Aggregating multiple columns with custom function

问题:

I was wondering if there is some way to specify a custom aggregation function for spark dataframes over multiple columns.

I have a table like this of the type (name, item, price):

john | tomato | 1.99
john | carrot | 0.45
bill | apple  | 0.99
john | banana | 1.29
bill | taco   | 2.59

to:

I would like to aggregate the item and it's cost for each person into a list like this:

john | (tomato, 1.99), (carrot, 0.45), (banana, 1.29)
bill | (apple, 0.99), (taco, 2.59)

Is this possible in dataframes? I recently learned about collect_list but it appears to only work for one column.

回答1:

The easiest way to do this as a DataFrame is to first collect two lists, and then use a UDF to zip the two lists together. Something like:

import org.apache.spark.sql.functions.{collect_list, udf}
import sqlContext.implicits._

val zipper = udf[Seq[(String, Double)], Seq[String], Seq[Double]](_.zip(_))

val df = Seq(
  ("john", "tomato", 1.99),
  ("john", "carrot", 0.45),
  ("bill", "apple", 0.99),
  ("john", "banana", 1.29),
  ("bill", "taco", 2.59)
).toDF("name", "food", "price")

val df2 = df.groupBy("name").agg(
  collect_list(col("food")) as "food",
  collect_list(col("price")) as "price" 
).withColumn("food", zipper(col("food"), col("price"))).drop("price")

df2.show(false)
# +----+---------------------------------------------+
# |name|food                                         |
# +----+---------------------------------------------+
# |john|[[tomato,1.99], [carrot,0.45], [banana,1.29]]|
# |bill|[[apple,0.99], [taco,2.59]]                  |
# +----+---------------------------------------------+

回答2:

Consider using the struct function to group the columns together before collecting as a list:

import org.apache.spark.sql.functions.{collect_list, struct}
import sqlContext.implicits._

val df = Seq(
  ("john", "tomato", 1.99),
  ("john", "carrot", 0.45),
  ("bill", "apple", 0.99),
  ("john", "banana", 1.29),
  ("bill", "taco", 2.59)
).toDF("name", "food", "price")

df.groupBy($"name")
  .agg(collect_list(struct($"food", $"price")).as("foods"))
  .show(false)

Outputs:

+----+---------------------------------------------+
|name|foods                                        |
+----+---------------------------------------------+
|john|[[tomato,1.99], [carrot,0.45], [banana,1.29]]|
|bill|[[apple,0.99], [taco,2.59]]                  |
+----+---------------------------------------------+

回答3:

Maybe a better way than the zip function (since UDF and UDAF are very bad to performance) is to wrap the two columns into Struct.

This would probably work as well:

df.select('name, struct('food, 'price).as("tuple"))
  .groupBy('name)
  .agg(collect_list('tuple).as("tuples"))

回答4:

Here is an option by converting the data frame to a RDD of Map and then call a groupByKey on it. The result would be a list of key-value pairs where value is a list of tuples.

df.show
+----+------+----+
|  _1|    _2|  _3|
+----+------+----+
|john|tomato|1.99|
|john|carrot|0.45|
|bill| apple|0.99|
|john|banana|1.29|
|bill|  taco|2.59|
+----+------+----+


val tuples = df.map(row => row(0) -> (row(1), row(2)))
tuples: org.apache.spark.rdd.RDD[(Any, (Any, Any))] = MapPartitionsRDD[102] at map at <console>:43

tuples.groupByKey().map{ case(x, y) => (x, y.toList) }.collect
res76: Array[(Any, List[(Any, Any)])] = Array((bill,List((apple,0.99), (taco,2.59))), (john,List((tomato,1.99), (carrot,0.45), (banana,1.29))))