What is the fastest way to bulk load data into HBase programmatically?

Question

I have a Plain text file with possibly millions of lines which needs custom parsing and I want to load it into an HBase table as fast as possible (using Hadoop or HBase Java client).

My current solution is based on a MapReduce job without the Reduce part. I use FileInputFormat to read the text file so that each line is passed to the map method of my Mapper class. At this point the line is parsed to form a Put object which is written to the context. Then, TableOutputFormat takes the Put object and inserts it to table.

This solution yields an average insertion rate of 1,000 rows per second, which is less than what I expected. My HBase setup is in pseudo distributed mode on a single server.

One interesting thing is that during insertion of 1,000,000 rows, 25 Mappers (tasks) are spawned but they run serially (one after another); is this normal?

Here is the code for my current solution:

public static class CustomMap extends Mapper<LongWritable, Text, ImmutableBytesWritable, Put> {

    protected void map(LongWritable key, Text value, Context context) throws IOException {
        Map<String, String> parsedLine = parseLine(value.toString());

        Put row = new Put(Bytes.toBytes(parsedLine.get(keys[1])));
        for (String currentKey : parsedLine.keySet()) {
            row.add(Bytes.toBytes(currentKey),Bytes.toBytes(currentKey),Bytes.toBytes(parsedLine.get(currentKey)));
        }

        try {
            context.write(new ImmutableBytesWritable(Bytes.toBytes(parsedLine.get(keys[1]))), row);
        } catch (InterruptedException e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        }
    }
}

public int run(String[] args) throws Exception {
    if (args.length != 2) {
        return -1;
    }

    conf.set("hbase.mapred.outputtable", args[1]);

    // I got these conf parameters from a presentation about Bulk Load
    conf.set("hbase.hstore.blockingStoreFiles", "25");
    conf.set("hbase.hregion.memstore.block.multiplier", "8");
    conf.set("hbase.regionserver.handler.count", "30");
    conf.set("hbase.regions.percheckin", "30");
    conf.set("hbase.regionserver.globalMemcache.upperLimit", "0.3");
    conf.set("hbase.regionserver.globalMemcache.lowerLimit", "0.15");

    Job job = new Job(conf);
    job.setJarByClass(BulkLoadMapReduce.class);
    job.setJobName(NAME);
    TextInputFormat.setInputPaths(job, new Path(args[0]));
    job.setInputFormatClass(TextInputFormat.class);
    job.setMapperClass(CustomMap.class);
    job.setOutputKeyClass(ImmutableBytesWritable.class);
    job.setOutputValueClass(Put.class);
    job.setNumReduceTasks(0);
    job.setOutputFormatClass(TableOutputFormat.class);

    job.waitForCompletion(true);
    return 0;
}

public static void main(String[] args) throws Exception {
    Long startTime = Calendar.getInstance().getTimeInMillis();
    System.out.println("Start time : " + startTime);

    int errCode = ToolRunner.run(HBaseConfiguration.create(), new BulkLoadMapReduce(), args);

    Long endTime = Calendar.getInstance().getTimeInMillis();
    System.out.println("End time : " + endTime);
    System.out.println("Duration milliseconds: " + (endTime-startTime));

    System.exit(errCode);
}

Answer 1

I've gone through a process that is probably very similar to yours of attempting to find an efficient way to load data from an MR into HBase. What I found to work is using HFileOutputFormat as the OutputFormatClass of the MR.

Below is the basis of my code that I have to generate the job and the Mapper map function which writes out the data. This was fast. We don't use it anymore, so I don't have numbers on hand, but it was around 2.5 million records in under a minute.

Here is the (stripped down) function I wrote to generate the job for my MapReduce process to put data into HBase

private Job createCubeJob(...) {
    //Build and Configure Job
    Job job = new Job(conf);
    job.setJobName(jobName);
    job.setMapOutputKeyClass(ImmutableBytesWritable.class);
    job.setMapOutputValueClass(Put.class);
    job.setMapperClass(HiveToHBaseMapper.class);//Custom Mapper
    job.setJarByClass(CubeBuilderDriver.class);
    job.setInputFormatClass(TextInputFormat.class);
    job.setOutputFormatClass(HFileOutputFormat.class);

    TextInputFormat.setInputPaths(job, hiveOutputDir);
    HFileOutputFormat.setOutputPath(job, cubeOutputPath);

    Configuration hConf = HBaseConfiguration.create(conf);
    hConf.set("hbase.zookeeper.quorum", hbaseZookeeperQuorum);
    hConf.set("hbase.zookeeper.property.clientPort", hbaseZookeeperClientPort);

    HTable hTable = new HTable(hConf, tableName);

    HFileOutputFormat.configureIncrementalLoad(job, hTable);
    return job;
}

This is my map function from the HiveToHBaseMapper class (slightly edited ).

public void map(WritableComparable key, Writable val, Context context)
        throws IOException, InterruptedException {
    try{
        Configuration config = context.getConfiguration();
        String[] strs = val.toString().split(Constants.HIVE_RECORD_COLUMN_SEPARATOR);
        String family = config.get(Constants.CUBEBUILDER_CONFIGURATION_FAMILY);
        String column = strs[COLUMN_INDEX];
        String Value = strs[VALUE_INDEX];
        String sKey = generateKey(strs, config);
        byte[] bKey = Bytes.toBytes(sKey);
        Put put = new Put(bKey);
        put.add(Bytes.toBytes(family), Bytes.toBytes(column), (value <= 0) 
                        ? Bytes.toBytes(Double.MIN_VALUE)
                        : Bytes.toBytes(value));

        ImmutableBytesWritable ibKey = new ImmutableBytesWritable(bKey);
        context.write(ibKey, put);

        context.getCounter(CubeBuilderContextCounters.CompletedMapExecutions).increment(1);
    }
    catch(Exception e){
        context.getCounter(CubeBuilderContextCounters.FailedMapExecutions).increment(1);    
    }

}

I pretty sure this isn't going to be a Copy&Paste solution for you. Obviously the data I was working with here didn't need any custom processing (that was done in a MR job before this one). The main thing I want to provide out of this is the HFileOutputFormat. The rest is just an example of how I used it. :)
I hope it gets you onto a solid path to a good solution. :

Answer 2

One interesting thing is that during insertion of 1,000,000 rows, 25 Mappers (tasks) are spawned but they run serially (one after another); is this normal?

mapreduce.tasktracker.map.tasks.maximum parameter which is defaulted to 2 determines the maximum number of tasks that can run in parallel on a node. Unless changed, you should see 2 map tasks running simultaneously on each node.