HBase高速导入数据--BulkLoad

Apache HBase是一个分布式的、面向列的开源数据库。它能够让我们随机的、实时的訪问大数据。可是如何有效的将数据导入到HBase呢？HBase有多种导入数据的方法。最直接的方法就是在MapReduce作业中使用TableOutputFormat作为输出。或者使用标准的客户端API，可是这些都不是很有效的方法。

Bulkload利用MapReduce作业输出HBase内部数据格式的表数据，然后将生成的StoreFiles直接导入到集群中。

与使用HBase API相比，使用Bulkload导入数据占用更少的CPU和网络资源。

Bulkload过程主要包含三部分：

1.从数据源(一般是文本文件或其它的数据库)提取数据并上传到HDFS

这一步不在HBase的考虑范围内，无论数据源是什么，仅仅要在进行下一步之前将数据上传到HDFS就可以。

2.利用一个MapReduce作业准备数据

这一步须要一个MapReduce作业。而且大多数情况下还须要我们自己编写Map函数，而Reduce函数不须要我们考虑。由HBase提供。

该作业须要使用rowkey(行键)作为输出Key，KeyValue、Put或者Delete作为输出Value。MapReduce作业须要使用HFileOutputFormat2来生成HBase数据文件。为了有效的导入数据，须要配置HFileOutputFormat2使得每个输出文件都在一个合适的区域中。为了达到这个目的，MapReduce作业会使用Hadoop的TotalOrderPartitioner类依据表的key值将输出切割开来。

HFileOutputFormat2的方法configureIncrementalLoad()会自己主动的完毕上面的工作。

3.告诉RegionServers数据的位置并导入数据

这一步是最简单的，通常须要使用LoadIncrementalHFiles(更为人所熟知是completebulkload工具)。将文件在HDFS上的位置传递给它，它就会利用RegionServer将数据导入到对应的区域。

下图简单明白的说明了整个过程

图片来自How-to: Use HBase Bulk Loading, and Why

Note:在进行BulkLoad之前，要在HBase中创建与程序中同名且结构同样的空表

Java实现例如以下：

BulkLoadDriver.java

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.conf.Configured;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.hbase.HBaseConfiguration;
import org.apache.hadoop.hbase.TableName;
import org.apache.hadoop.hbase.client.Connection;
import org.apache.hadoop.hbase.client.ConnectionFactory;
import org.apache.hadoop.hbase.client.Put;
import org.apache.hadoop.hbase.io.ImmutableBytesWritable;
import org.apache.hadoop.hbase.mapreduce.HFileOutputFormat2;
import org.apache.hadoop.mapreduce.Job;
import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
import org.apache.hadoop.mapreduce.lib.input.TextInputFormat;
import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
import org.apache.hadoop.util.Tool;
import org.apache.hadoop.util.ToolRunner;

/**
 * Created by shaobo on 15-6-9.
 */
public class BulkLoadDriver extends Configured implements Tool {
    private static final String DATA_SEPERATOR = "\\s+";
    private static final String TABLE_NAME = "temperature";//表名
    private static final String COLUMN_FAMILY_1="date";//列组1
    private static final String COLUMN_FAMILY_2="tempPerHour";//列组2

    public static void main(String[] args) {
        try {
            int response = ToolRunner.run(HBaseConfiguration.create(), new BulkLoadDriver(), args);
            if(response == 0) {
                System.out.println("Job is successfully completed...");
            } else {
                System.out.println("Job failed...");
            }
        } catch(Exception exception) {
            exception.printStackTrace();
        }
    }

    public int run(String[] args) throws Exception {
        String outputPath = args[1];
        /**
         * 设置作业參数
         */
        Configuration configuration = getConf();
        configuration.set("data.seperator", DATA_SEPERATOR);
        configuration.set("hbase.table.name", TABLE_NAME);
        configuration.set("COLUMN_FAMILY_1", COLUMN_FAMILY_1);
        configuration.set("COLUMN_FAMILY_2", COLUMN_FAMILY_2);
        Job job = Job.getInstance(configuration, "Bulk Loading HBase Table::" + TABLE_NAME);
        job.setJarByClass(BulkLoadDriver.class);
        job.setInputFormatClass(TextInputFormat.class);
        job.setMapOutputKeyClass(ImmutableBytesWritable.class);//指定输出键类
        job.setMapOutputValueClass(Put.class);//指定输出值类
        job.setMapperClass(BulkLoadMapper.class);//指定Map函数
        FileInputFormat.addInputPaths(job, args[0]);//输入路径
        FileSystem fs = FileSystem.get(configuration);
        Path output = new Path(outputPath);
        if (fs.exists(output)) {
            fs.delete(output, true);//假设输出路径存在，就将其删除
        }
        FileOutputFormat.setOutputPath(job, output);//输出路径
        Connection connection = ConnectionFactory.createConnection(configuration);
        TableName tableName = TableName.valueOf(TABLE_NAME);
        HFileOutputFormat2.configureIncrementalLoad(job, connection.getTable(tableName), connection.getRegionLocator(tableName));
        job.waitForCompletion(true);
        if (job.isSuccessful()){
            HFileLoader.doBulkLoad(outputPath, TABLE_NAME);//导入数据
            return 0;
        } else {
            return 1;
        }
    }

}

BulkLoadMapper.java

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.hbase.client.Put;
import org.apache.hadoop.hbase.io.ImmutableBytesWritable;
import org.apache.hadoop.hbase.util.Bytes;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapreduce.Mapper;

/**
 * Created by shaobo on 15-6-9.
 */
public class BulkLoadMapper extends Mapper<LongWritable, Text, ImmutableBytesWritable, Put> {
    private String hbaseTable;
    private String dataSeperator;
    private String columnFamily1;
    private String columnFamily2;

    public void setup(Context context) {
        Configuration configuration = context.getConfiguration();//获取作业參数
        hbaseTable = configuration.get("hbase.table.name");
        dataSeperator = configuration.get("data.seperator");
        columnFamily1 = configuration.get("COLUMN_FAMILY_1");
        columnFamily2 = configuration.get("COLUMN_FAMILY_2");
    }

    public void map(LongWritable key, Text value, Context context){
        try {
            String[] values = value.toString().split(dataSeperator);
            ImmutableBytesWritable rowKey = new ImmutableBytesWritable(values[0].getBytes());
            Put put = new Put(Bytes.toBytes(values[0]));
            put.addColumn(Bytes.toBytes(columnFamily1), Bytes.toBytes("month"), Bytes.toBytes(values[1]));
            put.addColumn(Bytes.toBytes(columnFamily1), Bytes.toBytes("day"), Bytes.toBytes(values[2]));
            for (int i = 3; i < values.length; ++i){
                put.addColumn(Bytes.toBytes(columnFamily2), Bytes.toBytes("hour : " + i), Bytes.toBytes(values[i]));
            }
            context.write(rowKey, put);
        } catch(Exception exception) {
            exception.printStackTrace();
        }

    }

}

HFileLoader.java

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.hbase.HBaseConfiguration;
import org.apache.hadoop.hbase.client.HTable;
import org.apache.hadoop.hbase.mapreduce.LoadIncrementalHFiles;

/**
 * Created by shaobo on 15-6-9.
 */
public class HFileLoader {
    public static void doBulkLoad(String pathToHFile, String tableName){
        try {
            Configuration configuration = new Configuration();
            HBaseConfiguration.addHbaseResources(configuration);
            LoadIncrementalHFiles loadFfiles = new LoadIncrementalHFiles(configuration);
            HTable hTable = new HTable(configuration, tableName);//指定表名
            loadFfiles.doBulkLoad(new Path(pathToHFile), hTable);//导入数据
            System.out.println("Bulk Load Completed..");
        } catch(Exception exception) {
            exception.printStackTrace();
        }

    }

}

将程序编译打包，提交到Hadoop执行

HADOOP_CLASSPATH=$(hbase mapredcp):/path/to/hbase/conf hadoop jar BulkLoad.jar inputpath outputpath

上述命令使用方法可參考44. HBase, MapReduce, and the CLASSPATH

作业执行情况：

15/06/14 14:31:07 INFO mapreduce.HFileOutputFormat2: Looking up current regions for table temperature(表名)
15/06/14 14:31:07 INFO mapreduce.HFileOutputFormat2: Configuring 1 reduce partitions to match current region count
15/06/14 14:31:07 INFO mapreduce.HFileOutputFormat2: Writing partition information to /home/shaobo/hadoop/tmp/partitions_5d464f1e-d412-4dbe-bb98-367f8431bdc9
15/06/14 14:31:07 INFO zlib.ZlibFactory: Successfully loaded & initialized native-zlib library
15/06/14 14:31:07 INFO compress.CodecPool: Got brand-new compressor [.deflate]
15/06/14 14:31:08 INFO mapreduce.HFileOutputFormat2: Incremental table temperature(表名) output configured.
15/06/14 14:31:08 INFO client.RMProxy: Connecting to ResourceManager at localhost/127.0.0.1:8032
15/06/14 14:31:15 INFO input.FileInputFormat: Total input paths to process : 2
15/06/14 14:31:15 INFO mapreduce.JobSubmitter: number of splits:2
15/06/14 14:31:16 INFO mapreduce.JobSubmitter: Submitting tokens for job: job_1434262360688_0002
15/06/14 14:31:17 INFO impl.YarnClientImpl: Submitted application application_1434262360688_0002
15/06/14 14:31:17 INFO mapreduce.Job: The url to track the job: http://shaobo-ThinkPad-E420:8088/proxy/application_1434262360688_0002/
15/06/14 14:31:17 INFO mapreduce.Job: Running job: job_1434262360688_0002
15/06/14 14:31:28 INFO mapreduce.Job: Job job_1434262360688_0002 running in uber mode : false
15/06/14 14:31:28 INFO mapreduce.Job:  map 0% reduce 0%
15/06/14 14:32:24 INFO mapreduce.Job:  map 49% reduce 0%
15/06/14 14:32:37 INFO mapreduce.Job:  map 67% reduce 0%
15/06/14 14:32:43 INFO mapreduce.Job:  map 100% reduce 0%
15/06/14 14:33:39 INFO mapreduce.Job:  map 100% reduce 67%
15/06/14 14:33:42 INFO mapreduce.Job:  map 100% reduce 70%
15/06/14 14:33:45 INFO mapreduce.Job:  map 100% reduce 88%
15/06/14 14:33:48 INFO mapreduce.Job:  map 100% reduce 100%
15/06/14 14:33:52 INFO mapreduce.Job: Job job_1434262360688_0002 completed successfully
...
...
...
15/06/14 14:34:02 WARN mapreduce.LoadIncrementalHFiles: Skipping non-directory hdfs://localhost:9000/user/output/_SUCCESS
15/06/14 14:34:03 INFO hfile.CacheConfig: CacheConfig:disabled
15/06/14 14:34:03 INFO hfile.CacheConfig: CacheConfig:disabled
15/06/14 14:34:07 INFO mapreduce.LoadIncrementalHFiles: Trying to load hfile=hdfs://localhost:9000/user/output/date/c64cd2524fba48738bab26630d550b61 first=AQW00061705 last=USW00094910
15/06/14 14:34:07 INFO mapreduce.LoadIncrementalHFiles: Trying to load hfile=hdfs://localhost:9000/user/output/tempPerHour/43af29456913444795a820544691eb3d first=AQW00061705 last=USW00094910
Bulk Load Completed..
Job is successfully completed...

BulLoad过程的第三步也能够在用MapReduce作业生成HBase数据文件后在命令行中进行，不一定要与MapReduce过程写在一起。

$ hadoop jar hbase-server-VERSION.jar completebulkload [-c /path/to/hbase/config/hbase-site.xml] outputpath tablename

若在提交作业是产生例如以下异常：

15/06/16 11:41:06 INFO mapreduce.Job: Job job_1434420992867_0003 failed with state FAILED due to: Application application_1434420992867_0003 failed 2 times due to AM Container for appattempt_1434420992867_0003_000002 exited with  exitCode: -1000
For more detailed output, check application tracking page:http://cdh1:8088/proxy/application_1434420992867_0003/Then, click on links to logs of each attempt.
Diagnostics: Rename cannot overwrite non empty destination directory /data/yarn/nm/usercache/hdfs/filecache/16
java.io.IOException: Rename cannot overwrite non empty destination directory /data/yarn/nm/usercache/hdfs/filecache/16
    at org.apache.hadoop.fs.AbstractFileSystem.renameInternal(AbstractFileSystem.java:716)
    at org.apache.hadoop.fs.FilterFs.renameInternal(FilterFs.java:228)
    at org.apache.hadoop.fs.AbstractFileSystem.rename(AbstractFileSystem.java:659)
    at org.apache.hadoop.fs.FileContext.rename(FileContext.java:909)
    at org.apache.hadoop.yarn.util.FSDownload.call(FSDownload.java:364)
    at org.apache.hadoop.yarn.util.FSDownload.call(FSDownload.java:60)
    at java.util.concurrent.FutureTask.run(FutureTask.java:262)
    at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:471)
    at java.util.concurrent.FutureTask.run(FutureTask.java:262)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
    at java.lang.Thread.run(Thread.java:745)

Failing this attempt. Failing the application.
15/06/16 11:41:06 INFO mapreduce.Job: Counters: 0

将cdh2和cdh3机器的/data/yarn/nm/usercache/hdfs/filecache下的文件删除就可以。可參考http://stackoverflow.com/questions/30857413/hadoop-complains-about-attempting-to-overwrite-nonempty-destination-directory

參考资料：

http://hbase.apache.org/book.html#arch.bulk.load

http://blog.cloudera.com/blog/2013/09/how-to-use-hbase-bulk-loading-and-why/