HBase原子性保证

HBase提供基于单 行 数据操作的原子性保证

即：对同一行的变更操作（包括针对一列/多列/多column family的操作），要么完全成功，要么完全失败，不会有其他状态
示例：
A客户端针对rowkey=10的行发起操作：dim1:a = 1  dim2:b=1
B客户端针对rowkey=10的行发起操作：dim1:a = 2  dim2:b=2
dim1、dim2为column family， a、b为column

A客户端和B客户端同时发起请求，最终rowkey=10的行各个列的值可能是dim1:a = 1  dim2:b=1，也可能是dim1:a = 2  dim2:b=2
但绝对不会是dim1:a = 1  dim2:b=2

HBase基于行锁来保证单行操作的原子性，可以看下HRegion put的代码(base: HBase 0.94.20):：
org.apache.hadoop.hbase.regionserver.HRegion：

  /**
   * @param put
   * @param lockid
   * @param writeToWAL
   * @throws IOException
   * @deprecated row locks (lockId) held outside the extent of the operation are deprecated.
   */
  public void put(Put put, Integer lockid, boolean writeToWAL)
  throws IOException {
    checkReadOnly();


    // Do a rough check that we have resources to accept a write.  The check is
    // 'rough' in that between the resource check and the call to obtain a
    // read lock, resources may run out.  For now, the thought is that this
    // will be extremely rare; we'll deal with it when it happens.
    checkResources();
    startRegionOperation();
    this.writeRequestsCount.increment();
    this.opMetrics.setWriteRequestCountMetrics(this.writeRequestsCount.get());
    try {
      // We obtain a per-row lock, so other clients will block while one client
      // performs an update. The read lock is released by the client calling
      // #commit or #abort or if the HRegionServer lease on the lock expires.
      // See HRegionServer#RegionListener for how the expire on HRegionServer
      // invokes a HRegion#abort.
      byte [] row = put.getRow();
      // If we did not pass an existing row lock, obtain a new one
      Integer lid = getLock(lockid, row, true);


      try {
        // All edits for the given row (across all column families) must happen atomically.
        internalPut(put, put.getClusterId(), writeToWAL);
      } finally {
        if(lockid == null) releaseRowLock(lid);
      }
    } finally {
      closeRegionOperation();
    }
  }

getLock调用了internalObtainRowLock：

 private Integer internalObtainRowLock(final HashedBytes rowKey, boolean waitForLock)
      throws IOException {
    checkRow(rowKey.getBytes(), "row lock");
    startRegionOperation();
    try {
      CountDownLatch rowLatch = new CountDownLatch(1);

      // loop until we acquire the row lock (unless !waitForLock)
      while (true) {
        CountDownLatch existingLatch = lockedRows.putIfAbsent(rowKey, rowLatch);
        if (existingLatch == null) {
          break;
        } else {
          // row already locked
          if (!waitForLock) {
            return null;
          }
          try {
            if (!existingLatch.await(this.rowLockWaitDuration,
                            TimeUnit.MILLISECONDS)) {
              throw new IOException("Timed out on getting lock for row=" + rowKey);
            }
          } catch (InterruptedException ie) {
            // Empty
          }
        }
      }

      // loop until we generate an unused lock id
      while (true) {
        Integer lockId = lockIdGenerator.incrementAndGet();
        HashedBytes existingRowKey = lockIds.putIfAbsent(lockId, rowKey);
        if (existingRowKey == null) {
          return lockId;
        } else {
          // lockId already in use, jump generator to a new spot
          lockIdGenerator.set(rand.nextInt());
        }
      }
    } finally {
      closeRegionOperation();
    }
  }

HBase行锁的实现细节推荐下： hbase源码解析之行锁  

HBase也提供API（lockRow/unlockRow）显示的获取行锁，但不推荐使用。原因是两个客户端很可能在拥有对方请求的锁时，又同时请求对方已拥有的锁，这样便形成了死锁，在锁超时前，两个被阻塞的客户端都会占用一个服务端的处理线程，而服务器线程是非常稀缺的资源

HBase提供了几个特别的原子操作接口：
 checkAndPut/checkAndDelete/increment/append，这几个接口非常有用，内部实现也是基于行锁
checkAndPut/checkAndDelete内部调用代码片段：

      // Lock row
      Integer lid = getLock(lockId, get.getRow(), true);
      ......
      // get and compare
      try {
        result = get(get, false);
        ......
        //If matches put the new put or delete the new delete
        if (matches) {
          if (isPut) {
            internalPut(((Put) w), HConstants.DEFAULT_CLUSTER_ID, writeToWAL);
          } else {
            Delete d = (Delete)w;
            prepareDelete(d);
            internalDelete(d, HConstants.DEFAULT_CLUSTER_ID, writeToWAL);
          }
          return true;
        }
        return false;
      } finally {
        // release lock
        if(lockId == null) releaseRowLock(lid);
      }

实现逻辑：加锁=>get=>比较=>put/delete

checkAndPut在实际应用中非常有价值，我们线上生成Dpid的项目，多个客户端会并行生成DPID，如果有一个客户端已经生成了一个DPID，则其他客户端不能生成新的DPID，只能获取该DPID
代码片段：

			ret = hbaseUse.checkAndPut("bi.dpdim_mac_dpid_mapping", mac, "dim",
					"dpid", null, dpid);
			if(false == ret){
				String retDpid = hbaseUse.query("bi.dpdim_mac_dpid_mapping", mac, "dim", "dpid");
				if(!retDpid.equals(ABNORMAL)){
					return retDpid;
				}
			}else{
				columnList.add("mac");
				valueList.add(mac);
			}

checkAndPut详细试用可以参考： HBaseEveryDay_Atomic_compare_and_set  

Reference：
HBase - Apache HBase (TM) ACID Properties
hbase源码解析之行锁  
HBase权威指南
HBaseEveryDay_Atomic_compare_and_set