hbase写入性能影响续

阅读更多

    今天发现hbase在写入一张新表时，写入过程中时常会出现一段连续的时间无法写入数据，时间长达10s以上。经过查看region server日志发现有如下记录：

2011-04-18 14:50:22,942 INFO org.apache.hadoop.hbase.regionserver.HRegion: Blocking updates for 'IPC Server handler 6 on 60020' on region data1,70712707089004,1303109282990.da1eccd9d9ebf0f8bfe1116fe7046763.: memstore size 128.1m is >= than blocking 128.0m size

2011-04-18 14:50:22,944 INFO org.apache.hadoop.hbase.regionserver.HRegion: Blocking updates for 'IPC Server handler 4 on 60020' on region data1,70712707089004,1303109282990.da1eccd9d9ebf0f8bfe1116fe7046763.: memstore size 128.3m is >= than blocking 128.0m size

2011-04-18 14:50:22,955 INFO org.apache.hadoop.hbase.regionserver.HRegion: Blocking updates for 'IPC Server handler 8 on 60020' on region data1,70712707089004,1303109282990.da1eccd9d9ebf0f8bfe1116fe7046763.: memstore size 128.3m is >= than blocking 128.0m size

2011-04-18 14:50:22,955 INFO org.apache.hadoop.hbase.regionserver.HRegion: Blocking updates for 'IPC Server handler 7 on 60020' on region data1,70712707089004,1303109282990.da1eccd9d9ebf0f8bfe1116fe7046763.: memstore size 128.3m is >= than blocking 128.0m size

    查找了原代码，在HRegion中发现如下片断：

private void checkResources() {

    // If catalog region, do not impose resource constraints or block updates.
    if (this.getRegionInfo().isMetaRegion()) return;

    boolean blocked = false;
    while (this.memstoreSize.get() > this.blockingMemStoreSize) {
      requestFlush();
      if (!blocked) {
        LOG.info("Blocking updates for '" + Thread.currentThread().getName() +
          "' on region " + Bytes.toStringBinary(getRegionName()) +
          ": memstore size " +
          StringUtils.humanReadableInt(this.memstoreSize.get()) +
          " is >= than blocking " +
          StringUtils.humanReadableInt(this.blockingMemStoreSize) + " size");
      }
      blocked = true;
      synchronized(this) {
        try {
          wait(threadWakeFrequency);
        } catch (InterruptedException e) {
          // continue;
        }
      }
    }
    if (blocked) {
      LOG.info("Unblocking updates for region " + this + " '"
          + Thread.currentThread().getName() + "'");
    }
  }

    原来是因为region server在写入时会检查每个region对应的memstore的总大小是否超过了memstore默认大小的2倍(hbase.hregion.memstore.block.multiplier决定)，如果超过了则锁住memstore不让新写请求进来并触发flush，避免产生OOM。由于在flush时还会触发compact/split等操作。因此这个过程通常比较长，必须要紧持到对应的memstore完全刷新到磁盘才会结束，因此regionserver会睡眠10s再检查memstore是不是低于阀值。
    对于在线应用来说，10s的时间是不可接受的，但这个过程确实非常漫长，因此可以调整以下配置以减少或者避免这种情况的出现。

    hbase.hregion.memstore.block.multiplier
    8//内存充足确保不会产生OOM的情况下，调大此值
  
  
    hbase.server.thread.wakefrequency
    100//减少睡眠等待时间，默认值为10000