Chinaunix首页 | 论坛 | 博客
  • 博客访问: 8033190
  • 博文数量: 594
  • 博客积分: 13065
  • 博客等级: 上将
  • 技术积分: 10324
  • 用 户 组: 普通用户
  • 注册时间: 2008-03-26 16:44
个人简介

推荐: blog.csdn.net/aquester https://github.com/eyjian https://www.cnblogs.com/aquester http://blog.chinaunix.net/uid/20682147.html

文章分类

全部博文(594)

分类: HADOOP

2017-05-19 10:25:55

正常情况下,什么时候上报blocks,是由NameNode通过回复心跳响应的方式触发的。
一次机房搬迁中,原机房hadoop版本为2.7.2,新机房版本为2.8.0,采用先扩容再缩容的方式搬迁。由于新旧机房机型不同和磁盘数不同,操作过程搞混过hdfs-site.xml,因为两种不同的机型,hdfs-site.xml不便做到相同,导致了NameNode报大量“missing block”。

然而依据NameNode所报信息,在DataNode能找到那些被标记为“missing”的blocks。修复配置问题后,“missing block”并没有消失。结合DataNode源代码,估计是因为DataNode没有向NameNode上报blocks。

结合DataNode的源代码,发现了HDFS自带的工具triggerBlockReport,它可以强制指定的DataNode向NameNode上报块,使用方法为:
hdfs dfsadmin -triggerBlockReport datanode_host:ipc_port
如:hdfs dfsadmin -triggerBlockReport 192.168.31.35:50020

正常情况下NameNode启动时,会要求DataNode上报一次blocks(通过fullBlockReportLeaseId值来控制),相关源代码如下:

DataNode相关代码(BPServiceActor.java):
private void offerService() throws Exception {
    HeartbeatResponse resp = sendHeartBeat(requestBlockReportLease); // 向NameNode发向心跳
    long fullBlockReportLeaseId = resp.getFullBlockReportLeaseId(); // 心跳响应
    boolean forceFullBr = scheduler.forceFullBlockReport.getAndSet(false); // triggerBlockReport强制上报仅一次有效
    if (forceFullBr) {
        LOG.info("Forcing a full block report to " + nnAddr);
    }
    if ((fullBlockReportLeaseId != 0) || forceFullBr) {
        cmds = blockReport(fullBlockReportLeaseId);
        fullBlockReportLeaseId = 0;
    }
}

// NameNode相关代码(FSNamesystem.java):
/**
* The given node has reported in.  This method should:
* 1) Record the heartbeat, so the datanode isn't timed out
* 2) Adjust usage stats for future block allocation

* If a substantial amount of time passed since the last datanode 
* heartbeat then request an immediate block report.  

* @return an array of datanode commands 
* @throws IOException
*/
HeartbeatResponse handleHeartbeat(DatanodeRegistration nodeReg,
  StorageReport[] reports, long cacheCapacity, long cacheUsed,
  int xceiverCount, int xmitsInProgress, int failedVolumes,
  VolumeFailureSummary volumeFailureSummary,
  boolean requestFullBlockReportLease) throws IOException {
    readLock();
    try {
        //get datanode commands
        final int maxTransfer = blockManager.getMaxReplicationStreams() - xmitsInProgress;
        DatanodeCommand[] cmds = blockManager.getDatanodeManager().handleHeartbeat(
            nodeReg, reports, blockPoolId, cacheCapacity, cacheUsed,
            xceiverCount, maxTransfer, failedVolumes, volumeFailureSummary);

        long fullBlockReportLeaseId = 0;
        if (requestFullBlockReportLease) {
            fullBlockReportLeaseId =  blockManager.requestBlockReportLeaseId(nodeReg);
        }
        //create ha status
        final NNHAStatusHeartbeat haState = new NNHAStatusHeartbeat(
            haContext.getState().getServiceState(),
            getFSImage().getCorrectLastAppliedOrWrittenTxId());

        return new HeartbeatResponse(cmds, haState, rollingUpgradeInfo, fullBlockReportLeaseId);
    } finally {
        readUnlock("handleHeartbeat");
    }
}


阅读(9293) | 评论(0) | 转发(0) |
给主人留下些什么吧!~~