kafka-connect-hive sink插件实现要点小结

九州暮云

发布于 2019-08-21 10:51:31

1.2K0

发布于 2019-08-21 10:51:31

kafka-connect-hive sink插件实现了以ORC和Parquet两种方式向Hive表中写入数据。Connector定期从Kafka轮询数据并将其写入HDFS，来自每个Kafka主题的数据由提供的分区字段进行分区并划分为块，每个数据块都表示为一个HDFS文件，文件名由topic名称+分区编号+offset构成。如果配置中没有指定分区，则使用默认分区方式，每个数据块的大小由已写入HDFS的文件长度、写入HDFS的时间和未写入HDFS的记录数决定。

在阅读该插件的源码过程中，觉得有很多值得学习的地方，特总结如下以备后忘。

一、分区策略

该插件可以配置两种分区策略：

STRICT：要求必须已经创建了所有分区
DYNAMIC：根据PARTITIONBY指定的分区字段创建分区

STRICT策略

实现代码及注释如下：

package com.landoop.streamreactor.connect.hive.sink.partitioning

import com.landoop.streamreactor.connect.hive.{DatabaseName, Partition, TableName}
import org.apache.hadoop.fs.{FileSystem, Path}
import org.apache.hadoop.hive.metastore.IMetaStoreClient

import scala.collection.JavaConverters._
import scala.util.control.NonFatal
import scala.util.{Failure, Success, Try}

/**
  * A [[PartitionHandler]] that requires any partition
  * to already exist in the metastore.
  *
  * 要求分区已经在metastore中存在
  */
object StrictPartitionHandler extends PartitionHandler {

  override def path(partition: Partition,
                    db: DatabaseName,
                    tableName: TableName)
                   (client: IMetaStoreClient,
                    fs: FileSystem): Try[Path] = {
    try {
      // 获取Hive metastore中表的存储位置，成功则返回
      val part = client.getPartition(db.value, tableName.value, partition.entries.map(_._2).toList.asJava)
      Success(new Path(part.getSd.getLocation))
    } catch { // 未找到表的存储位置，返回异常
      case NonFatal(e) =>
        Failure(new RuntimeException(s"Partition '${partition.entries.map(_._2).toList.mkString(",")}' does not exist and strict policy requires upfront creation", e))
    }
  }
}

DYNAMIC策略

实现代码及注释如下：

package com.landoop.streamreactor.connect.hive.sink.partitioning

import com.landoop.streamreactor.connect.hive.{DatabaseName, Partition, TableName}
import com.typesafe.scalalogging.slf4j.StrictLogging
import org.apache.hadoop.fs.{FileSystem, Path}
import org.apache.hadoop.hive.metastore.IMetaStoreClient
import org.apache.hadoop.hive.metastore.api.{StorageDescriptor, Table}

import scala.collection.JavaConverters._
import scala.util.{Failure, Success, Try}

/**
  * A [[PartitionHandler]] that creates partitions
  * on the fly as required.
  *
  * The path of the partition is determined by the given
  * [[PartitionPathPolicy]] parameter. By default this will
  * be an implementation that uses the standard hive
  * paths of key1=value1/key2=value2.
  */
class DynamicPartitionHandler(pathPolicy: PartitionPathPolicy = DefaultMetastorePartitionPathPolicy)
  extends PartitionHandler with StrictLogging {

  override def path(partition: Partition,
                    db: DatabaseName,
                    tableName: TableName)
                   (client: IMetaStoreClient,
                    fs: FileSystem): Try[Path] = {

    def table: Table = client.getTable(db.value, tableName.value)

    def create(path: Path, table: Table): Unit = {
      logger.debug(s"New partition will be created at $path")

      // 设置的表的存储位置信息
      val sd = new StorageDescriptor(table.getSd)
      sd.setLocation(path.toString)

      val params = new java.util.HashMap[String, String]
      // 获取分区key的值、分区创建时间
      val values = partition.entries.map(_._2).toList.asJava
      val ts = (System.currentTimeMillis / 1000).toInt

      // 给表设置并创建新分区
      val p = new org.apache.hadoop.hive.metastore.api.Partition(values, db.value, tableName.value, ts, 0, sd, params)
      logger.debug(s"Updating hive metastore with partition $p")
      client.add_partition(p)

      logger.info(s"Partition has been created in metastore [$partition]")
    }

    // 获取分区信息
    Try(client.getPartition(db.value, tableName.value, partition.entries.toList.map(_._2).asJava)) match {
      case Success(p) => Try { // 成功则返回
        new Path(p.getSd.getLocation)
      }
      case Failure(_) => Try { // 失败则根据分区路径创建策略生成分区路径并返回
        val t = table
        val tableLocation = new Path(t.getSd.getLocation)
        val path = pathPolicy.path(tableLocation, partition)
        create(path, t)
        path
      }
    }
  }
}

该方式会以标准的Hive分区路径来创建分区，也就是分区字段=分区字段值的方式。

二、文件命名和大小控制

Kafka轮询数据并将其写入HDFS，来自每个Kafka主题的数据由提供的分区字段进行分区并划分为块，每个数据块都表示为一个HDFS文件，这里涉及到两个细节：

如何给文件命名
文件如何分块，文件大小及数量如何控制

接下来逐一看一下相关代码实现，文件命名部分实现代码如下：

package com.landoop.streamreactor.connect.hive.sink.staging

import com.landoop.streamreactor.connect.hive.{Offset, Topic}

import scala.util.Try

trait FilenamePolicy {
  val prefix: String
}

object DefaultFilenamePolicy extends FilenamePolicy {
  val prefix = "streamreactor"
}

object CommittedFileName {

  private val Regex = s"(.+)_(.+)_(\\d+)_(\\d+)_(\\d+)".r

  def unapply(filename: String): Option[(String, Topic, Int, Offset, Offset)] = {
    filename match {
      case Regex(prefix, topic, partition, start, end) =>
        // 返回主题名称、分区编号、起始offset和结束offset
        Try((prefix, Topic(topic), partition.toInt, Offset(start.toLong), Offset(end.toLong))).toOption
      case _ => None
    }
  }
}

从上面代码可以看出，文件名由topic名称+分区编号+offset构成。假设文件前缀是streamreactor，topic名称是hive_sink_orc，分布编号是0，当前最大的offset是1168，那么最终生成的文件名称就是streamreactor_hive_sink_orc_0_1168。

接下来看看文件的大小是如何控制的。在HDFS中一个块通常是64M、128M、256M，小文件会占用NameNode的大量元数据存储内存，增加文件数据块的寻址时间。文件的大小主要由sink插件的三个配置项决定，这些配置项信息如下：

WITH_FLUSH_INTERVAL：long类型，表示文件提交的时间间隔，单位是毫秒
WITH_FLUSH_SIZE：long类型，表示执行提交操作之前，已提交到HDFS的文件长度，单位是字节
WITH_FLUSH_COUNT：long类型，表示执行提交操作之前，未提交到HDFS的记录数，一条数据算一个记录

这些参数在CommitPolicy特质中被使用，该特质的信息及实现类如下：

package com.landoop.streamreactor.connect.hive.sink.staging

import com.landoop.streamreactor.connect.hive.TopicPartitionOffset
import com.typesafe.scalalogging.slf4j.StrictLogging
import org.apache.hadoop.fs.{FileSystem, Path}
import org.apache.kafka.connect.data.Struct

import scala.concurrent.duration.FiniteDuration

/**
  * The [[CommitPolicy]] is responsible for determining when
  * a file should be flushed (closed on disk, and moved to be visible).
  *
  * Typical implementations will flush based on number of records,
  * file size, or time since the file was opened.
  *
  * 负责决定文件何时被刷新(在磁盘上关闭，以及移动到可见)，一般情况下基于记录数量、文件大小和文件被打开的时间来刷新
  */
trait CommitPolicy {

  /**
    * This method is invoked after a file has been written.
    *
    * If the output file should be committed at this time, then this
    * method should return true, otherwise false.
    *
    * Once a commit has taken place, a new file will be opened
    * for the next record.
    *
    * 该方法在文件被写入之后调用，在这时如果文件应该被提交，该方法返回true，否则返回false。一旦发生了提交，新文件将为下一个记录打开
    *
    * @param tpo   the [[TopicPartitionOffset]] of the last record written 最后一次记录的TopicPartitionOffset
    * @param path  the path of the file that the struct was written to 文件写入的路径
    * @param count the number of records written thus far to the file 到目前为止写入文件的记录数
    *
    */
  def shouldFlush(struct: Struct, tpo: TopicPartitionOffset, path: Path, count: Long)
                 (implicit fs: FileSystem): Boolean
}

/**
  * Default implementation of [[CommitPolicy]] that will flush the
  * output file under the following circumstances:
  * - file size reaches limit
  * - time since file was created
  * - number of files is reached
  *
  * CommitPolicy 的默认实现，将根据以下场景刷新输出文件：
  * 文件大小达到限制
  * 文件创建以来的时间
  * 达到文件数量
  *
  * @param interval in millis 毫秒间隔
  */
case class DefaultCommitPolicy(fileSize: Option[Long],
                               interval: Option[FiniteDuration],
                               fileCount: Option[Long]) extends CommitPolicy with StrictLogging {
  require(fileSize.isDefined || interval.isDefined || fileCount.isDefined)

  override def shouldFlush(struct: Struct, tpo: TopicPartitionOffset, path: Path, count: Long)
                          (implicit fs: FileSystem): Boolean = {
    // 返回文件状态
    val stat = fs.getFileStatus(path)
    val open_time = System.currentTimeMillis() - stat.getModificationTime // 计算文件打开时间

    /**
      * stat.getLen：文件长度，以字节为单位
      * stat.getModificationTime：文件修改时间，以毫秒为单位
      */
    fileSize.exists(_ <= stat.getLen) || interval.exists(_.toMillis <= open_time) || fileCount.exists(_ <= count)
  }
}

现在来分析一下DefaultCommitPolicy类的实现逻辑：

首先，返回HDFS上文件的状态，接着计算文件被打开的时间，最后使用exists函数来执行以下逻辑判断：

fileSize.exists(_ <= stat.getLen)：已提交到HDFS的文件长度stat.getLen是否大于设置的文件长度阈值fileSize
interval.exists(_.toMillis <= open_time)：文件打开时间open_time是否大于设置的文件打开时间阈值interval
fileCount.exists(_ <= count)：未提交到HDFS的记录数count是否大于设置未提交到HDFS的记录数阈值fileCount

以上三个判断条件只要任何一个成立，就返回true，接着执行flush操作，将文件刷新到HDFS的对应目录中，这样就很好地控制了文件的大小以及数量，避免过多小文件的产生。

三、异常处理策略

异常处理不当的话，会直接影响服务的高可用，产生不可预估的损失。kafka-connect在处理数据读写的过程中产生的异常默认是直接抛出的，这类异常容易使负责读写的task停止服务，示例异常信息如下：

[2019-02-25 11:03:56,170] ERROR WorkerSinkTask{id=hive-sink-example-0} Task threw an uncaught and unrecoverable exception (org.apache.kafka.connect.runtime.WorkerTask:177)
MetaException(message:Could not connect to meta store using any of the URIs provided. Most recent failure: org.apache.thrift.transport.TTransportException: java.net.ConnectException: Operation timed out (Connection timed out)
	at org.apache.thrift.transport.TSocket.open(TSocket.java:226)
	at org.apache.hadoop.hive.metastore.HiveMetaStoreClient.open(HiveMetaStoreClient.java:477)
	at org.apache.hadoop.hive.metastore.HiveMetaStoreClient.<init>(HiveMetaStoreClient.java:285)
	at org.apache.hadoop.hive.metastore.HiveMetaStoreClient.<init>(HiveMetaStoreClient.java:210)
	at com.landoop.streamreactor.connect.hive.sink.HiveSinkTask.start(HiveSinkTask.scala:56)
	at org.apache.kafka.connect.runtime.WorkerSinkTask.initializeAndStart(WorkerSinkTask.java:302)
	at org.apache.kafka.connect.runtime.WorkerSinkTask.execute(WorkerSinkTask.java:191)
	at org.apache.kafka.connect.runtime.WorkerTask.doRun(WorkerTask.java:175)
	at org.apache.kafka.connect.runtime.WorkerTask.run(WorkerTask.java:219)
	at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
	at java.util.concurrent.FutureTask.run(FutureTask.java:266)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
	at java.lang.Thread.run(Thread.java:748)
Caused by: java.net.ConnectException: Operation timed out (Connection timed out)
	at java.net.PlainSocketImpl.socketConnect(Native Method)
	at java.net.AbstractPlainSocketImpl.doConnect(AbstractPlainSocketImpl.java:350)
	at java.net.AbstractPlainSocketImpl.connectToAddress(AbstractPlainSocketImpl.java:206)
	at java.net.AbstractPlainSocketImpl.connect(AbstractPlainSocketImpl.java:188)
	at java.net.SocksSocketImpl.connect(SocksSocketImpl.java:392)
	at java.net.Socket.connect(Socket.java:589)
	at org.apache.thrift.transport.TSocket.open(TSocket.java:221)
	... 13 more
)
	at org.apache.hadoop.hive.metastore.HiveMetaStoreClient.open(HiveMetaStoreClient.java:525)
	at org.apache.hadoop.hive.metastore.HiveMetaStoreClient.<init>(HiveMetaStoreClient.java:285)
	at org.apache.hadoop.hive.metastore.HiveMetaStoreClient.<init>(HiveMetaStoreClient.java:210)
	at com.landoop.streamreactor.connect.hive.sink.HiveSinkTask.start(HiveSinkTask.scala:56)
	at org.apache.kafka.connect.runtime.WorkerSinkTask.initializeAndStart(WorkerSinkTask.java:302)
	at org.apache.kafka.connect.runtime.WorkerSinkTask.execute(WorkerSinkTask.java:191)
	at org.apache.kafka.connect.runtime.WorkerTask.doRun(WorkerTask.java:175)
	at org.apache.kafka.connect.runtime.WorkerTask.run(WorkerTask.java:219)
	at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
	at java.util.concurrent.FutureTask.run(FutureTask.java:266)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
	at java.lang.Thread.run(Thread.java:748)
[2019-02-25 11:03:56,172] ERROR WorkerSinkTask{id=hive-sink-example-0} Task is being killed and will not recover until manually restarted (org.apache.kafka.connect.runtime.WorkerTask:178)

在以上异常信息可以看到，由于连接Hive metastore超时，因此相关的Task被杀死，需要我们手动重启。当然这只是kafka-connect在运行中发生的一个异常，对于这类容易使Task停止工作的异常，需要设置相关的异常处理策略，sink插件在实现中定义了三种异常处理策略，分别如下：

NOOP：表示在异常发生后，不处理异常，继续工作
THROW：表示在异常发生后，直接抛出异常，这样会使服务停止
RETRY：表示在异常发生后，进行重试，相应地，需要定义重试次数，来避免无限重试情况的发生

基于以上三种异常处理策略，sink插件相关的实现类如下：

/*
 *  Copyright 2017 Datamountaineer.
 *
 *  Licensed under the Apache License, Version 2.0 (the "License");
 *  you may not use this file except in compliance with the License.
 *  You may obtain a copy of the License at
 *
 *  http://www.apache.org/licenses/LICENSE-2.0
 *
 *  Unless required by applicable law or agreed to in writing, software
 *  distributed under the License is distributed on an "AS IS" BASIS,
 *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 *  See the License for the specific language governing permissions and
 *  limitations under the License.
 */

package com.datamountaineer.streamreactor.connect.errors

import java.util.Date

import com.datamountaineer.streamreactor.connect.errors.ErrorPolicyEnum.ErrorPolicyEnum
import com.typesafe.scalalogging.slf4j.StrictLogging
import org.apache.kafka.connect.errors.RetriableException

/**
  * Created by andrew@datamountaineer.com on 19/05/16. 
  * kafka-connect-common
  */
object ErrorPolicyEnum extends Enumeration {
  type ErrorPolicyEnum = Value
  val NOOP, THROW, RETRY = Value
}

case class ErrorTracker(retries: Int, maxRetries: Int, lastErrorMessage: String, lastErrorTimestamp: Date, policy: ErrorPolicy)

trait ErrorPolicy extends StrictLogging {
  def handle(error: Throwable, sink: Boolean = true, retryCount: Int = 0)
}

object ErrorPolicy extends StrictLogging {
  def apply(policy: ErrorPolicyEnum): ErrorPolicy = {
    policy match {
      case ErrorPolicyEnum.NOOP => NoopErrorPolicy()
      case ErrorPolicyEnum.THROW => ThrowErrorPolicy()
      case ErrorPolicyEnum.RETRY => RetryErrorPolicy()
    }
  }
}

/**
 * 不处理异常策略
 */
case class NoopErrorPolicy() extends ErrorPolicy {
  override def handle(error: Throwable, sink: Boolean = true, retryCount: Int = 0){
    logger.warn(s"Error policy NOOP: ${error.getMessage}. Processing continuing.")
  }
}

/**
 * 异常抛出处理策略
 */
case class ThrowErrorPolicy() extends ErrorPolicy {
  override def handle(error: Throwable, sink: Boolean = true, retryCount: Int = 0){
    throw new RuntimeException(error)
  }
}

/**
 * 异常重试处理策略
 */
case class RetryErrorPolicy() extends ErrorPolicy {

  override def handle(error: Throwable, sink: Boolean = true, retryCount: Int) = {
    if (retryCount == 0) {
      throw new RuntimeException(error)
    }
    else {
      logger.warn(s"Error policy set to RETRY. Remaining attempts $retryCount")
      throw new RetriableException(error)
    }
  }
}