本文转载自：https://blog.csdn.net/xnlej/article/details/79037145
spark streaming 用direct 的方式有优势，但是也容易丢失数据，只能保证at least one ,不能保证exactly one ,要想保证后者，只能手动保存kafka的offset数据。实现方面参考了一位大神的java代码，把它改写成scala 代码，并修复相关bug。在此基础上进一步实现了事务机制
我们从Spark的官方文档可以知道，维护Spark内部维护Kafka便宜了信息是存储在HasOffsetRanges类的offsetRanges中，我们可以在Spark Streaming程序里面获取这些信息：
1 val offsetsList = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
这样我们就可以获取所以分区消费信息，只需要遍历offsetsList
import java.net.URLDecoder
import com.dianyou.utl.PropertiesUtil
import org.apache.hadoop.hbase._
import org.apache.hadoop.hbase.client._
import org.apache.hadoop.hbase.filter.CompareFilter.CompareOp
import org.apache.hadoop.hbase.filter.{BinaryComparator, RowFilter}
import org.apache.hadoop.hbase.util.Bytes
import org.apache.kafka.clients.consumer.ConsumerRecord
import org.apache.kafka.common.TopicPartition
import org.apache.kafka.common.serialization.StringDeserializer
import org.apache.spark.rdd.RDD
import org.apache.spark.streaming.dstream.InputDStream
import org.apache.spark.streaming.kafka010.{ConsumerStrategies, HasOffsetRanges, KafkaUtils, LocationStrategies}
import org.apache.spark.streaming.{Seconds, StreamingContext, Time}
import org.apache.spark.{SparkConf, SparkContext}
import scala.collection.mutable
/**
* 手工操作offset
* 1 从hbase获取offset，从kafka拉取数据
* 2 数据处理完后，把until offset 保存到hbase
* 3 kafka 长时间挂掉之后，从kafka最早的offset 开始读取 此处还需要处理
* Created by Administrator on 2017/12/28.
*/
object OffsetOperate {
var hbaseProp = PropertiesUtil.getProperties("hbase")
var kafkaconsumePro = PropertiesUtil.getProperties("kafkaconsume")
def main(args: Array[String]): Unit = {
val conf = new SparkConf().setAppName("sparkStreaming - offset operate")
.setMaster("local[2]") // --master local[2] | spark://xx:7077 | yarn
.set("spark.testing.memory", "2147480000")
val sc = new SparkContext(conf)
val ssc = new StreamingContext(sc,Seconds(5))
//kafka配置
val kafkaParams = Map[String, Object](
"bootstrap.servers" -> kafkaconsumePro.getProperty("bootstrap.servers"),
"key.deserializer" -> classOf[StringDeserializer],
"value.deserializer" -> classOf[StringDeserializer],
"group.id" -> kafkaconsumePro.getProperty("group"),
"auto.offset.reset" -> "earliest", // 第一次读取时从topic 首位置开始读取
"enable.auto.commit" -> (false: java.lang.Boolean)// kafka 不保存消费的offset
)
//监听频道
val topics = Array(kafkaconsumePro.getProperty("topics"))
// 获取hbase连接
val hbaseConf = HBaseConfiguration.create()
hbaseConf.set("hbase.zookeeper.quorum",hbaseProp.getProperty("quorum")) //zookeeper 集群
hbaseConf.set("hbase.zookeeper.property.client","2181")
hbaseConf.set("hbase.master", hbaseProp.getProperty("hbase_master"))
hbaseConf.set("hbase.defaults.for.version.skip", "true")
//获取连接对象
val conn = ConnectionFactory.createConnection(hbaseConf)
val admin = conn.getAdmin
val tn = TableName.valueOf("hbase_consumer_offset") //hbase 表名
val isExist = admin.tableExists(tn)
val streams : InputDStream[ConsumerRecord[String,String]]= {
if(isExist) {
val table = new HTable(hbaseConf, "hbase_consumer_offset")
val filter = new RowFilter(CompareOp.GREATER_OR_EQUAL, new BinaryComparator(Bytes.toBytes(topics + "_")))
println("============ 过滤器已经创建 ==========")
val s = new Scan()
s.setFilter(filter)
val rs = table.getScanner(s)
// 设置 offset
val fromOffsets = scala.collection.mutable.Map[TopicPartition, Long]()
var s1 = ""
var s2 = 0
var s3: Long = 0
for (r: Result <- rs.next(200)) {
println("rowKey : " + new String(r.getRow))
for (keyvalue: KeyValue <- r.raw()) {
if ("topic".equals(new String(keyvalue.getQualifier))) {
s1 = new String(keyvalue.getValue)
println("columnFamily :" + new String(keyvalue.getFamily) + " column :" +new String( keyvalue.getQualifier) + s1)
} else if ("partition".equals(new String(keyvalue.getQualifier))){
s2 = Bytes.toInt(keyvalue.getValue)
println("columnFamily :" + new String(keyvalue.getFamily) + " column :" + new String( keyvalue.getQualifier) + s2)
} else if("offset".equals(new String(keyvalue.getQualifier))) { //if("offset".equals(new String(keyvalue.getQualifier)))
s3 = Bytes.toLong(keyvalue.getValue)
println("columnFamily :" + new String(keyvalue.getFamily) + " column :" + new String( keyvalue.getQualifier) + s3)
}
}
fromOffsets.put(new TopicPartition(s1, s2), s3)
}
println("fromOffset is : "+fromOffsets)
KafkaUtils.createDirectStream(ssc, LocationStrategies.PreferConsistent,
ConsumerStrategies.Assign(fromOffsets.keySet, kafkaParams, fromOffsets)) //(fromOffsets.keySet,kafkaParams,fromOffsets))
}else{ //Hbase 里面不存在offset表，从topic首位置开始消费
val htable = new HTableDescriptor(TableName.valueOf("hbase_consumer_offset"))
htable.addFamily(new HColumnDescriptor(("topic_partition_offset")))
admin.createTable(htable)
println("表已经创建成功========" + htable)
KafkaUtils.createDirectStream(ssc, LocationStrategies.PreferConsistent, ConsumerStrategies.Subscribe(topics, kafkaParams))
}
}
// val dstream = streams.map(x=>URLDecoder.decode(x.value()))
// 操作成功后更新offset
streams.foreachRDD{ rdd =>
//if(!rdd.isEmpty()){
// 打成一个事务，把业务计算和offset保存放在一起，要么成功，要么一起失败，实现精确一次的消费
import scala.collection.JavaConversions._
val table = new HTable(hbaseConf,"hbase_consumer_offset")
table.setAutoFlush(false, false)
var putList:List[Put] = List()
val offsetRanges = rdd.asInstanceOf[HasOffsetRanges].offsetRanges // RDD[ConsumerRecord[String,String]] 强转成offsetRanges
for(offsetRange <- offsetRanges){
println("the topic is "+offsetRange.topic)
println("the partition is "+offsetRange.partition)
println("the fromOffset is "+offsetRange.fromOffset)
println("the untilOffset is "+offsetRange.untilOffset)
println("the object is "+offsetRange)
// val table = new HTable(hbaseConf,"hbase_consumer_offset")
// table.setAutoFlush(false, false)
val put = new Put(Bytes.toBytes(offsetRange.topic+"_"+offsetRange.partition))
put.add(Bytes.toBytes("topic_partition_offset"),Bytes.toBytes("topic"),Bytes.toBytes(offsetRange.topic))
put.add(Bytes.toBytes("topic_partition_offset"),Bytes.toBytes("partition"),Bytes.toBytes(offsetRange.partition))
put.add(Bytes.toBytes("topic_partition_offset"),Bytes.toBytes("offset"),Bytes.toBytes(offsetRange.untilOffset))
putList = put+:putList
// println("add data success !")
}
println("the RDD records are "+rdd.map{x =>URLDecoder.decode(x.value())}.collect.foreach(println)) // 程序的计算逻辑
// }
table.put(putList)
table.flushCommits()
println("add and compute data success !")
}
ssc.start()
ssc.awaitTermination()
}
}
此处还有一个问题，当kafka 数据清掉之后，从hbase中读取offset数据到kafka查询时，会读不到数据 报数组越界的错误，有时间再来完善下
参考链接 ：https://www.jianshu.com/p/667e0f58b7b9
网友评论