Spark 结构化流异常处理

Question

我使用 Spark Structured Streaming 从 MQTT 流源读取数据 API。

val lines:= spark.readStream
  .format("org.apache.bahir.sql.streaming.mqtt.MQTTStreamSourceProvider")
  .option("topic", "Employee")
  .option("username", "username")
  .option("password", "passwork")
  .option("clientId", "employee11")
  .load("tcp://localhost:8000").as[(String, Timestamp)]

我将流数据转换为案例class员工

case class Employee(Name: String, Department: String)    
val ds = lines.map {
        row =>
          implicit val format = DefaultFormats
          parse(row._1).extract[Employee]
      }
  ....some transformations
 df.writeStream
        .outputMode("append")
        .format("es")
        .option("es.resource", "spark/employee")
        .option("es.nodes", "localhost")
        .option("es.port", 9200)
        .start()
        .awaitTermination()

现在队列中有一些消息的结构不同于 Employee 案例 class。假设缺少一些必需的列。我的流式处理作业因未找到字段异常而失败。

现在我想处理这样的异常，也想为此发送警报通知。我试着放一个 try/catch 块。

case class ErrorMessage(row: String)        
catch {

      case e: Exception =>
        val ds = lines.map {
          row =>
            implicit val format = DefaultFormats
            parse(row._1).extract[ErrorMessage]
        }
        val error = lines.foreach(row => {
          sendErrorMail(row._1)
        })
    }
  }

得到异常 Exception in thread "main" org.apache.spark.sql.AnalysisException: Queries with streaming sources must be executed with writeStream.start();; mqtt 如有任何帮助，我们将不胜感激。

Answer 1

我在 catch 块中创建了一个 foreach 接收器，并且能够处理异常并发送邮件警报。

catch {
    case e: Exception =>
      val foreachWriter = new ForeachWriter[Row] {
        override def open(partitionId: Timestamp, version: Timestamp): Boolean = {
          true
        }

        override def process(value: Row): Unit = {
         code for sending mail.........
        }

        override def close(errorOrNull: Throwable): Unit = {}
      }
      val df = lines.selectExpr("cast (value as string) as json")
      df.writeStream
        .foreach(foreachWriter)
        .outputMode("append")
        .start()
        .awaitTermination()
  }

Answer 2

我认为您应该使用 start() 方法的 return 对象，如 Spark streaming doc 中所述。类似于：

val query = df.writeStream. ... .start()
try {
   //If the query has terminated with an exception, then the exception will be thrown.
   query.awaitTermination()
catch {
   case ex: Exception => /*code to send mail*/
}

实现自己的 foreach sink 会导致频繁打开和关闭连接的开销。

Spark 结构化流异常处理

Spark Structured Streaming exception handling

spark-structured-streaming