Hadoop World 大會上的文章選摘（一）

Compression

         LZOor Snappy(20% better than LZO)——Block(default)

 

Serialization

         AVROdidn’t work well——deserialization issue

         Developedconfigurable serialization mechanism that uses JSON except Data type

 

Secondary Indexes

         Wereusing ITHBase and IHBase from contrib——doesn’t work well

         Redesignedschema without need for index

         Westill need it though

 

Performance

         Severaltunable parameters

        

Hardware(Hadoop+HBase)

         DataNode -24GB RAM,8Cores,4*1TB(64GB,24Cores,8*2TB)

         6mappers and 6 reducers per node (16 mappers,4reducers)

         Memoryallocation by process

         DataNode——1GB(2GB)

         TaskTracker——1GB(2GB)

         MapTasks——6*1GB(16*1.5GB)

         ReduceTasks——6*1GB(4*1.5GB)

         RegionServer——8GB(24GB)

         TotalAllocation——24GB(64GB)

 

Deployment

         Donot run ZK instances on DN,have a separate ZK quorum(3 minimum)

         Donot run HMaster on NN

         AvoidSPOF for HMaster(run additional master(s))