Hadoop生成HFile直接入库HBase心得

转载请标明出处：http://blackwing.iteye.com/blog/1991380 


hbase自带了ImportTsv类，可以直接把tsv格式（官方教材显示，是\t分割各个字段的文本格式）生成HFile，并且使用另外一个类org.apache.hadoop.hbase.mapreduce.LoadIncrementalHFiles直接把HFile移动到hbase对应的hdfs目录。 

PS:网上看到一个XD说，直接生成HFile并入库HBase效率不如先生成HFile，再通过LoadIncrementalHFiles移动文件到hbase目录高，这点没有验证，我的做法也是先生成，再move。 

官方教材在此： 

 http://hbase.apache.org/book/ops_mgt.html#importtsv

但ImportTsv功能对我来说不适合，例如文件格式为： 

 topsid   uid   roler_num   typ        time
 10      111111   255         0       1386553377000

ImportTsv导入的命令为： 

 bin/hbase org.apache.hadoop.hbase.mapreduce.ImportTsv -Dimporttsv.columns=HBASE_ROW_KEY,kq:topsid,kq:uid,kq:roler_num,kq:type -Dimporttsv.bulk.output=hdfs://storefile-outputdir

它生成的表格式为： 

 row :　10
 cf  :  kq
 qualifier: topsid
 value: 10
 .....

而我要求的格式是： 

 row :　10-111111-255
 cf  :  kq
 qualifier: 0
 value: 1

所以还是自己写MR处理数据方便。 
Mapper： 

 /*
  * adminOnOff.log 文件格式：
  * topsid   uid   roler_num   typ   time
  * */
 public class HFileImportMapper2 extends
         Mapper {
     protected SimpleDateFormat sdf = new SimpleDateFormat("yyyyMMdd");
     protected final String CF_KQ="kq";//考勤
     protected final int ONE=1;
     @Override
     protected void map(LongWritable key, Text value,Context context)
             throws IOException, InterruptedException {
         String line = value.toString();
         System.out.println("line : "+line);
         String[] datas = line.split("\\s+");
         // row格式为:yyyyMMdd-sid-uid-role_num-timestamp-typ
         String row = sdf.format(new Date(Long.parseLong(datas[4])))
                 + "-" + datas[0] + "-" + datas[1] + "-" + datas[2]
                 + "-" + datas[4] + "-" + datas[3];
         ImmutableBytesWritable rowkey = new ImmutableBytesWritable(
                 Bytes.toBytes(row));
         KeyValue kv = new KeyValue(Bytes.toBytes(row),this.CF_KQ.getBytes(), datas[3].getBytes(),Bytes.toBytes(this.ONE));
         context.write(rowkey, kv);
         }
 }

job: 

 public class GenHFile2 {
     public static void main(String[] args) {
         Configuration conf = new Configuration();
         conf.addResource("myConf.xml");
         String input = conf.get("input");
         String output = conf.get("output");
         String tableName = conf.get("source_table");
         System.out.println("table : "+tableName);
         HTable table;
         try {
             //运行前，删除已存在的中间输出目录
             try {
                 FileSystem fs = FileSystem.get(URI.create(output), conf);
                 fs.delete(new Path(output),true);
                 fs.close();
             } catch (IOException e1) {
                 e1.printStackTrace();
             }

             table = new HTable(conf,tableName.getBytes());
             Job job = new Job(conf);
             job.setJobName("Generate HFile");

             job.setJarByClass(HFileImportMapper2.class);
             job.setInputFormatClass(TextInputFormat.class);
             job.setMapperClass(HFileImportMapper2.class);
             FileInputFormat.setInputPaths(job, input);

 //job.setReducerClass(KeyValueSortReducer.class);
 //job.setMapOutputKeyClass(ImmutableBytesWritable.class);
 //job.setMapOutputValueClass(KeyValue.class);
             job.getConfiguration().set("mapred.mapoutput.key.class", "org.apache.hadoop.hbase.io.ImmutableBytesWritable");
             job.getConfiguration().set("mapred.mapoutput.value.class", "org.apache.hadoop.hbase.KeyValue");

 //job.setOutputFormatClass(HFileOutputFormat.class);
 FileOutputFormat.setOutputPath(job, new Path(output));
     //job.setPartitionerClass(SimpleTotalOrderPartitioner.class);
 HFileOutputFormat.configureIncrementalLoad(job,table);
             try {
                 job.waitForCompletion(true);
             } catch (InterruptedException e) {
                 e.printStackTrace();
             } catch (ClassNotFoundException e) {
                 e.printStackTrace();
             }
         } catch (IOException e) {
             e.printStackTrace();
         }
     }
 }

生成的HFile文件在hdfs的/output目录下，已经根据cf名称建好文件目录： 

 hdfs://namenode/output/kq/601c5029fb264dc8869a635043c24560

其中： 

 HFileOutputFormat.configureIncrementalLoad(job,table);

根据其源码知道，会自动为job设置好以下参数： 

 public static void configureIncrementalLoad(Job job, HTable table)
 throws IOException {
   Configuration conf = job.getConfiguration();

   job.setOutputKeyClass(ImmutableBytesWritable.class);
   job.setOutputValueClass(KeyValue.class);
   job.setOutputFormatClass(HFileOutputFormat.class);

   // Based on the configured map output class, set the correct reducer to properly
   // sort the incoming values.
   // TODO it would be nice to pick one or the other of these formats.
   if (KeyValue.class.equals(job.getMapOutputValueClass())) {
     job.setReducerClass(KeyValueSortReducer.class);
   } else if (Put.class.equals(job.getMapOutputValueClass())) {
     job.setReducerClass(PutSortReducer.class);
   } else if (Text.class.equals(job.getMapOutputValueClass())) {
     job.setReducerClass(TextSortReducer.class);
   } else {
     LOG.warn("Unknown map output value type:" + job.getMapOutputValueClass());
   }

   conf.setStrings("io.serializations", conf.get("io.serializations"),
       MutationSerialization.class.getName(), ResultSerialization.class.getName(),
       KeyValueSerialization.class.getName());

   // Use table's region boundaries for TOP split points.
   LOG.info("Looking up current regions for table " + Bytes.toString(table.getTableName()));
   List startKeys = getRegionStartKeys(table);
   LOG.info("Configuring " + startKeys.size() + " reduce partitions " +
       "to match current region count");
   job.setNumReduceTasks(startKeys.size());

   configurePartitioner(job, startKeys);
   // Set compression algorithms based on column families
   configureCompression(table, conf);
   configureBloomType(table, conf);
   configureBlockSize(table, conf);

   TableMapReduceUtil.addDependencyJars(job);
   TableMapReduceUtil.initCredentials(job);
   LOG.info("Incremental table " + Bytes.toString(table.getTableName()) + " output configured.");
 }

HFileOutputFormat只支持写单个column family，如果有多个cf，则需要写多个job来实现了。