Tachyon Cluster: 基于Zookeeper的Master High Availability（HA）高可用配置实现

1.Tachyon简介

Tachyon是一个高容错的分布式文件系统，允许文件以内存的速度在集群框架中进行可靠的共享，就像Spark和 MapReduce那样。通过利用信息继承，内存侵入，Tachyon获得了高性能。Tachyon工作集文件缓存在内存中，并且让不同的 Jobs/Queries以及框架都能内存的速度来访问缓存文件。因此，Tachyon可以减少那些需要经常使用的数据集通过访问磁盘来获得的次数。

2.Tachyon能解决什么问题：（摘自Tachyon 分布式内存文件系统）

1.不同FrameWork之间共享内存数据Slow问题
     给定一个场景，MapReduce任务的输出结果会存入到Tachyon里，Spark的Job会从Tachyon里读取MapReduce任务的输出作为输入。如果把Disk当作文件的落地，那么写的性能是十分低下的。但是如果把Memory当作落地，写的性能是非常高的，fast write，以便Spark Job不会感觉到这是2个计算框架的操作，因为写和读的速度都非常快。同样的，你可以用Impala的输出结果当作Spark的输入。
2.Spark的Executor Crash问题
     Spark的执行引擎和存储引擎都是在Executor进程里，即在一个Executor内会有多个Task在运行，并且这个Executor的内存会放入cache的RDD里。
问题来了，一旦我的Executor挂了，那么Tasks会失败，并且这些cache的RDD的Block也会丢失，这就会有ReCompute的过程，重新去取数据，根据血缘关系递归的去计算丢失的数据，这当然会很耗费资源，而且性能低下。
3.内存冗余问题
     这里说的内存冗余是说，Spark中不同Job之间可能同时读取了同一个文件，比如：job1和job2的计算任务都需要读取到账号信息表中的数据，那么我们都在他们各自的Executor里都cache了这一张账号表，是不是就出现了一个数据，2个内存副本，其实这样做是完全没有必要的，是冗余的。
4.GC时间过长
     有时候影响程序执行的并不是代码本身，而是由于内存中存了太多的Java Objects，如果Executor这个Jvm里cache的对象太多，比如：达到80G UP，这个时候出现几次FULL GC，你就会很纳闷我的程序怎么不动了？你去看GC log，原来在GC。

3.基于Zookeeper的Fault Tolerant Tachyon Cluster 实现

3.0 配置前提

• hadoop version：2.2.0.2.0.6.0-101
• zookeeper version：2.3.5
• Tachyon version: 0.4.1

集群情况：

Cluster	Masters	Slaves
Tachyon	bigdata001,bigdata002	bigdata001,bigdata002,bigdata003,bigdata004,bigdata005,bigdata006,bigdata007,bigdata008

zookeeper url: bigdata001:2181,bigdata002:2181,bigdata003:2181

3.1 HA架构

3.2 配置（conf/tachyon-env.sh ）

1.参考官方文档：Fault Tolerant Tachyon Cluster

①HDFS

export TACHYON_UNDERFS_ADDRESS=hdfs://[namenodeserver]:[namenodeport]

②ZooKeeper:

Property Name	Example	Meaning
tachyon.usezookeeper	true	Whether or not Master processes should use ZooKeeper.
tachyon.zookeeper.address	localhost:2181	The hostname and port ZooKeeper is running on.

③Master Node Configuration

export TACHYON_MASTER_ADDRESS=[externally visible address of this machine]

TACHYON_JAVA_OPTS to include:

-Dtachyon.master.journal.folder=hdfs://[namenodeserver]:[namenodeport]/tachyon/journal

④Worker Node Configuration

export TACHYON_MASTER_ADDRESS=[address of one of the master nodes in the system]

2.集群配置

Master节点配置：bigdata001节点的tachyon/conf/tachyon-env.sh添加如下（下划线部分）

export TACHYON_MASTER_ADDRESS=192.168.1.101
export TACHYON_UNDERFS_ADDRESS=hdfs://192.168.1.101:8020
export TACHYON_JAVA_OPTS+="
  -Dlog4j.configuration=file:$CONF_DIR/log4j.properties
  -Dtachyon.debug=false
  -Dtachyon.underfs.address=$TACHYON_UNDERFS_ADDRESS
  -Dtachyon.underfs.hdfs.impl=$TACHYON_UNDERFS_HDFS_IMPL
  -Dtachyon.data.folder=$TACHYON_UNDERFS_ADDRESS/tmp/tachyon/data
  -Dtachyon.workers.folder=$TACHYON_UNDERFS_ADDRESS/tmp/tachyon/workers
  -Dtachyon.worker.memory.size=$TACHYON_WORKER_MEMORY_SIZE
  -Dtachyon.worker.data.folder=$TACHYON_RAM_FOLDER/tachyonworker/
  -Dtachyon.master.worker.timeout.ms=
  -Dtachyon.master.hostname=$TACHYON_MASTER_ADDRESS
  -Dtachyon.master.journal.folder=$TACHYON_UNDERFS_ADDRESS/tachyon/journal/
  -Dtachyon.master.pinlist=/pinfiles;/pindata
  -Dorg.apache.jasper.compiler.disablejsr199=true
  -Dtachyon.user.default.block.size.byte=
  -Dtachyon.user.file.buffer.bytes=
  -Dtachyon.usezookeeper=true
  -Dtachyon.zookeeper.address=bigdata001:2181,bigdata002:2181,bigdata003:2181
"

配置同步到所有的slave节点：bigdata002,bigdata003,bigdata004,bigdata005,bigdata006,bigdata007,bigdata008

由于我们要将bigdata002作为另外一个master，因此，此节点的配置需要做修改TACHYON_MASTER_ADDRESS的值，如下

export TACHYON_MASTER_ADDRESS=192.168.1.102
export TACHYON_UNDERFS_ADDRESS=hdfs://192.168.1.101:8020
export TACHYON_JAVA_OPTS+="
  -Dlog4j.configuration=file:$CONF_DIR/log4j.properties
  -Dtachyon.debug=false
  -Dtachyon.underfs.address=$TACHYON_UNDERFS_ADDRESS
  -Dtachyon.underfs.hdfs.impl=$TACHYON_UNDERFS_HDFS_IMPL
  -Dtachyon.data.folder=$TACHYON_UNDERFS_ADDRESS/tmp/tachyon/data
  -Dtachyon.workers.folder=$TACHYON_UNDERFS_ADDRESS/tmp/tachyon/workers
  -Dtachyon.worker.memory.size=$TACHYON_WORKER_MEMORY_SIZE
  -Dtachyon.worker.data.folder=$TACHYON_RAM_FOLDER/tachyonworker/
  -Dtachyon.master.worker.timeout.ms=
  -Dtachyon.master.hostname=$TACHYON_MASTER_ADDRESS
  -Dtachyon.master.journal.folder=$TACHYON_UNDERFS_ADDRESS/tachyon/journal/
  -Dtachyon.master.pinlist=/pinfiles;/pindata
  -Dorg.apache.jasper.compiler.disablejsr199=true
  -Dtachyon.user.default.block.size.byte=
  -Dtachyon.user.file.buffer.bytes=
  -Dtachyon.usezookeeper=true
  -Dtachyon.zookeeper.address=bigdata001:,bigdata002:,bigdata003:
"

3.启动集群

[root@bigdata001 tachyon]# ./bin/tachyon-stop.sh

Killed  processes
Killed  processes
192.168.1.103: Killed  processes
192.168.1.101: Killed 0 processes
192.168.1.102: Killed  processes
192.168.1.104: Killed  processes
192.168.1.106: Killed  processes
192.168.1.105: Killed  processes
192.168.1.107: Killed  processes
192.168.1.108: Killed  processes

[root@bigdata001 tachyon]# ./bin/tachyon format

192.168.1.101: Formatting Tachyon Worker @ bigdata001
192.168.1.102: Formatting Tachyon Worker @ bigdata002
192.168.1.103: Formatting Tachyon Worker @ bigdata003
192.168.1.104: Formatting Tachyon Worker @ bigdata004
192.168.1.105: Formatting Tachyon Worker @ bigdata005
192.168.1.106: Formatting Tachyon Worker @ bigdata006
192.168.1.107: Formatting Tachyon Worker @ bigdata007
192.168.1.102: Removing local data under folder: /mnt/ramdisk/tachyonworker/
192.168.1.101: Removing local data under folder: /mnt/ramdisk/tachyonworker/
192.168.1.103: Removing local data under folder: /mnt/ramdisk/tachyonworker/
192.168.1.104: Removing local data under folder: /mnt/ramdisk/tachyonworker/
192.168.1.108: Formatting Tachyon Worker @ bigdata008
192.168.1.105: Removing local data under folder: /mnt/ramdisk/tachyonworker/
192.168.1.106: Removing local data under folder: /mnt/ramdisk/tachyonworker/
192.168.1.107: Removing local data under folder: /mnt/ramdisk/tachyonworker/
192.168.1.108: Removing local data under folder: /mnt/ramdisk/tachyonworker/
Formatting Tachyon Master @ 192.168.1.101
Formatting JOURNAL_FOLDER: hdfs://192.168.1.101:8020/tachyon/journal/
Formatting UNDERFS_DATA_FOLDER: hdfs://192.168.1.101:8020/tmp/tachyon/data
Formatting UNDERFS_WORKERS_FOLDER: hdfs://192.168.1.101:8020/tmp/tachyon/workers

[root@bigdata001 tachyon]# ./bin/tachyon-start.sh all Mount

Killed 0 processes
Killed 0 processes
192.168.1.103: Killed 0 processes
192.168.1.101: Killed 0 processes
192.168.1.105: Killed 0 processes
192.168.1.102: Killed 0 processes
192.168.1.107: Killed 0 processes
192.168.1.106: Killed 0 processes
192.168.1.104: Killed 0 processes
192.168.1.108: Killed 0 processes
Starting master @ 192.168.1.101
192.168.1.101: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.102: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.101: Starting worker @ bigdata001
192.168.1.103: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.102: Starting worker @ bigdata002
192.168.1.103: Starting worker @ bigdata003
192.168.1.104: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.105: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.104: Starting worker @ bigdata004
192.168.1.105: Starting worker @ bigdata005
192.168.1.106: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.106: Starting worker @ bigdata006
192.168.1.107: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.107: Starting worker @ bigdata007
192.168.1.108: Formatting RamFS: /mnt/ramdisk (2gb)
192.168.1.108: Starting worker @ bigdata008

[root@bigdata001 tachyon]# jps

可以看到Master和Worker的进程号：8315 Master  8458 Worker

在另个master节点bigdata002上启动另外一个master：

[root@bigdata002 tachyon]# ./bin/tachyon-start.sh master

Starting master @ 192.168.1.102

4.测试HA

web界面查看：http://bigdata001:19999

kill掉bigdata001的master进程，切换时间大概需要20s，再次查看新的Web UI：http://bigdata002:19999/home

5.Zk上查看

[root@bigdata002 conf]# zkCli.sh

[zk: localhost:2181(CONNECTED) 61] ls /election
[_c_ae6213f4-a2e3-46f9-8fc0-5c5c64d7e773-lock-0000000027, _c_12297d87-56fc-4cd9-8f8d-7312a6af4cc2-lock-0000000026]

[zk: localhost:2181(CONNECTED) 63] ls /leader
[bigdata001:19998, bigdata002:19998]

6.遇到的问题及解决

1 ./bin/tachyon format 报错

Formatting Tachyon Master @ 0.0.0.0
Exception in thread "main" java.lang.RuntimeException: java.lang.ClassNotFoundException: Class  not found
        at org.apache.hadoop.conf.Configuration.getClass(Configuration.java:1720)
        at org.apache.hadoop.fs.FileSystem.getFileSystemClass(FileSystem.java:2415)
        at org.apache.hadoop.fs.FileSystem.createFileSystem(FileSystem.java:2428)
        at org.apache.hadoop.fs.FileSystem.access$200(FileSystem.java:88)
        at org.apache.hadoop.fs.FileSystem$Cache.getInternal(FileSystem.java:2467)
        at org.apache.hadoop.fs.FileSystem$Cache.get(FileSystem.java:2449)
        at org.apache.hadoop.fs.FileSystem.get(FileSystem.java:367)
        at org.apache.hadoop.fs.Path.getFileSystem(Path.java:287)
        at tachyon.UnderFileSystemHdfs.<init>(UnderFileSystemHdfs.java:69)
        at tachyon.UnderFileSystemHdfs.getClient(UnderFileSystemHdfs.java:53)
        at tachyon.UnderFileSystem.get(UnderFileSystem.java:53)
        at tachyon.Format.main(Format.java:42)
Caused by: java.lang.ClassNotFoundException: Class  not found
        at org.apache.hadoop.conf.Configuration.getClassByName(Configuration.java:1626)
        at org.apache.hadoop.conf.Configuration.getClass(Configuration.java:1718)
        ... 11 more

原因及解决：配置文件tachyon/conf/tachyon-env.sh的此项未配置：export TACHYON_UNDERFS_HDFS_IMPL=org.apache.hadoop.hdfs.DistributedFileSystem