[SPARK-36917] java.lang.ClassCastException: cannot assign instance of java.lang.invoke.SerializedLambda to field org.apache.spark.rdd.MapPartitionsRDD.f of type scala.Function3 in instance of org.apache.spark.rdd.MapPartitionsRDD

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Not A Problem
• Affects Version/s: 3.1.2
• Component/s: Spark Core, Spark Submit
• Labels: hadoop java serializable spark spark-conf ubuntu

Environment

Ubuntu 20
Spark3.1.2-hadoop3.2
Hadoop 3.1

Language

java

• Assignee: Unassigned
• Reporter: Davide Benedetto
• Created: 03/Oct/21 20:42
• Updated: 14/Aug/23 23:36
• Resolved: 14/Aug/23 23:36

Description

My spark Job fails with this error:

org.apache.spark.SparkException: Job aborted due to stage failure: Task 0 in stage 0.0 failed 4 times, most recent failure: Lost task 0.3 in stage 0.0 (TID 3) (davben-lubuntu executor 2): java.lang.ClassCastException: cannot assign instance of java.lang.invoke.SerializedLambda to field org.apache.spark.rdd.MapPartitionsRDD.f of type scala.Function3 in instance of org.apache.spark.rdd.MapPartitionsRDD

My OS Linux Ubuntu 20 is in this way organized: I have two user: /home/davben and /home/hadoop. Into hadoop user I have installed hadoop 3.1 and spark-3.1.2-hadoop3.2.  Both users refers to java-8-openjdk Java installation. The Spark job is launched from user davben on eclipse IDE  in this way:
I create the spark conf and the spark session

 

System.setProperty("hadoop.home.dir", "/home/hadoop/hadoop");
SparkConf sparkConf = new SparkConf()
.setAppName("simple")
.setMaster("yarn")
.set("spark.executor.memory", "1g")
.set("deploy.mode", "cluster")
.set("spark.yarn.stagingDir", "hdfs://localhost:9000/user/hadoop/") .set("spark.hadoop.fs.defaultFS","hdfs://localhost:9000") .set("spark.hadoop.yarn.resourcemanager.hostname","localhost") .set("spark.hadoop.yarn.resourcemanager.scheduler.address","localhost:8030") .set("spark.hadoop.yarn.resourcemanager.address ","localhost:8032") .set("spark.hadoop.yarn.resourcemanager.webapp.address","localhost:8088") .set("spark.hadoop.yarn.resourcemanager.admin.address","localhost:8083")
SparkSession spark = SparkSession.builder().config(sparkConf).getOrCreate();

Then I create a dataset with two entries:

 

List<Row> rows = new ArrayList<>(); 
rows.add(RowFactory.create("a", "b"));
rows.add(RowFactory.create("a", "a"));
StructType structType = new StructType(); 
structType = structType.add("edge_1", DataTypes.StringType, false);
structType = structType.add("edge_2", DataTypes.StringType, false); ExpressionEncoder<Row> edgeEncoder = RowEncoder.apply(structType);
Dataset<Row> edge = spark.createDataset(rows, edgeEncoder);

Then I print the content of the current dataset edge

 edge.show();

 

Then I perform a map transformation on edge that upper cases the values of the two entries and return the result in edge2

 Dataset<Row> edge2 = edge.map(new MyFunction2(), edgeEncoder);

The following is the code of MyFunction2

public class MyFunction2 implements MapFunction<Row, Row>, scala.Serializable { 
private static final long serialVersionUID = 1L;

@Override public Row call(Row v1) throws Exception { 
String el1 = v1.get(0).toString().toUpperCase(); 
String el2 = v1.get(1).toString().toUpperCase(); 
return RowFactory.create(el1,el2); 
}
}

Finally I show the content of edge2

edge2.show();

I can confirm that, checking on the hadoop UI a localhost:8088, the job is submitted correctly, and
what sounds strange is that the first show is returned correctly in my console, but the second one fails returning the up mentioned error.