Parse CSV as DataFrame/DataSet with Apache Spark a

问题:

I am new to spark, and I want to use group-by & reduce to find the following from CSV (one line by employed):

  Department, Designation, costToCompany, State
  Sales, Trainee, 12000, UP
  Sales, Lead, 32000, AP
  Sales, Lead, 32000, LA
  Sales, Lead, 32000, TN
  Sales, Lead, 32000, AP
  Sales, Lead, 32000, TN 
  Sales, Lead, 32000, LA
  Sales, Lead, 32000, LA
  Marketing, Associate, 18000, TN
  Marketing, Associate, 18000, TN
  HR, Manager, 58000, TN

I would like to simplify the about CSV with group by Department, Designation, State with additional columns with sum(costToCompany) and TotalEmployeeCount

Should get a result like:

  Dept, Desg, state, empCount, totalCost
  Sales,Lead,AP,2,64000
  Sales,Lead,LA,3,96000  
  Sales,Lead,TN,2,64000

Is there any way to achieve this using transformations and actions. Or should we go for RDD operations?

回答1:

Procedure

• Create a Class (Schema) to encapsulate your structure (it’s not required for the approach B, but it would make your code easier to read if you are using Java)

  public class Record implements Serializable {
    String department;
    String designation;
    long costToCompany;
    String state;
    // constructor , getters and setters  
  }
  

• Loading CVS (JSON) file

  JavaSparkContext sc;
  JavaRDD<String> data = sc.textFile("path/input.csv");
  //JavaSQLContext sqlContext = new JavaSQLContext(sc); // For previous versions 
  SQLContext sqlContext = new SQLContext(sc); // In Spark 1.3 the Java API and Scala API have been unified
  
  
  JavaRDD<Record> rdd_records = sc.textFile(data).map(
    new Function<String, Record>() {
        public Record call(String line) throws Exception {
           // Here you can use JSON
           // Gson gson = new Gson();
           // gson.fromJson(line, Record.class);
           String[] fields = line.split(",");
           Record sd = new Record(fields[0], fields[1], fields[2].trim(), fields[3]);
           return sd;
        }
  });
  

At this point you have 2 approaches:

A. SparkSQL

• Register a table (using the your defined Schema Class)

  JavaSchemaRDD table = sqlContext.applySchema(rdd_records, Record.class);
  table.registerAsTable("record_table");
  table.printSchema();
  

• Query the table with your desired Query-group-by

  JavaSchemaRDD res = sqlContext.sql("
    select department,designation,state,sum(costToCompany),count(*) 
    from record_table 
    group by department,designation,state
  ");
  

• Here you would also be able to do any other query you desire, using a SQL approach

B. Spark

• Mapping using a composite key: Department,Designation,State

  JavaPairRDD<String, Tuple2<Long, Integer>> records_JPRDD = 
  rdd_records.mapToPair(new
    PairFunction<Record, String, Tuple2<Long, Integer>>(){
      public Tuple2<String, Tuple2<Long, Integer>> call(Record record){
        Tuple2<String, Tuple2<Long, Integer>> t2 = 
        new Tuple2<String, Tuple2<Long,Integer>>(
          record.Department + record.Designation + record.State,
          new Tuple2<Long, Integer>(record.costToCompany,1)
        );
        return t2;
  }
  

  });

• reduceByKey using the composite key, summing costToCompany column, and accumulating the number of records by key

  JavaPairRDD<String, Tuple2<Long, Integer>> final_rdd_records = 
   records_JPRDD.reduceByKey(new Function2<Tuple2<Long, Integer>, Tuple2<Long,
   Integer>, Tuple2<Long, Integer>>() {
      public Tuple2<Long, Integer> call(Tuple2<Long, Integer> v1,
      Tuple2<Long, Integer> v2) throws Exception {
          return new Tuple2<Long, Integer>(v1._1 + v2._1, v1._2+ v2._2);
      }
  });
  

回答2:

CSV file can be parsed with Spark built-in CSV reader. It will return DataFrame/DataSet on the successful read of the file. On top of DataFrame/DataSet, you apply SQL-like operations easily.

Using Spark 2.x(and above) with Java

Create SparkSession object aka spark

import org.apache.spark.sql.SparkSession;

SparkSession spark = SparkSession
    .builder()
    .appName("Java Spark SQL Example")
    .getOrCreate();

Create Schema for Row with StructType

import org.apache.spark.sql.types.StructType;

StructType schema = new StructType()
    .add("department", "string")
    .add("designation", "string")
    .add("ctc", "long")
    .add("state", "string");

Create dataframe from CSV file and apply schema to it

Dataset<Row> df = spark.read()
    .option("mode", "DROPMALFORMED")
    .schema(schema)
    .csv("hdfs://path/input.csv");

more option on reading data from CSV file

Now we can aggregation on data in 2 ways

1. SQL way

Register a table in spark sql metastore to perform SQL operation

df.createOrReplaceTempView("employee");

Run SQL query on registered dataframe

Dataset<Row> sqlResult = spark.sql(
    "SELECT department, designation, state, SUM(ctc), COUNT(department)" 
        + " FROM employee GROUP BY department, designation, state");

sqlResult.show(); //for testing

We can even execute SQL directly on CSV file with out creating table with Spark SQL

---

2. Object chaining or Programming or Java-like way

Do the necessary import for sql functions

import static org.apache.spark.sql.functions.count;
import static org.apache.spark.sql.functions.sum;

Use groupBy and agg on dataframe/dataset to perform count and sum on data

Dataset<Row> dfResult = df.groupBy("department", "designation", "state")
    .agg(sum("ctc"), count("department"));
// After Spark 1.6 columns mentioned in group by will be added to result by default

dfResult.show();//for testing

dependent libraries

"org.apache.spark" % "spark-core_2.11" % "2.0.0" 
"org.apache.spark" % "spark-sql_2.11" % "2.0.0"

回答3:

The following might not be entirely correct, but it should give you some idea of how to juggle data. It's not pretty, should be replaced with case classes etc, but as a quick example of how to use the spark api, I hope it's enough :)

val rawlines = sc.textfile("hdfs://.../*.csv")
case class Employee(dep: String, des: String, cost: Double, state: String)
val employees = rawlines
  .map(_.split(",") /*or use a proper CSV parser*/
  .map( Employee(row(0), row(1), row(2), row(3) )

# the 1 is the amount of employees (which is obviously 1 per line)
val keyVals = employees.map( em => (em.dep, em.des, em.state), (1 , em.cost))

val results = keyVals.reduceByKey{ a,b =>
    (a._1 + b._1, b._1, b._2) # (a.count + b.count , a.cost + b.cost )
}

#debug output
results.take(100).foreach(println)

results
  .map( keyval => someThingToFormatAsCsvStringOrWhatever )
  .saveAsTextFile("hdfs://.../results")

Or you can use SparkSQL:

val sqlContext = new SQLContext(sparkContext)

# case classes can easily be registered as tables
employees.registerAsTable("employees")

val results = sqlContext.sql("""select dep, des, state, sum(cost), count(*) 
  from employees 
  group by dep,des,state"""

回答4:

For JSON, if your text file contains one JSON object per line, you can use sqlContext.jsonFile(path) to let Spark SQL load it as a SchemaRDD (the schema will be automatically inferred). Then, you can register it as a table and query it with SQL. You can also manually load the text file as an RDD[String] containing one JSON object per record and use sqlContext.jsonRDD(rdd) to turn it as a SchemaRDD. jsonRDD is useful when you need to pre-process your data.