如何在 spark 3.0 结构化流媒体中使用 kafka.group.id 和检查点以继续从 Kafka 中读取它在重启后停止的位置？

Question

Based on the introduction in Spark 3.0, https://spark.apache.org/docs/latest/structured-streaming-kafka-integration.html . It should be possible to set "kafka.group.id" to track the offset. For our use case, I want to avoid the potential data loss if the streaming spark job failed and restart. Based on my previous questions, I have a feeling that kafka.group.id in Spark 3.0 is something that will help.

How to specify the group id of kafka consumer for spark structured streaming?

How to ensure no data loss for kafka data ingestion through Spark Structured Streaming?

However, I tried the settings in spark 3.0 as below.

package com.example

/**
 * @author ${user.name}
 */
import scala.math.random

import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.types.{StructType, StructField, StringType, IntegerType, BooleanType, LongType}
import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions._
import org.apache.spark.sql.Row
import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.SaveMode
import org.apache.spark.SparkFiles
import java.util.Properties
import org.postgresql.Driver
import org.apache.spark.sql.streaming.Trigger
import java.time.Instant
import org.apache.hadoop.fs.{FileSystem, Path}
import java.net.URI
import java.sql.Connection
import java.sql.DriverManager
import java.sql.ResultSet
import java.sql.SQLException
import java.sql.Statement


//import org.apache.spark.sql.hive.HiveContext

import scala.io.Source

import java.nio.charset.StandardCharsets

import com.amazonaws.services.kms.{AWSKMS, AWSKMSClientBuilder}
import com.amazonaws.services.kms.model.DecryptRequest
import java.nio.ByteBuffer
import com.google.common.io.BaseEncoding


object App {
    
    def main(args: Array[String]): Unit = {
      
      val spark: SparkSession = SparkSession.builder()
        .appName("MY-APP")
        .getOrCreate()

      import spark.sqlContext.implicits._

      spark.catalog.clearCache()
      spark.conf.set("spark.sql.autoBroadcastJoinThreshold", -1)
      spark.conf.set("spark.sql.legacy.timeParserPolicy", "LEGACY")

      spark.sparkContext.setLogLevel("ERROR")
      spark.sparkContext.setCheckpointDir("/home/ec2-user/environment/spark/spark-local/checkpoint")
      
      System.gc()
      
      val df = spark.readStream
        .format("kafka")
          .option("kafka.bootstrap.servers", "mybroker.io:6667")
          .option("subscribe", "mytopic")
          .option("kafka.security.protocol", "SASL_SSL")
          .option("kafka.ssl.truststore.location", "/home/ec2-user/environment/spark/spark-local/creds/cacerts")
          .option("kafka.ssl.truststore.password", "changeit")
          .option("kafka.ssl.truststore.type", "JKS")
          .option("kafka.sasl.kerberos.service.name", "kafka")
          .option("kafka.sasl.mechanism", "GSSAPI")
          .option("kafka.group.id","MYID")
          .load()

      df.printSchema()

      
      val schema = new StructType()
        .add("id", StringType)
        .add("x", StringType)
        .add("eventtime", StringType)

      val idservice = df.selectExpr("CAST(value AS STRING)")
        .select(from_json(col("value"), schema).as("data"))
        .select("data.*")

       
      val monitoring_df = idservice
                .selectExpr("cast(id as string) id", 
                            "cast(x as string) x",
                            "cast(eventtime as string) eventtime")              

      val monitoring_stream = monitoring_df.writeStream
                              .trigger(Trigger.ProcessingTime("120 seconds"))
                              .foreachBatch { (batchDF: DataFrame, batchId: Long) =>
                                if(!batchDF.isEmpty) 
                                {
                                    batchDF.persist()
                                    printf("At %d, the %dth microbatch has %d records and %d partitions \n", Instant.now.getEpochSecond, batchId, batchDF.count(), batchDF.rdd.partitions.size)                                    
                                    batchDF.show()

                                    batchDF.write.mode(SaveMode.Overwrite).option("path", "/home/ec2-user/environment/spark/spark-local/tmp").saveAsTable("mytable")
                                    spark.catalog.refreshTable("mytable")
                                    
                                    batchDF.unpersist()
                                    spark.catalog.clearCache()
                                }
                            }
                            .start()
                            .awaitTermination()
    }
   
}

The spark job is tested in the standalone mode by using below spark-submit command, but the same problem exists when I deploy in cluster mode in AWS EMR.

spark-submit --master local[1] --files /home/ec2-user/environment/spark/spark-local/creds/client_jaas.conf,/home/ec2-user/environment/spark/spark-localreds/cacerts,/home/ec2-user/environment/spark/spark-local/creds/krb5.conf,/home/ec2-user/environment/spark/spark-local/creds/my.keytab --driver-java-options "-Djava.security.auth.login.config=/home/ec2-user/environment/spark/spark-local/creds/client_jaas.conf -Djava.security.krb5.conf=/home/ec2-user/environment/spark/spark-local/creds/krb5.conf" --conf spark.dynamicAllocation.enabled=false --conf "spark.executor.extraJavaOptions=-Djava.security.auth.login.config=/home/ec2-user/environment/spark/spark-local/creds/client_jaas.conf -Djava.security.krb5.conf=/home/ec2-user/environment/spark/spark-local/creds/krb5.conf" --conf "spark.driver.extraJavaOptions=-Djava.security.auth.login.config=/home/ec2-user/environment/spark/spark-local/creds/client_jaas.conf -Djava.security.krb5.conf=/home/ec2-user/environment/spark/spark-local/creds/krb5.conf" --conf spark.yarn.maxAppAttempts=1000 --packages org.apache.spark:spark-sql-kafka-0-10_2.11:2.4.0 --class com.example.App ./target/sparktest-1.0-SNAPSHOT-jar-with-dependencies.jar

Then, I started the streaming job to read the streaming data from Kafka topic. After some time, I killed the spark job. Then, I wait for 1 hour to start the job again. If I understand correctly, the new streaming data should start from the offset when I killed the spark job. However, it still starts as the latest offset, which caused data loss during the time I stopped the job.

Do I need to configure more options to avoid data loss? Or do I have some misunderstanding for the Spark 3.0? Thanks!

Problem solved

The key issue here is that the checkpoint must be added to the query specifically. To just add checkpoint for SparkContext is not enough. After adding the checkpoint, it is working. In the checkpoint folder, it will create a offset subfolder, which contains offset file, 0, 1, 2, 3.... For each file, it will show the offset information for different partition.

{"8":109904920,"2":109905750,"5":109905789,"4":109905621,"7":109905330,"1":109905746,"9":109905750,"3":109905936,"6":109905531,"0":109905583}}

One suggestion is to put the checkpoint to some external storage, such as s3. It can help recover the offset even when you need to rebuild the EMR cluster itself in case.

Answer 1

According to the Spark Structured Integration Guide , Spark itself is keeping track of the offsets and there are no offsets committed back to Kafka. That means if your Spark Streaming job fails and you restart it all necessary information on the offsets is stored in Spark's checkpointing files.

Even if you set the ConsumerGroup name with kafka.group.id , your application will still not commit the messages back to Kafka. The information on the next offset to read is only available in the checkpointing files of your Spark application.

If you stop and restart your application without a re-deployment and ensure that you do not delete old checkpoint files, your application will continue reading from where it left off.

In the Spark Structured Streaming documentation on Recovering from Failures with Checkpointing it is written that:

"In case of a failure or intentional shutdown, you can recover the previous progress and state of a previous query, and continue where it left off. This is done using checkpointing and write-ahead logs. You can configure a query with a checkpoint location, and the query will save all the progress information (ie range of offsets processed in each trigger) [...]"

This can be achieved by setting the following option in your writeStream query (it is not sufficient to set the checkpoint directory in your SparkContext configurations):

.option("checkpointLocation", "path/to/HDFS/dir")

In the docs it is also noted that "This checkpoint location has to be a path in an HDFS compatible file system , and can be set as an option in the DataStreamWriter when starting a query."

In addition, the fault tolerance capabilities of Spark Structured Streaming also depends on your output sink as described in section Output Sinks .

As you are currently using the ForeachBatch Sink, you might not have restart capabilities in your application.