自定义Spark Partitioner提升es-hadoop Bulk效率

http://www.jianshu.com/p/cccc56e39429/comments/2022782 和 https://github.com/elastic/elasticsearch-hadoop/issues/745 都有提到通过自定义Spark Partitioner提升es-hadoop Bulk效率，但是无可运行代码，自己针对其思路在spark-shell里实现了一份。

思路：

spark streming监控/tmp/data下的新文件，并将文中每行内容存储到ES的web/blog索引里！

注意：代码里使用了doc ID来定制路由，该id为自动生成的uuid！因此在启动ES后，需要：

curl -s -XPUT localhost:/web -d '
{
    "mappings": {
        "blog": {
            "_id": {
                "path": "uuid"
            },
            "properties": {
                "title": {
                    "type":   "string",
                    "index":  "analyzed"
                }
            }
        }
    }
}'

告诉ES使用blog document中的uuid字段作为_id。ES 2.0以后见 http://stackoverflow.com/questions/32334709/how-to-set-id-in-elasticsearch-2-0

下面是spark-shell代码：

import org.apache.spark._
import org.apache.spark.streaming._
import org.elasticsearch.spark._
import org.apache.spark.Partitioner
import org.elasticsearch.hadoop.cfg.PropertiesSettings
import org.elasticsearch.spark.cfg.SparkSettingsManager
import org.elasticsearch.hadoop.cfg.Settings
import org.elasticsearch.hadoop.rest.RestRepository
import scala.collection.JavaConversions._

// 为方便测试，下面是自己用scala实现的es hash函数
// 尤其注意：在生产环境下，使用ES jar包里的函数，位置为：
// https://github.com/elastic/elasticsearch/blob/master/core/src/main/java/org/elasticsearch/cluster/routing/Murmur3HashFunction.java
object Murmur3HashFunction {
  def hash(routing: String): Int = {
    val bytesToHash = Array.ofDim[Byte](routing.length * 2)
    for (i <- 0 until routing.length) {
      val c = routing.charAt(i)
      val b1 = c.toByte
      val b2 = (c >>> 8).toByte
      assert(((b1 & 0xFF) | ((b2 & 0xFF) << 8)) == c)
      bytesToHash(i * 2) = b1
      bytesToHash(i * 2 + 1) = b2
    }
    hash(bytesToHash, 0, bytesToHash.length)
  }

  def hash(bytes: Array[Byte], offset: Int, length: Int): Int = {
    murmurhash3_x86_32(bytes, offset, length, 0)
  }

  def murmurhash3_x86_32(data: Array[Byte],
      offset: Int,
      len: Int,
      seed: Int): Int = {
    val c1 = 0xcc9e2d51
    val c2 = 0x1b873593
    var h1 = seed
    val roundedEnd = offset + (len & 0xfffffffc)
    var i = offset
    while (i < roundedEnd) {
      var k1 = (data(i) & 0xff) | ((data(i + 1) & 0xff) << 8) | ((data(i + 2) & 0xff) << 16) |
        (data(i + 3) << 24)
      k1 *= c1
      k1 = (k1 << 15) | (k1 >>> 17)
      k1 *= c2
      h1 ^= k1
      h1 = (h1 << 13) | (h1 >>> 19)
      h1 = h1 * 5 + 0xe6546b64
      i += 4
    }
    var k1 = 0
    len & 0x03 match {
      case 3 => k1 = (data(roundedEnd + 2) & 0xff) << 16
      case 2 => k1 |= (data(roundedEnd + 1) & 0xff) << 8
      case 1 =>
        k1 |= (data(roundedEnd) & 0xff)
        k1 *= c1
        k1 = (k1 << 15) | (k1 >>> 17)
        k1 *= c2
        h1 ^= k1
      case _ => //break
    }
    h1 ^= len
    h1 ^= h1 >>> 16
    h1 *= 0x85ebca6b
    h1 ^= h1 >>> 13
    h1 *= 0xc2b2ae35
    h1 ^= h1 >>> 16
    h1
  }
}

// 自定义Partitioner
class ESShardPartitioner(settings: String) extends Partitioner {
      protected var _numPartitions = -1

      override def numPartitions: Int = {
        val newSettings = new PropertiesSettings().load(settings)
        // 生产环境下，需要自行设置索引的 index/type，我是以web/blog作为实验的index
        newSettings.setResourceRead("web/blog") // ******************** !!! modify it !!! ********************
        newSettings.setResourceWrite("web/blog") // ******************** !!! modify it !!! ********************
        val repository = new RestRepository(newSettings)
        val targetShards = repository.getWriteTargetPrimaryShards(newSettings.getNodesClientOnly())
        repository.close()
        _numPartitions = targetShards.size()
        _numPartitions
      } 

      override def getPartition(docID: Any): Int = {
        var shardId = Murmur3HashFunction.hash(docID.toString()) % _numPartitions;
        if (shardId < 0) {
            shardId += _numPartitions;
        }
        shardId
      }
}

sc.getConf.setMaster("local").setAppName("RDDTest").set("es.nodes", "127.0.0.1").set("spark.serializer", "org.apache.spark.serializer.KryoSerializer").set("es.index.auto.create", "true");
val ssc = new StreamingContext(sc, Seconds(2));
val fileStream = ssc.textFileStream("/tmp/data");

fileStream.foreachRDD { rdd => {
    def makeItem(content: String) : (String, Map[String,String]) = {
        val uuid = java.util.UUID.randomUUID.toString();
        (uuid, Map("content"->content, "uuid"->uuid))
    }
    println("********************start*************************");
    var r2 = rdd.map(makeItem);
    val sparkCfg = new SparkSettingsManager().load(rdd.sparkContext.getConf)
    val settings = sparkCfg.save();
    var r3 = r2.partitionBy(new ESShardPartitioner(settings));
    r3.map(x=>x._2).saveToEs("web/blog")
    println("data count: " + rdd.count.toString);
    println("*********************end************************");
}};

ssc.start();
ssc.awaitTermination();

运行方法：

./spark-shell --jars ../lib/elasticsearch-spark-1.2_2.10-2.1.2.jar

然后在spark shell里运行上述代码。

通过shell 伪造数据：

mkdir /mmp/data
#rm -rf  /tmp/ ata"
rm -f "/tmp/data/*"
for ((j=;j<;j++)); do
        {
        for ((i=;i<;i++)); do
        file_name=`python -c 'import random;print random.random()'`
        echo "$j $i is sad story." >"/tmp/data/$file_name.log"
        done
        sleep
        }
done
echo "OK, waiting..."
echo "done"

运行上述脚本，看到spark shell里显示：

见http://www.cnblogs.com/bonelee/p/6078956.html ES路由底层实现！