故障。参与者RM会一直阻塞下去。尤其在第二阶段,协调者发生故障,那么所有的参与者还都处于锁定事务资源的状态中,而无法继续完成事务操作。(如果是协调者挂掉,可以重新选举一个协调者,但是无法解决因为协调者宕机导致的参与者处于阻塞状态的问题)

问题二 :数据不一致

数据不一致。在二阶段提交的阶段二中,当协调者向参与者发送commit请求之后,发生了局部网络异常或者在发送commit请求过程中协调者发生了故障,这回导致只有一部分参与者接受到了commit请求。而在这部分参与者接到commit请求之后就会执行commit操作。但是其他部分未接到commit请求的机器则无法执行事务提交。于是整个分布式系统便出现了数据不一致性的现象。

如何解决?

解决的方案简单,就是我们在事务的操作的每一步,我们都需要对事务状态的日志进行人为的记录,我们可以把日志记录存储在我们想存储的地方,可以是本地存储,也可以中心化的存储。atomikos的开源版本,我们之前也分析了,它是使用内存 + file的方式,存储在本地,这样的话,如果在一个集群系统里面,如果有节点宕机,日志又存储在本地,所以事务不能及时的恢复(需要重启服务)。

Atomikos 多场景下事务恢复。

Atomikos 提供了二种方式,来应对不同场景下的异常情况。

  • 场景一: 服务节点不宕机,因为其他的原因,产生需要事务恢复的情况。 这个时候才要定时任务进行恢复。 具体的代码 com.atomikos.icatch.imp.TransactionServiceImp.init() 方法,会初始化一个定时任务,进行事务的恢复。
public synchronized void init ( Properties properties ) throws SysException
    {
        shutdownInProgress_ = false;
        control_ = new com.atomikos.icatch.admin.imp.LogControlImp ( (AdminLog) this.recoveryLog );
		ConfigProperties configProperties = new ConfigProperties(properties);
		long recoveryDelay = configProperties.getRecoveryDelay();  
        recoveryTimer = new PooledAlarmTimer(recoveryDelay);  
        recoveryTimer.addAlarmTimerListener(new AlarmTimerListener() {
			@Override
			public void alarm(AlarmTimer timer) {
				//进行事务恢复
				performRecovery();

			}
		});

        TaskManager.SINGLETON.executeTask(recoveryTimer);
        initialized_ = true;
    }
  • 最终会进入com.atomikos.datasource.xa.XATransactionalResource.recover() 方法。
   public void recover() {
    	XaResourceRecoveryManager xaResourceRecoveryManager = XaResourceRecoveryManager.getInstance();
    	if (xaResourceRecoveryManager != null) { //null for LogCloud recovery
    		try {
				xaResourceRecoveryManager.recover(getXAResource());
			} catch (Exception e) {
				refreshXAResource(); //cf case 156968
			}

    	}
    }
  • 场景二: 当服务节点宕机重启动过程中进行事务的恢复。具体实现在com.atomikos.datasource.xa.XATransactionalResource.setRecoveryService()方法里面
 @Override
	public void setRecoveryService ( RecoveryService recoveryService )
            throws ResourceException
    {

        if ( recoveryService != null ) {
            if ( LOGGER.isTraceEnabled() ) LOGGER.logTrace ( "Installing recovery service on resource "
                    + getName () );
            this.branchIdentifier=recoveryService.getName();
         //进行事务恢复
            recover();
        }

    }

com.atomikos.datasource.xa.XATransactionalResource.recover() 流程详解。

Atomikos事务恢复流程源码解析_java

主代码:

	public void recover(XAResource xaResource) throws XAException {
      // 根据XA recovery 协议获取 xid
		List<XID> xidsToRecover = retrievePreparedXidsFromXaResource(xaResource);
		Collection<XID> xidsToCommit;
		try {
            // xid 与日志记录的xid进行匹配
			xidsToCommit = retrieveExpiredCommittingXidsFromLog();
			for (XID xid : xidsToRecover) {
				if (xidsToCommit.contains(xid)) {
            //执行 XA commit xid 进行提交                 
					replayCommit(xid, xaResource);
				} else {
					attemptPresumedAbort(xid, xaResource);
				}
			}
		} catch (LogException couldNotRetrieveCommittingXids) {
			LOGGER.logWarning("Transient error while recovering - will retry later...", couldNotRetrieveCommittingXids);
		}
	}
  • 我们来看一下如何根据 XA recovery 协议获取RM端存储的xid。 进入方法 retrievePreparedXidsFromXaResource(xaResource), 最后进入 com.atomikos.datasource.xa.RecoveryScan.recoverXids()方法。
public static List<XID> recoverXids(XAResource xaResource, XidSelector selector) throws XAException {
		List<XID> ret = new ArrayList<XID>();

        boolean done = false;
        int flags = XAResource.TMSTARTRSCAN;
        Xid[] xidsFromLastScan = null;
        List<XID> allRecoveredXidsSoFar = new ArrayList<XID>();
        do {
        	xidsFromLastScan = xaResource.recover(flags);
            flags = XAResource.TMNOFLAGS;
            done = (xidsFromLastScan == null || xidsFromLastScan.length == 0);
            if (!done) {

                // TEMPTATIVELY SET done TO TRUE
                // TO TOLERATE ORACLE 8.1.7 INFINITE
                // LOOP (ALWAYS RETURNS SAME RECOVER
                // SET). IF A NEW SET OF XIDS IS RETURNED
                // THEN done WILL BE RESET TO FALSE

                done = true;
                for ( int i = 0; i < xidsFromLastScan.length; i++ ) {
                	XID xid = new XID ( xidsFromLastScan[i] );
                    // our own XID implements equals and hashCode properly
                    if (!allRecoveredXidsSoFar.contains(xid)) {
                        // a new xid is returned -> we can not be in a recovery loop -> go on
                        allRecoveredXidsSoFar.add(xid);
                        done = false;
                        if (selector.selects(xid)) {
                        	ret.add(xid);
                        }
                    }
                }
            }
        } while (!done);

		return ret;
	}
  • 我们重点关注xidsFromLastScan = xaResource.recover(flags); 这个方法,如果我们使用MySQL,那么久会进入 MysqlXAConnection.recover()方法。执行 XA recovery xid 语句来获取 xid。
 protected static Xid[] recover(Connection c, int flag) throws XAException {
        /*
         * The XA RECOVER statement returns information for those XA transactions on the MySQL server that are in the PREPARED state. (See Section 13.4.7.2, ???XA
         * Transaction States???.) The output includes a row for each such XA transaction on the server, regardless of which client started it.
         *
         * XA RECOVER output rows look like this (for an example xid value consisting of the parts 'abc', 'def', and 7):
         *
         * mysql> XA RECOVER;
         * +----------+--------------+--------------+--------+
         * | formatID | gtrid_length | bqual_length | data |
         * +----------+--------------+--------------+--------+
         * | 7 | 3 | 3 | abcdef |
         * +----------+--------------+--------------+--------+
         *
         * The output columns have the following meanings:
         *
         * formatID is the formatID part of the transaction xid
         * gtrid_length is the length in bytes of the gtrid part of the xid
         * bqual_length is the length in bytes of the bqual part of the xid
         * data is the concatenation of the gtrid and bqual parts of the xid
         */

        boolean startRscan = ((flag & TMSTARTRSCAN) > 0);
        boolean endRscan = ((flag & TMENDRSCAN) > 0);

        if (!startRscan && !endRscan && flag != TMNOFLAGS) {
            throw new MysqlXAException(XAException.XAER_INVAL, Messages.getString("MysqlXAConnection.001"), null);
        }

        //
        // We return all recovered XIDs at once, so if not  TMSTARTRSCAN, return no new XIDs
        //
        // We don't attempt to maintain state to check for TMNOFLAGS "outside" of a scan
        //

        if (!startRscan) {
            return new Xid[0];
        }

        ResultSet rs = null;
        Statement stmt = null;

        List<MysqlXid> recoveredXidList = new ArrayList<MysqlXid>();

        try {
            // TODO: Cache this for lifetime of XAConnection
            stmt = c.createStatement();

            rs = stmt.executeQuery("XA RECOVER");

            while (rs.next()) {
                final int formatId = rs.getInt(1);
                int gtridLength = rs.getInt(2);
                int bqualLength = rs.getInt(3);
                byte[] gtridAndBqual = rs.getBytes(4);

                final byte[] gtrid = new byte[gtridLength];
                final byte[] bqual = new byte[bqualLength];

                if (gtridAndBqual.length != (gtridLength + bqualLength)) {
                    throw new MysqlXAException(XAException.XA_RBPROTO, Messages.getString("MysqlXAConnection.002"), null);
                }

                System.arraycopy(gtridAndBqual, 0, gtrid, 0, gtridLength);
                System.arraycopy(gtridAndBqual, gtridLength, bqual, 0, bqualLength);

                recoveredXidList.add(new MysqlXid(gtrid, bqual, formatId));
            }
        } catch (SQLException sqlEx) {
            throw mapXAExceptionFromSQLException(sqlEx);
        } finally {
            if (rs != null) {
                try {
                    rs.close();
                } catch (SQLException sqlEx) {
                    throw mapXAExceptionFromSQLException(sqlEx);
                }
            }

            if (stmt != null) {
                try {
                    stmt.close();
                } catch (SQLException sqlEx) {
                    throw mapXAExceptionFromSQLException(sqlEx);
                }
            }
        }

        int numXids = recoveredXidList.size();

        Xid[] asXids = new Xid[numXids];
        Object[] asObjects = recoveredXidList.toArray();

        for (int i = 0; i < numXids; i++) {
            asXids[i] = (Xid) asObjects[i];
        }

        return asXids;
    }
  • 这里要注意如果Mysql的版本 <5.7.7 ,则不会有任何数据,在以后的版本中Mysql进行了修复,因此如果我们想要使用MySQL充当RM,版本必须 >= 5.7.7 ,原因是:

MySQL 5.6版本在客户端退出的时候,自动把已经prepare的事务回滚了,那么MySQL为什么要这样做?这主要取决于MySQL的内部实现,MySQL 5.7以前的版本,对于prepare的事务,MySQL是不会记录binlog的(官方说是减少fsync,起到了优化的作用)。只有当分布式事务提交的时候才会把前面的操作写入binlog信息,所以对于binlog来说,分布式事务与普通的事务没有区别,而prepare以前的操作信息都保存在连接的IO_CACHE中,如果这个时候客户端退出了,以前的binlog信息都会被丢失,再次重连后允许提交的话,会造成Binlog丢失,从而造成主从数据的不一致,所以官方在客户端退出的时候直接把已经prepare的事务都回滚了!

  • 回到主线再从自己记录的事务日志里面获取XID
  Collection<XID> xidsToCommit = retrieveExpiredCommittingXidsFromLog();
  • 我们来看下获取事务日志里面的XID的retrieveExpiredCommittingXidsFromLog()方法。 然后进入com.atomikos.recovery.imp.RecoveryLogImp.getCommittingParticipants()方法。
public Collection<ParticipantLogEntry> getCommittingParticipants()
			throws LogReadException {
		Collection<ParticipantLogEntry> committingParticipants = new HashSet<ParticipantLogEntry>();
		Collection<CoordinatorLogEntry> committingCoordinatorLogEntries = repository.findAllCommittingCoordinatorLogEntries();

		for (CoordinatorLogEntry coordinatorLogEntry : committingCoordinatorLogEntries) {
			for (ParticipantLogEntry participantLogEntry : coordinatorLogEntry.participants) {
				committingParticipants.add(participantLogEntry);
			}
		}
		return committingParticipants;
	}

到这里我们来简单介绍一下,事务日志的存储结构。首先是 CoordinatorLogEntry,这是一次XA事务的所有信息实体类。

public class CoordinatorLogEntry implements Serializable {

  //全局事务id
 	public final String id;

   //是否已经提交
	public final boolean wasCommitted;

	/**
	 * Only for subtransactions, null otherwise.
	 */
	public final String superiorCoordinatorId;

   //参与者集合
	public final ParticipantLogEntry[] participants;
}
  • 再来看一下参与者实体类 ParticipantLogEntry :
public class ParticipantLogEntry implements Serializable {

	private static final long serialVersionUID = 1728296701394899871L;

	/**
	 * The ID of the global transaction as known by the transaction core.
	 */

	public final String coordinatorId;

	/**
	 * Identifies the participant within the global transaction.
	 */

	public final String uri;

	/**
	 * When does this participant expire (expressed in millis since Jan 1, 1970)?
	 */

	public final long expires;

	/**
	 * Best-known state of the participant.
	 */
	public final TxState state;

	/**
	 * For diagnostic purposes, null if not relevant.
	 */
	public final String resourceName;
}
  • 回到com.atomikos.recovery.xa.DefaultXaRecoveryLog.getExpiredCommittingXids() 方法,可以到获取了一次XA事务过程中,存储的事务日志中的xid。
public Set<XID> getExpiredCommittingXids() throws LogReadException {
		Set<XID> ret = new HashSet<XID>();
		Collection<ParticipantLogEntry> entries = log.getCommittingParticipants();
		for (ParticipantLogEntry entry : entries) {
			if (expired(entry) && !http(entry)) {
				XID xid = new XID(entry.coordinatorId, entry.uri);
				ret.add(xid);
			}
		}
		return ret;
	}
  • 如果从RM中通过XA recovery取出的XID,包含在从事务日志中取出的XID,则进行commit,否则进行rollback.
List<XID> xidsToRecover = retrievePreparedXidsFromXaResource(xaResource);
		Collection<XID> xidsToCommit;
		try {
			xidsToCommit = retrieveExpiredCommittingXidsFromLog();
			for (XID xid : xidsToRecover) {
				if (xidsToCommit.contains(xid)) {
					replayCommit(xid, xaResource);
				} else {
					attemptPresumedAbort(xid, xaResource);
				}
			}
		} catch (LogException couldNotRetrieveCommittingXids) {
			LOGGER.logWarning("Transient error while recovering - will retry later...", couldNotRetrieveCommittingXids);
		}
  • replayCommit 方法如下:
private void replayCommit(XID xid, XAResource xaResource) {
		if (LOGGER.isDebugEnabled()) LOGGER.logDebug("Replaying commit of xid: " + xid);
		try {
      //进行事务提交
			xaResource.commit(xid, false);
     //更新事务日志
			log.terminated(xid);
		} catch (XAException e) {
			if (alreadyHeuristicallyTerminatedByResource(e)) {
				handleHeuristicTerminationByResource(xid, xaResource, e, true);
			} else if (xidTerminatedInResourceByConcurrentCommit(e)) {
				log.terminated(xid);
			} else {
				LOGGER.logWarning("Transient error while replaying commit - will retry later...", e);
			}
		}
	}
  • attemptPresumedAbort(xid, xaResource); 方法如下:
private void attemptPresumedAbort(XID xid, XAResource xaResource) {
		try {
			log.presumedAborting(xid);
			if (LOGGER.isDebugEnabled()) LOGGER.logDebug("Presumed abort of xid: " + xid);
			try {
         //进行回滚
				xaResource.rollback(xid);
        //更新日志状态
				log.terminated(xid);
			} catch (XAException e) {
				if (alreadyHeuristicallyTerminatedByResource(e)) {
					handleHeuristicTerminationByResource(xid, xaResource, e, false);
				} else if (xidTerminatedInResourceByConcurrentRollback(e)) {
					log.terminated(xid);
				} else {
					LOGGER.logWarning("Unexpected exception during recovery - ignoring to retry later...", e);
				}
			}
		} catch (IllegalStateException presumedAbortNotAllowedInCurrentLogState) {
			// ignore to retry later if necessary
		} catch (LogException logWriteException) {
			LOGGER.logWarning("log write failed for Xid: "+xid+", ignoring to retry later", logWriteException);
		}
	}

文章到此,已经写的很长很多了,我们分析了ShardingSphere对于XA方案,提供了一套SPI解决方案,对Atomikos进行了整合,也分析了Atomikos初始化流程,开始事务流程,获取连接流程,提交事务流程,回滚事务流程,事务恢复流程。 希望对大家理解XA的原理有所帮助。

作者介绍: 肖宇,Apache ShardingSphere Committer,开源hmily分布式事务框架作者, 开源soul网关作者,热爱开源,追求写优雅代码。目前就职入京东数科,参与ShardingSphere的开源建设,以及分布式数据库的研发工作。