AWS Glue集成Kafka/MSK

简单总结一下AWS Glue如何集成Kafka/MSK吧，因为当前（Glue 2.0）Glue对Kafka的集成和使用在我看来是比较“扭曲”的，并不能像原生Spark那样以spark.readStream的方式接入Kafka，所以暂时只做一下简单的归纳，后续根据新的进展再做补充。

在Glue中接入Kafka的消息需要先创建Glue Connection，只有通过Connection，Glue才能访问到外部Kakfa/MSK集群，因为在配置Connection的过程中需要提供VPC信息，Glue会透明地给后台节点创建ENI（虚拟网卡）联通到指定的VPC上，注意这一操作是通明的，不可控的。

有了Connection之后，还得建一张数据表（就是一张普通的Hive表），并在表的属性配置（TBLPROPERTIES）中设定所使用的Connection。以下是一个示例：

CREATE EXTERNAL TABLE `default.user`(
  `id` bigint COMMENT 'from deserializer',
  `name` string COMMENT 'from deserializer', 
  `age` int COMMENT 'from deserializer', 
  `updatedtime` bigint COMMENT 'from deserializer')
ROW FORMAT SERDE 
  'org.openx.data.jsonserde.JsonSerDe' 
WITH SERDEPROPERTIES ( 
  'paths'='') 
STORED AS INPUTFORMAT 
  'org.apache.hadoop.mapred.TextInputFormat' 
OUTPUTFORMAT 
  'org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat'
LOCATION
  'user'
TBLPROPERTIES (
  'classification'='json', 
  'connectionName'='MSK')

在这个建表语句中，需要注意的是最后一行：'connectionName'='MSK' ，这就是将connection配置给数据表的地方，这里的MSK是我创建的一个Connection的名字，稍后会给出这个Connection的具体信息。但这里要关注的重点在于：**Glue将对Kafka的接入“畸形”地映射成为了对一张Hive表的读取，这是很不好的。**这意味着，在Glue中, 像下面这种用标准Spark API读取Kafka消息的代码是跑不通的：

val stream = spark
      .readStream
      .format("kafka")
      .option("kafka.bootstrap.servers", "broker1:9092,broker2:9092")
      .option("subscribe", "user")
      .option("startingOffsets", "earliest")
      .load()

然后，看一下前面提到的名为MSK的Connection配置吧：

从这个Connection中可以看到，关键性的网络连接信息都是在Connection中配置的，Glue会基于这些信息在Glue作业节点上创建ENI（虚拟网卡），以便作业节点可以连接到Kafka的Broker。这就是为什么在Glue中我们不能脱离开Connection直接使用option("kafka.bootstrap.servers", "broker1:9092,broker2:9092") 的形式连接Kafka的原因，因为Connection包含了网络相关的配置，而读取数据表生成Dataframe的过程会触发Connection的初始化，也就是打通网络，所以导致了Glue集成Kafka变成了一种类似"读表"的操作。

作为一份参考，以下是一份自动生成的读取MSK的参考脚本：

package com.github.glueboot.core.job

import com.amazonaws.services.glue.DynamicFrame
import com.amazonaws.services.glue.GlueContext
import com.amazonaws.services.glue.MappingSpec
import com.amazonaws.services.glue.errors.CallSite
import com.amazonaws.services.glue.util.GlueArgParser
import com.amazonaws.services.glue.util.Job
import com.amazonaws.services.glue.util.JsonOptions
import java.util.Calendar
import org.apache.spark.SparkContext
import org.apache.spark.sql.Dataset
import org.apache.spark.sql.Row
import org.apache.spark.sql.SaveMode
import org.apache.spark.sql.streaming.Trigger
import scala.collection.JavaConverters._

object GlueApp {
     
  def main(sysArgs: Array[String]) {
     
    val spark: SparkContext = new SparkContext()
    val glueContext: GlueContext = new GlueContext(spark)
    // @params: [JOB_NAME]
    val args = GlueArgParser.getResolvedOptions(sysArgs, Seq("JOB_NAME").toArray)
    Job.init(args("JOB_NAME"), glueContext, args.asJava)
    // @type: DataSource
    // @args: [stream_type = kafka, stream_batch_time = "100 seconds", database = "default", additionalOptions = {"startingOffsets": "earliest", "inferSchema": "false"}, stream_checkpoint_location = "s3:///user/checkpoint/", table_name = "user"]
    // @return: datasource0
    // @inputs: []
    val data_frame_datasource0 = glueContext.getCatalogSource(database = "default", tableName = "user", redshiftTmpDir = "", transformationContext = "datasource0", additionalOptions = JsonOptions("""{"startingOffsets": "earliest", "inferSchema": "false"}""")).getDataFrame()
    glueContext.forEachBatch(data_frame_datasource0, (dataFrame: Dataset[Row], batchId: Long) => {
     
      if (dataFrame.count() > 0) {
     
        val datasource0 = DynamicFrame(dataFrame, glueContext)
        // @type: ApplyMapping
        // @args: [mapping = [("id", "long", "id", "long"), ("name", "string", "name", "string"), ("age", "int", "age", "int"), ("updatedtime", "long", "updatedtime", "long")], transformation_ctx = "applymapping0"]
        // @return: applymapping0
        // @inputs: [frame = datasource0]
        val applymapping0 = datasource0.applyMapping(mappings = Seq(("id", "long", "id", "long"), ("name", "string", "name", "string"), ("age", "int", "age", "int"), ("updatedtime", "long", "updatedtime", "long")), caseSensitive = false, transformationContext = "applymapping0")
        // @type: DataSink
        // @args: [stream_batch_time = "100 seconds", stream_checkpoint_location = "s3:///user/checkpoint/", connection_type = "s3", path = "s3:///user", format = "csv", transformation_ctx = "datasink1"]
        // @return: datasink1
        // @inputs: [frame = applymapping0]
        val year: Int = Calendar.getInstance().get(Calendar.YEAR)
        val month: Int = Calendar.getInstance().get(Calendar.MONTH) + 1
        val day: Int = Calendar.getInstance().get(Calendar.DATE)
        val hour: Int = Calendar.getInstance().get(Calendar.HOUR_OF_DAY)
        val minute: Int = Calendar.getInstance().get(Calendar.MINUTE)
        val path_datasink1 = "s3:///user" + "/ingest_year=" + "%04d".format(year) + "/ingest_month=" + "%02d".format(month) + "/ingest_day=" + "%02d".format(day) + "/ingest_hour=" + "%02d".format(hour) + "/"
        val datasink1 = glueContext.getSinkWithFormat(connectionType = "s3", options = JsonOptions(s"""{"path": "$path_datasink1"}"""), transformationContext = "datasink1", format = "csv").writeDynamicFrame(applymapping0)
      }
    }, JsonOptions("""{"windowSize" : "100 seconds", "checkpointLocation" : "s3:///user/checkpoint/"}"""))
    Job.commit()
  }
}

---

推荐：博主历时三年倾注大量心血创作的《大数据平台架构与原型实现：数据中台建设实战》一书已由知名IT图书品牌电子工业出版社博文视点出版发行，真诚推荐给每一位读者！点击《重磅推荐：建大数据平台太难了！给我发个工程原型吧！》了解图书详情，扫码进入京东手机购书页面！