经验总结：高性能的数据同步

     最近在做一个银行的生产数据脱敏系统，今天写代码时遇到了一个“瓶颈”，脱敏系统需要将生产环境上Infoxmix里的数据原封不动的Copy到另一台Oracle数据库服务器上，然后对Copy后的数据作些漂白处理。为了将人为干预的因素降到最低，在系统设计时采用Java代码对数据作Copy,思路如图

    首先在代码与生产库间建立一个Connection，将读取到的数据放在ResultSet对象，然后再与开发库建立一个Connection。从ResultSet取出数据后通过TestConnection插入到开发库，以此来实现Copy。代码写完后运行程序，速度太慢了，一秒钟只能Copy一千条数据，生产库上有上亿条数据，按照这个速度同步完要到猴年马月呀,用PreparedStatement批处理速度也没有提交多少。我想能不能用多线程处理，多个人干活总比一个人干活速度要快。

    假设生产库有1万条数据，我开5个线程，每个线程分2000条数据，同时向开发库里插数据，Oracle支持高并发这样的话速度至少会提高好多倍，按照这个思路重新进行了编码，批处理设置为1万条一提交，统计插入数量的变量使用java.util.concurrent.atomic.AtomicLong，程序一运行，传输速度飞快CPU利用率在70%~90%，现在一秒钟可以拷贝50万条记录，没过几分钟上亿条数据一条不落地全部Copy到目标库。

 

在查询的时候我用了如下语句

String queryStr = "SELECT * FROM xx";

ResultSet coreRs = PreparedStatement.executeQuery(queryStr);

实习生问如果xx表里有上千万条记录，你全部查询出来放到ResultSet,那内存不溢出了么？Java在设计的时候已经考虑到这个问题了，并没有查询出所有的数据，而是只查询了一部分数据放到ResultSet,数据“用完”它会自动查询下一批数据，你可以用setFetchSize(int rows)方法设置一个建议值给ResultSet,告诉它每次从数据库Fetch多少条数据。但我不赞成，因为JDBC驱动会根据实际情况自动调整Fetch的数量。另外性能也与网线的带宽有直接的关系。

相关代码

package com.dlbank.domain;

import java.sql.Connection;
import java.sql.PreparedStatement;
import java.sql.ResultSet;
import java.sql.Statement;
import java.util.List;
import java.util.concurrent.atomic.AtomicLong;

import org.apache.log4j.Logger;

/**
 *<p>title: 数据同步类 </p>  
 *<p>Description: 该类用于将生产核心库数据同步到开发库</p>  
 *@author Tank Zhang 
 */
public class CoreDataSyncImpl implements CoreDataSync {
	
	private List<String> coreTBNames; //要同步的核心库表名
	private ConnectionFactory connectionFactory;
	private Logger log = Logger.getLogger(getClass());
	
	private AtomicLong currentSynCount = new AtomicLong(0L); //当前已同步的条数
	
	private int syncThreadNum;  //同步的线程数

	@Override
	public void syncData(int businessType) throws Exception {
		
		for (String tmpTBName : coreTBNames) {
			log.info("开始同步核心库" + tmpTBName + "表数据");
			// 获得核心库连接
			Connection coreConnection = connectionFactory.getDMSConnection(4);
			Statement coreStmt = coreConnection.createStatement();
			//为每个线程分配结果集
			ResultSet coreRs = coreStmt.executeQuery("SELECT count(*) FROM "+tmpTBName);
			coreRs.next();
			//总共处理的数量
			long totalNum = coreRs.getLong(1);
			//每个线程处理的数量
			long ownerRecordNum =(long) Math.ceil((totalNum / syncThreadNum)); 
			log.info("共需要同步的数据量："+totalNum);
			log.info("同步线程数量："+syncThreadNum);
			log.info("每个线程可处理的数量："+ownerRecordNum);
			// 开启五个线程向目标库同步数据
			for(int i=0; i < syncThreadNum; i ++){
				StringBuilder sqlBuilder = new StringBuilder();
				//拼装后SQL示例
				//Select * From dms_core_ds Where id between 1 And 657398
				//Select * From dms_core_ds Where id between 657399 And 1314796
				//Select * From dms_core_ds Where id between 1314797 And 1972194
				//Select * From dms_core_ds Where id between 1972195 And 2629592
				//Select * From dms_core_ds Where id between 2629593 And 3286990
				//..
				sqlBuilder.append("Select * From ").append(tmpTBName)
						.append(" Where id between " ).append(i * ownerRecordNum +1)
						.append( " And ")
						.append((i * ownerRecordNum + ownerRecordNum));
				Thread workThread = new Thread(
						new WorkerHandler(sqlBuilder.toString(),businessType,tmpTBName));
				workThread.setName("SyncThread-"+i);
				workThread.start();
			}
			while (currentSynCount.get() < totalNum);
			//休眠一会儿让数据库有机会commit剩余的批处理(只针对JUnit单元测试,因为单元测试完成后会关闭虚拟器，使线程里的代码没有机会作提交操作);
			//Thread.sleep(1000 * 3);
			log.info( "核心库"+tmpTBName+"表数据同步完成，共同步了" + currentSynCount.get() + "条数据");
		}
	}// end for loop
	
	public void setCoreTBNames(List<String> coreTBNames) {
		this.coreTBNames = coreTBNames;
	}

	public void setConnectionFactory(ConnectionFactory connectionFactory) {
		this.connectionFactory = connectionFactory;
	}
	
	public void setSyncThreadNum(int syncThreadNum) {
		this.syncThreadNum = syncThreadNum;
	}
	
	//数据同步线程
	final class WorkerHandler implements Runnable {
		ResultSet coreRs;
		String queryStr;
		int businessType;
		String targetTBName;
		public WorkerHandler(String queryStr,int businessType,String targetTBName) {
			this.queryStr = queryStr;
			this.businessType = businessType;
			this.targetTBName = targetTBName;
		}
		@Override
		public void run() {
			try {
				//开始同步
				launchSyncData();
			} catch(Exception e){
				log.error(e);
				e.printStackTrace();
			}
		}
		//同步数据方法
		void launchSyncData() throws Exception{
			// 获得核心库连接
			Connection coreConnection = connectionFactory.getDMSConnection(4);
			Statement coreStmt = coreConnection.createStatement();
			// 获得目标库连接
			Connection targetConn = connectionFactory.getDMSConnection(businessType);
			targetConn.setAutoCommit(false);// 设置手动提交
			PreparedStatement targetPstmt = targetConn.prepareStatement("INSERT INTO " + targetTBName+" VALUES (?,?,?,?,?)");
			ResultSet coreRs = coreStmt.executeQuery(queryStr);
			log.info(Thread.currentThread().getName()+"'s Query SQL::"+queryStr);
			int batchCounter = 0; //累加的批处理数量
			while (coreRs.next()) {
				targetPstmt.setString(1, coreRs.getString(2));
				targetPstmt.setString(2, coreRs.getString(3));
				targetPstmt.setString(3, coreRs.getString(4));
				targetPstmt.setString(4, coreRs.getString(5));
				targetPstmt.setString(5, coreRs.getString(6));
				targetPstmt.addBatch();
				batchCounter++;
				currentSynCount.incrementAndGet();//递增
				if (batchCounter % 10000 == 0) { //1万条数据一提交
					targetPstmt.executeBatch();
					targetPstmt.clearBatch();
					targetConn.commit();
				}
			}
			//提交剩余的批处理
			targetPstmt.executeBatch();
			targetPstmt.clearBatch();
			targetConn.commit();
			//释放连接 
			connectionFactory.release(targetConn, targetPstmt,coreRs);
		}
	}
}