[HBase]Table create

让我们从创建表开始探索hbase内部机制。假设hbase.root目录为/new，测试表名为t1。

client使用HBaseAdmin的createTable接口，过程如下

1. 建立HMasterRPC连接，并调用之，由于hmaster端创建table是异步的，所以这里是一个异步操作。如果不指定split规则，默认会创建一个空region。

 getMaster().createTable(desc, splitKeys);

 2. client线程全表扫描meta表，检查t1表的region是否都分配好。默认重试100次，每次失败sleep。

  MetaScannerVisitor visitor = new MetaScannerVisitorBase() {
        @Override
        public boolean processRow(Result rowResult) throws IOException {
          HRegionInfo info = Writables.getHRegionInfoOrNull(
              rowResult.getValue(HConstants.CATALOG_FAMILY,
                  HConstants.REGIONINFO_QUALIFIER));
          ......
	   //拿'server'列，如果有值，则认为分配成功
          byte [] value = rowResult.getValue(HConstants.CATALOG_FAMILY,
              HConstants.SERVER_QUALIFIER);
            // Make sure that regions are assigned to server
          if (value != null && value.length > 0) {
            hostAndPort = Bytes.toString(value);
          }
          if (!(info.isOffline() || info.isSplit()) && hostAndPort != null) {
            actualRegCount.incrementAndGet();
          }
          return true;
        }
      };
      MetaScanner.metaScan(conf, visitor, desc.getName());

 来看HMaster的create table RCP接口

1.构造CreateTableHandler

    1.1 等待META表就位，如果就位，则获取META表第一个region的location，并建立RPC连接

    

public ServerName waitForMeta(long timeout)
  throws InterruptedException, IOException, NotAllMetaRegionsOnlineException {
    long stop = System.currentTimeMillis() + timeout;
    long waitTime = Math.min(50, timeout);
    synchronized (metaAvailable) {
      while(!stopped && (timeout == 0 || System.currentTimeMillis() < stop)) {
        if (getMetaServerConnection() != null) {
          return metaLocation;
        }
        // perhaps -ROOT- region isn't available, let us wait a bit and retry.
        metaAvailable.wait(waitTime);
      }
      if (getMetaServerConnection() == null) {
        throw new NotAllMetaRegionsOnlineException("Timed out (" + timeout + "ms)");
      }
      return metaLocation;
    }
  }

         1.2 判断t1表是否已存在

         1.3 创建并设置t1表在zk中的节点状态为‘enabling’，节点路径/hbase/table/t1

         

 private void setTableState(final String tableName, final TableState state)
  throws KeeperException {
    String znode = ZKUtil.joinZNode(this.watcher.tableZNode, tableName);
    if (ZKUtil.checkExists(this.watcher, znode) == -1) {
      ZKUtil.createAndFailSilent(this.watcher, znode);
    }
    synchronized (this.cache) {
      ZKUtil.setData(this.watcher, znode, Bytes.toBytes(state.toString()));
      this.cache.put(tableName, state);
    }
  }

 2.异步提交CreateTableHandler

this.executorService.submit(new CreateTableHandler(this,
      this.fileSystemManager, this.serverManager, hTableDescriptor, conf,
      newRegions, catalogTracker, assignmentManager));

 3.CreateTableHandler运行

              3.1 将table的元信息写入HDFS下的.tableinfo文件中，文件目录/new/t1/.tableinfo.0000000001。

           

private static Path writeTableDescriptor(final FileSystem fs,
      final HTableDescriptor hTableDescriptor, final Path tableDir,
      final FileStatus status)
  throws IOException {
    // Get temporary dir into which we'll first write a file to avoid
    // half-written file phenomeon.
	//先写tmp目录
    Path tmpTableDir = new Path(tableDir, ".tmp");
    //顺序号，从0开始
    int currentSequenceid =
      status == null? 0: getTableInfoSequenceid(status.getPath());
    int sequenceid = currentSequenceid;
    // Put arbitrary upperbound on how often we retry
    int retries = 10;
    int retrymax = currentSequenceid + retries;
    Path tableInfoPath = null;
    do {
      sequenceid += 1;
	//HDFS文件名，类是.tableinfo.0000000001
      Path p = getTableInfoFileName(tmpTableDir, sequenceid);
      if (fs.exists(p)) {
        LOG.debug(p + " exists; retrying up to " + retries + " times");
        continue;
      }
      try {
	//写内容
        writeHTD(fs, p, hTableDescriptor);
        tableInfoPath = getTableInfoFileName(tableDir, sequenceid);
	//重命名成最终文件
        if (!fs.rename(p, tableInfoPath)) {
          throw new IOException("Failed rename of " + p + " to " + tableInfoPath);
        }
      } 
	.......
      break;
    } while (sequenceid < retrymax);
    return tableInfoPath;
  }

              3.2 创建region

       

HRegion region = HRegion.createHRegion(newRegion,
        this.fileSystemManager.getRootDir(), this.conf,
        this.hTableDescriptor, null, false, true);

             3.3 META表新增记录，写入regioninfo列信息

 private static Put makePutFromRegionInfo(HRegionInfo regionInfo)
  throws IOException {
    Put put = new Put(regionInfo.getRegionName());
    put.add(HConstants.CATALOG_FAMILY, HConstants.REGIONINFO_QUALIFIER,
        Writables.getBytes(regionInfo));
    return put;
  }

            3.4 close region

            3.5 从zk获取活着的region server

//从/hbase/rs下获取并过滤掉那些dead的机器
List<ServerName> servers = serverManager.getOnlineServersList();
    // Remove the deadNotExpired servers from the server list.
    assignmentManager.removeDeadNotExpiredServers(servers);

           3.6 region分配，默认随机均匀分配，使用多线程批量分配，业务线程等待直到所有region都分配成功，详细的分配过程将在下一篇介绍

this.assignmentManager.assignUserRegions(Arrays.asList(newRegions),
        servers);

          3.7 设置t1表在zk中的节点状态为‘enabled’，节点路径/hbase/table/t1

 

小节

create table主要涉及，table元数据写入，region分配，zk状态信息修改，meta表修改和检查。

评论

1 楼 leibnitz 2014-07-17  

想问下你hbase版本是多少？我的是94.2，看到有些少少差异。
另外，我用的版本在regions分配严格来说不是随机，而是round-robin.
总体来说写得很好！
thanks