hadoop loadBalance源码分析

项目hbase数据库出现很诡异的assignment ，region移动的src和dest都是同一台regionserver，不过时间戳不同，启动的只有一个regionserver， 不知道怎么出现了两个时间戳

分析下源码解决一下 

loadbalance只有一个实现 org.apache.hadoop.hbase.master.DefaultLoadBalancer

在HMaster中会启动一个线程 org.apache.hadoop.hbase.Chore，按照设定的hbase.balancer.period（默认300000ms，五分钟），遍历所有表，根据每个表在regionserver中的region数量做balance，有一个平衡系数hbase.regions.slop（默认0.2），根据region总数算出平均region值，avg×0.8 取整作为最小值，avg×1.2取整作为最大值，regionserver上超过最大值要移走，小于最小值要移动region过来。否则打印目前的平衡状态。

 

assignmentManager 根据上述步骤生成的RegionPlan, 从src移动region到desc  src和desc都是ServerName对象

 

HMaster启动时会等待region servers注册到serverManager

// Wait for region servers to report in.
this.serverManager.waitForRegionServers(status);
// Check zk for regionservers that are up but didn't register
for (ServerName sn: this.regionServerTracker.getOnlineServers()) {
    if (!this.serverManager.isServerOnline(sn)) {
    // Not registered; add it.
       LOG.info("Registering server found up in zk but who has not yet " +
　　　　"reported in: " + sn);
　　　　this.serverManager.recordNewServer(sn, HServerLoad.EMPTY_HSERVERLOAD);
　　}
}

serverManager线程sleep一定时间，等待HRegionServer注册

HRegionServer.java:

      // Try and register with the Master; tell it we are here.  Break if
      // server is stopped or the clusterup flag is down or hdfs went wacky.
      while (keepLooping()) {
        MapWritable w = reportForDuty();
        if (w == null) {
          LOG.warn("reportForDuty failed; sleeping and then retrying.");
          this.sleeper.sleep();
        } else {
          handleReportForDutyResponse(w);
          break;
        }
      }

HRegionServer 注册之后进入mainloop

 // The main run loop.
 while (!this.stopped && isHealthy()) {
　　long now = System.currentTimeMillis();

　　　if ((now - lastMsg) >= msgInterval) {
　　　　　　doMetrics();
　　　　　　tryRegionServerReport();
　　　　　　lastMsg = System.currentTimeMillis();
　　　　}

  }

每隔hbase.regionserver.msginterval时间（默认3秒），进行一次注册尝试,如果服务器ip和端口不在已注册列表中，则添加ServerName进map

ServerManager.java

 void regionServerReport(ServerName sn, HServerLoad hsl)
  throws YouAreDeadException, PleaseHoldException {
    checkIsDead(sn, "REPORT");
    if (!this.onlineServers.containsKey(sn)) {
      // Already have this host+port combo and its just different start code?
      checkAlreadySameHostPort(sn);
      // Just let the server in. Presume master joining a running cluster.
      // recordNewServer is what happens at the end of reportServerStartup.
      // The only thing we are skipping is passing back to the regionserver
      // the ServerName to use. Here we presume a master has already done
      // that so we'll press on with whatever it gave us for ServerName.
      recordNewServer(sn, hsl);
    } else {
      this.onlineServers.put(sn, hsl);
    }
  }

recordNewServer 会打印 ServerName对象的ip 端口和时间戳信息

同一个region server注册的ServerName对象 会拥有同样的时间戳 

this.startcode = System.currentTimeMillis();

...

result = this.hbaseMaster.regionServerStartup(port, this.startcode, now);

...

this.serverNameFromMasterPOV = new ServerName(hostnameFromMasterPOV,
this.isa.getPort(), this.startcode);

...

this.hbaseMaster.regionServerReport(this.serverNameFromMasterPOV.getVersionedBytes(), hsl);

region server启动时startCode是固定死的，按照这个流程是不会出现相同IP和端口，但时间戳不同的region server跑在线上的 

如果一台机器上启动了两个region server 会把时间戳小的移出，下次添加进时间戳大的进去

我们遇到的问题是时间戳不同的regionserver被注册在了master上，并且相互之间做region move