hadoop3.x 新特性

Features	Hadoop 2.x	Hadoop 3.x
Minimum Required Java Version	JDK 6 and above.	JDK 8 is the minimum runtime version of JAVA required to run Hadoop 3.x as many dependency library files have been used from JDK 8.
Fault Tolerance	Fault Tolerance is handled through replication leading to storage and network bandwidth overhead.(3个副本)	Support for Erasure Coding（纠错码） in HDFS improves fault tolerance （0.5纠错码+1数据 = 1.5倍磁盘占用）
Storage Scheme	Follows a 3x Replication Scheme for data recovery leading to 200% storage overhead. For instance, if there are 8 data blocks then a total of 24 blocks will occupy the storage space because of the 3x replication scheme.	Storage overhead in Hadoop 3.0 is reduced to 50% with support for Erasure Coding. In this case, if here are 8 data blocks then a total of only 12 blocks will occupy the storage space.
Change in Port Numbers	Hadoop HDFS NameNode -8020 Hadoop HDFS DataNode -50010 Secondary NameNode HTTP -50091	Hadoop HDFS NameNode -9820 Hadoop HDFS DataNode -9866 Secondary NameNode HTTP -9869
YARN Timeline Service	YARN timeline service introduced in Hadoop 2.0 has some scalability issues.	YARN Timeline service has been enhanced with ATS v2 which improves the scalability and reliability.
Intra DataNode Balancing	HDFS Balancer in Hadoop 2.0 caused skew within a DataNode because of addition or replacement of disks.	Intra DataNode Balancing has been introduced in Hadoop 3.0 to address the intra-DataNode skews which occur when disks are added or replaced.
Number of NameNodes	Hadoop 2.0 introduced a secondary namenode as standby.(一主一备)	Hadoop 3.0 supports 2 or more NameNodes.（一主多备）
Heap Size	In Hadoop 2.0 , for Java and Hadoop tasks, the heap size needs to be set through two similar properties mapreduce.{map,reduce}.java. Opts and mapreduce.{map,reduce}.memory.mb	In Hadoop 3.0, heap size or mapreduce.*.memory.mb is derived automatically.

hdfs HA 逻辑

1. 增加用于主备之间信息共享推送的 JournalNode
JournalNode 是 hadoop 根据 paxos 协议实现的日志服务
2. 增加用于选主决策的 zookeeper 集群：ha.zookeeper.quorum 配置
3. 增加用于监控同机器上的 namenode，试图选举，切换本地 namenode 的 active,standby 状态的zookeeper failover controller（zkfc）进程：QuorumPeerMain