HBase client访问ZooKeeper获取rootregionserver DeadLock问题（zookeeper.ClientCnxn Unable to get data of znode /hbase/rootregionserver）

2012年11月28日 出现故障，" Unable to get data of znode /hbase/root-region-server"

问题比较诡异，两个机房，只有一个机房故障，5台服务器相续故障，错误日志相同。使用的HBase客户端版本为0.94.0

1）分析步骤：

1 jstack jmap 查看是否有死锁、block或内存溢出

jmap 看内存回收状况没有什么异常，内存和CPU占用都不多

 jstack pid > test.log
pid: Unable to open socket file: target process not responding or HotSpot VM not loaded
The -F option can be used when the target process is not responding

出现这种错误，使用-F参数
jstack -l -F pid >jstack.log

jstack -l pid >jstack.log

这时jstack.log并不会有有用信息，要去catalina.out里看，注意在服务下线后使用，容易造成tomcat僵死
catalina.out里发现deadlock死锁

Found one Java-level deadlock:
=============================
"catalina-exec-800":
  waiting to lock monitor 0x000000005f1f6530 (object 0x0000000731902200, a java.lang.Object),
  which is held by "catalina-exec-710"
"catalina-exec-710":
  waiting to lock monitor 0x00002aaab9a05bd0 (object 0x00000007321f8708, a org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation),
  which is held by "catalina-exec-29-EventThread"
"catalina-exec-29-EventThread":
  waiting to lock monitor 0x000000005f9f0af0 (object 0x0000000732a9c7e0, a org.apache.hadoop.hbase.zookeeper.RootRegionTracker),
  which is held by "catalina-exec-710"
Java stack information for the threads listed above:
===================================================
"catalina-exec-800":
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegionInMeta(HConnectionManager.java:943)
        - waiting to lock <0x0000000731902200> (a java.lang.Object)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegion(HConnectionManager.java:836)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.relocateRegion(HConnectionManager.java:807)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.getRegionLocation(HConnectionManager.java:725)
        at org.apache.hadoop.hbase.client.ServerCallable.connect(ServerCallable.java:82)
        at org.apache.hadoop.hbase.client.ServerCallable.withRetries(ServerCallable.java:162)
        at org.apache.hadoop.hbase.client.HTable.get(HTable.java:685)
        at org.apache.hadoop.hbase.client.HTablePool$PooledHTable.get(HTablePool.java:366)
        at com.weibo.api.commons.hbase.CustomHBase.get(CustomHBase.java:171)
        at com.weibo.api.commons.hbase.CustomHBase.get(CustomHBase.java:160)
        at com.weibo.api.commons.hbase.CustomHBase.get(CustomHBase.java:150)

"catalina-exec-710":
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.resetZooKeeperTrackers(HConnectionManager.java:599)
        - waiting to lock <0x00000007321f8708> (a org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.abort(HConnectionManager.java:1660)
        at org.apache.hadoop.hbase.zookeeper.ZooKeeperNodeTracker.getData(ZooKeeperNodeTracker.java:158)
        - locked <0x0000000732a9c7e0> (a org.apache.hadoop.hbase.zookeeper.RootRegionTracker)
        at org.apache.hadoop.hbase.zookeeper.RootRegionTracker.getRootRegionLocation(RootRegionTracker.java:62)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegion(HConnectionManager.java:821)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegion(HConnectionManager.java:801)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegionInMeta(HConnectionManager.java:933)
        ......
        at org.apache.hadoop.hbase.client.MetaScanner.metaScan(MetaScanner.java:123)
        at org.apache.hadoop.hbase.client.MetaScanner.metaScan(MetaScanner.java:99)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.prefetchRegionCache(HConnectionManager.java:894)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegionInMeta(HConnectionManager.java:948)
        - locked <0x0000000731902200> (a java.lang.Object)
  
 "catalina-exec-29-EventThread":
        at org.apache.hadoop.hbase.zookeeper.ZooKeeperNodeTracker.stop(ZooKeeperNodeTracker.java:98)
        - waiting to lock <0x0000000732a9c7e0> (a org.apache.hadoop.hbase.zookeeper.RootRegionTracker)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.resetZooKeeperTrackers(HConnectionManager.java:604)
        - locked <0x00000007321f8708> (a org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation)
        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.abort(HConnectionManager.java:1660)
        at org.apache.hadoop.hbase.zookeeper.ZooKeeperWatcher.connectionEvent(ZooKeeperWatcher.java:374)
        at org.apache.hadoop.hbase.zookeeper.ZooKeeperWatcher.process(ZooKeeperWatcher.java:271)
        at org.apache.zookeeper.ClientCnxn$EventThread.processEvent(ClientCnxn.java:521)
        at org.apache.zookeeper.ClientCnxn$EventThread.run(ClientCnxn.java:497)
Found 1 deadlock.

 warn.log中报告Interrupted异常 是由上述死锁引起的

2012-11-28 20:06:17 [WARN] hconnection-0x4384d0a47f41c63 Unable to get data of znode /hbase/root-region-server

java.lang.InterruptedException

        at java.lang.Object.wait(Native Method)

        at java.lang.Object.wait(Object.java:485)

        at org.apache.zookeeper.ClientCnxn.submitRequest(ClientCnxn.java:1253)

        at org.apache.zookeeper.ZooKeeper.getData(ZooKeeper.java:1129)

        at org.apache.hadoop.hbase.zookeeper.RecoverableZooKeeper.getData(RecoverableZooKeeper.java:264)

        at org.apache.hadoop.hbase.zookeeper.ZKUtil.getDataInternal(ZKUtil.java:522)

        at org.apache.hadoop.hbase.zookeeper.ZKUtil.getDataAndWatch(ZKUtil.java:498)

        at org.apache.hadoop.hbase.zookeeper.ZooKeeperNodeTracker.getData(ZooKeeperNodeTracker.java:156)

        at org.apache.hadoop.hbase.zookeeper.RootRegionTracker.getRootRegionLocation(RootRegionTracker.java:62)

        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegion(HConnectionManager.java:821)

        ......

        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.relocateRegion(HConnectionManager.java:807)

        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegionInMeta(HConnectionManager.java:1042)

        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.locateRegion(HConnectionManager.java:836)

        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.relocateRegion(HConnectionManager.java:807)

        at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.getRegionLocation(HConnectionManager.java:725)

        at org.apache.hadoop.hbase.client.ServerCallable.connect(ServerCallable.java:82)

        at org.apache.hadoop.hbase.client.ServerCallable.withRetries(ServerCallable.java:162)

        at org.apache.hadoop.hbase.client.HTable.get(HTable.java:685)

        at org.apache.hadoop.hbase.client.HTablePool$PooledHTable.get(HTablePool.java:366)

      

2012-11-28 20:06:17 [ERROR] [hbase_error]

java.io.IOException: Giving up after tries=1

        at org.apache.hadoop.hbase.client.ServerCallable.withRetries(ServerCallable.java:192)

        at org.apache.hadoop.hbase.client.HTable.get(HTable.java:685)

        at org.apache.hadoop.hbase.client.HTablePool$PooledHTable.get(HTablePool.java:366)

 

Caused by: java.lang.InterruptedException: sleep interrupted

        at java.lang.Thread.sleep(Native Method)

        at org.apache.hadoop.hbase.client.ServerCallable.withRetries(ServerCallable.java:189)

        ... 13 more

https://issues.apache.org/jira/browse/HBASE-5060  HBase client is blocked forever ，跟这个问题有点相似，但没有解决这问题

2 同进程另启动线程访问root-region-server

                    // check if zk is ok
                    ZooKeeper zk = null;
                    Watcher watch = new Watcher() {
                        public void process(WatchedEvent event) {
                        }
                    };
                    String zookeeperQuorum = wbobjectHBaseConfMap.get("hbase.zookeeper.quorum");
                    
                    if (StringUtils.isNotBlank(zookeeperQuorum)) {
                        try {
                            zk = new ZooKeeper(zookeeperQuorum, 30000, watch);
                            byte[] data = zk.getData("/hbase/root-region-server", watch, null);
                            ApiLogger.info(" get root-region-server success! ip:"
                                    + Util.toStr(data));
                        } catch (Exception e) {
                            ApiLogger.error(" get root-region-server error!" + e.getMessage());
                        } finally {
                            try {
                                zk.close();
                            } catch (InterruptedException e) {
                                ApiLogger.error("close zk error!");
                            }
                        }
                    }
                    

发现独立线程在整个进程死锁时还能正常工作，HBase的zookeeper实例异常死锁后就不能恢复，导致scan操作超时到30s+，并且不能恢复，正常应该在ms级别。

因此认为是HBase客户端连接ZooKeeper时出问题，流程：

出现网络抖动或者root表迁移，缓存表未命中，重新去获取root-region-server，结果获取失败，进行ZooKeeper重置操作

经过研究死锁代码

HConnectionImplementation 发现 get root-region-server异常，锁定RootRegionTracker，试图锁定自身，执行resetZooKeeperTrackers并调用RootRegionTracker.stop 去重置 ZooKeeper

而与此同时RootRegionTracker在执行stop方法时，已经锁定了HConnectionImplementation，还需要锁定它本身（synchronized方法），但其已经在HConnectionImplementation中被锁定。

HConnectionImplementation代码中：

 this.rootRegionTracker = new RootRegionTracker(this.zooKeeper, this);

调用时也没有超时cancel代码 这样就导致了两个资源循环等待的死锁

2）仿真模拟 

以下是个小程序，简单模拟这种死锁问题：

import java.util.concurrent.TimeUnit;

import org.apache.hadoop.hbase.ServerName;

class ZooKeeperNodeTracker {
    private boolean stopped = false;
    private AbortAble hConnectionImplementation;
    public ZooKeeperNodeTracker(AbortAble hConnectionImplementation) {
        this.hConnectionImplementation = hConnectionImplementation;
    }

    public synchronized void stop()  throws InterruptedException {
        this.stopped = true;
        System.out.println(Thread.currentThread()+"|"+Thread.currentThread().getId()+"stop zknode");
        TimeUnit.MICROSECONDS.sleep(100);
        notifyAll();
    }

    public boolean condition() {
        return stopped;
    }

    public boolean start() {
         stopped=false;
         return true;
    }
    
    public synchronized boolean getData(int i) throws InterruptedException {
        //error in get root region server
        if (i %100 ==0){
            hConnectionImplementation.resetZooKeeperTrackers();
            throw new InterruptedException("interrupted");            
        }
        return true;
    }

}

public class HConnectionManagerTest {

    static class HConnectionImplementation implements AbortAble{
        public HConnectionImplementation() {
            rootRegionTracker = new ZooKeeperNodeTracker(this);
        }

        private volatile ZooKeeperNodeTracker rootRegionTracker;
        @Override
        public synchronized void resetZooKeeperTrackers() {
            try{
                if (rootRegionTracker != null) {
                    rootRegionTracker.stop();
                    rootRegionTracker = null;
                    System.out.println(Thread.currentThread()+"|"+Thread.currentThread().getId()+"resetZooKeeperTrackers");
                }
            }catch(InterruptedException e){
                System.out.println(Thread.currentThread()+"----------resetZooKeeperTrackers Interrupted-----------");
                
            }
        }


        public void testGetData(String name) {
            int i = 1;
            while (i >0) {
                i++;
                try {
                    rootRegionTracker.getData(i);
                } catch (Exception e) {
                    resetZooKeeperTrackers();
                }
                if(i %100 ==0){
                    rootRegionTracker = new ZooKeeperNodeTracker(this);
                    System.out.println(name+" restart test");
                }
            }
        }
    }

    public static void main(String[] args)  {

        final HConnectionImplementation hcon = new HConnectionImplementation();
        Thread scan1 = new Thread(new Runnable() {
            public void run() {
                hcon.testGetData("[scan1]");
            }
        });
        
        Thread scan2 = new Thread(new Runnable() {
            public void run() {
                hcon.testGetData("[scan2]");
            }
        });
        try{
            scan1.start();
            scan2.start();
            TimeUnit.SECONDS.sleep(2);
        }catch (InterruptedException e){
            System.out.println("----------testgetdata -------interrupt");
        }
    }

}

类名构造函数都模拟HBase Client，并且放大getData error的情形，当同时并发两个scan操作时，前一个scan过程中，获取不到root-region-server，在ZooKeeperNodeTracker中做stop() 时，后一个scan也开始在HConnectionImplementation中执行resetZooKeeperTrackers（），两个资源ZooKeeperNodeTracker和HConnectionImplementation被各自分别占用等待，导致死锁。模拟程序的死锁解除可以更改

public synchronized boolean getData

方法为

public boolean getData

或

public synchronized void resetZooKeeperTrackers() {
            try{
                if (rootRegionTracker != null) {
                    rootRegionTracker.stop();

为

public void resetZooKeeperTrackers() {
            try{
                if (rootRegionTracker != null) {
　　　　　　　　　　　　synchronized(rootRegionTracker){
　　　　　　　　　　　　　　rootRegionTracker.stop();

解除互斥条件解决问题

 3）最终解决方案

通过Hadoop大会现场跟HBase开发者Ted Yu咨询，称0.94.0有很多bug不稳定，建议升级到0.94.2，通过查看relase note， 官方的两个patch地址已在0.94.1中修复 （通过对hbase源码进行分析找对问题对应点，再看对应源码svn详细修改记录）
1 通过避免嵌套重试循环来解决rpc线程卡死： https://issues.apache.org/jira/browse/HBASE-6326  .
2 通过等待-root-的region地址设置到root region tracker 来避免deadlock问题：https://issues.apache.org/jira/browse/HBASE-6115