使用 Java8 的 Spark 2.3 将行转换为列

Question

I am new to Spark 2.4 with Java 8. I need help. Here is example of instances:

Source DataFrame

+--------------+
| key | Value  |
+--------------+
| A   | John   |
| B   | Nick   |
| A   | Mary   |
| B   | Kathy  |
| C   | Sabrina|
| B   | George |
+--------------+

Meta DataFrame

+-----+
| key |
+-----+
| A   |
| B   |
| C   |
| D   |
| E   |
| F   |
+-----+

I would like to transform it to the following: Column names from Meta Dataframe and Rows will be transformed based on Source Dataframe

+-----------------------------------------------+
| A    | B      | C       | D     | E    | F    |
+-----------------------------------------------+
| John | Nick   | Sabrina | null  | null | null |
| Mary | Kathy  | null    | null  | null | null |
| null | George | null    | null  | null | null |
+-----------------------------------------------+

Need to write a code Spark 2.3 with Java8. Appreciated your help.

Answer 1

To make things clearer (and easily reproducible) let's define dataframes:

val df1 = Seq("A" -> "John", "B" -> "Nick", "A" -> "Mary", 
              "B" -> "Kathy", "C" -> "Sabrina", "B" -> "George")
          .toDF("key", "value")
val df2 = Seq("A", "B", "C", "D", "E", "F").toDF("key")

From what I see, you are trying to create one column by value in the key column of df2 . These columns should contain all the values of the value column that are associated to the key naming the column. If we take an example, column A 's first value should be the value of the first occurrence of A (if it exists, null otherwise): "John" . Its second value should be the value of the second occurrence of A: "Mary" . There is no third value so the third value of the column should be null .

I detailed it to show that we need a notion of rank of the values for each key (windowing function), and group by that notion of rank. It would go as follows:

import org.apache.spark.sql.expressions.Window
val df1_win = df1
    .withColumn("id", monotonically_increasing_id)
    .withColumn("rank", rank() over Window.partitionBy("key").orderBy("id"))
// the id is just here to maintain the original order.

// getting the keys in df2. Add distinct if there are duplicates.
val keys = df2.collect.map(_.getAs[String](0)).sorted

// then it's just about pivoting
df1_win
    .groupBy("rank")
    .pivot("key", keys) 
    .agg(first('value))
    .orderBy("rank")
    //.drop("rank") // I keep here it for clarity
    .show()
+----+----+------+-------+----+----+----+                                       
|rank|   A|     B|      C|   D|   E|   F|
+----+----+------+-------+----+----+----+
|   1|John|  Nick|Sabrina|null|null|null|
|   2|Mary| Kathy|   null|null|null|null|
|   3|null|George|   null|null|null|null|
+----+----+------+-------+----+----+----+

Here is the very same code in Java

Dataset<Row> df1_win = df1
    .withColumn("id", functions.monotonically_increasing_id())
    .withColumn("rank", functions.rank().over(Window.partitionBy("key").orderBy("id")));
    // the id is just here to maintain the original order.

// getting the keys in df2. Add distinct if there are duplicates.
// Note that it is a list of objects, to match the (strange) signature of pivot
List<Object> keys = df2.collectAsList().stream()
    .map(x -> x.getString(0))
    .sorted().collect(Collectors.toList());

// then it's just about pivoting
df1_win
    .groupBy("rank")
    .pivot("key", keys)
    .agg(functions.first(functions.col("value")))
    .orderBy("rank")
    // .drop("rank") // I keep here it for clarity
    .show();