How to select the first row of each group?

Rami picture Rami · Nov 23, 2015 · Viewed 113.4k times · Source

I have a DataFrame generated as follow:

df.groupBy($"Hour", $"Category")
  .agg(sum($"value") as "TotalValue")
  .sort($"Hour".asc, $"TotalValue".desc))

The results look like:

+----+--------+----------+
|Hour|Category|TotalValue|
+----+--------+----------+
|   0|   cat26|      30.9|
|   0|   cat13|      22.1|
|   0|   cat95|      19.6|
|   0|  cat105|       1.3|
|   1|   cat67|      28.5|
|   1|    cat4|      26.8|
|   1|   cat13|      12.6|
|   1|   cat23|       5.3|
|   2|   cat56|      39.6|
|   2|   cat40|      29.7|
|   2|  cat187|      27.9|
|   2|   cat68|       9.8|
|   3|    cat8|      35.6|
| ...|    ....|      ....|
+----+--------+----------+

As you can see, the DataFrame is ordered by Hour in an increasing order, then by TotalValue in a descending order.

I would like to select the top row of each group, i.e.

  • from the group of Hour==0 select (0,cat26,30.9)
  • from the group of Hour==1 select (1,cat67,28.5)
  • from the group of Hour==2 select (2,cat56,39.6)
  • and so on

So the desired output would be:

+----+--------+----------+
|Hour|Category|TotalValue|
+----+--------+----------+
|   0|   cat26|      30.9|
|   1|   cat67|      28.5|
|   2|   cat56|      39.6|
|   3|    cat8|      35.6|
| ...|     ...|       ...|
+----+--------+----------+

It might be handy to be able to select the top N rows of each group as well.

Any help is highly appreciated.

Answer

zero323 picture zero323 · Nov 23, 2015

Window functions:

Something like this should do the trick:

import org.apache.spark.sql.functions.{row_number, max, broadcast}
import org.apache.spark.sql.expressions.Window

val df = sc.parallelize(Seq(
  (0,"cat26",30.9), (0,"cat13",22.1), (0,"cat95",19.6), (0,"cat105",1.3),
  (1,"cat67",28.5), (1,"cat4",26.8), (1,"cat13",12.6), (1,"cat23",5.3),
  (2,"cat56",39.6), (2,"cat40",29.7), (2,"cat187",27.9), (2,"cat68",9.8),
  (3,"cat8",35.6))).toDF("Hour", "Category", "TotalValue")

val w = Window.partitionBy($"hour").orderBy($"TotalValue".desc)

val dfTop = df.withColumn("rn", row_number.over(w)).where($"rn" === 1).drop("rn")

dfTop.show
// +----+--------+----------+
// |Hour|Category|TotalValue|
// +----+--------+----------+
// |   0|   cat26|      30.9|
// |   1|   cat67|      28.5|
// |   2|   cat56|      39.6|
// |   3|    cat8|      35.6|
// +----+--------+----------+

This method will be inefficient in case of significant data skew.

Plain SQL aggregation followed by join:

Alternatively you can join with aggregated data frame:

val dfMax = df.groupBy($"hour".as("max_hour")).agg(max($"TotalValue").as("max_value"))

val dfTopByJoin = df.join(broadcast(dfMax),
    ($"hour" === $"max_hour") && ($"TotalValue" === $"max_value"))
  .drop("max_hour")
  .drop("max_value")

dfTopByJoin.show

// +----+--------+----------+
// |Hour|Category|TotalValue|
// +----+--------+----------+
// |   0|   cat26|      30.9|
// |   1|   cat67|      28.5|
// |   2|   cat56|      39.6|
// |   3|    cat8|      35.6|
// +----+--------+----------+

It will keep duplicate values (if there is more than one category per hour with the same total value). You can remove these as follows:

dfTopByJoin
  .groupBy($"hour")
  .agg(
    first("category").alias("category"),
    first("TotalValue").alias("TotalValue"))

Using ordering over structs:

Neat, although not very well tested, trick which doesn't require joins or window functions:

val dfTop = df.select($"Hour", struct($"TotalValue", $"Category").alias("vs"))
  .groupBy($"hour")
  .agg(max("vs").alias("vs"))
  .select($"Hour", $"vs.Category", $"vs.TotalValue")

dfTop.show
// +----+--------+----------+
// |Hour|Category|TotalValue|
// +----+--------+----------+
// |   0|   cat26|      30.9|
// |   1|   cat67|      28.5|
// |   2|   cat56|      39.6|
// |   3|    cat8|      35.6|
// +----+--------+----------+

With DataSet API (Spark 1.6+, 2.0+):

Spark 1.6:

case class Record(Hour: Integer, Category: String, TotalValue: Double)

df.as[Record]
  .groupBy($"hour")
  .reduce((x, y) => if (x.TotalValue > y.TotalValue) x else y)
  .show

// +---+--------------+
// | _1|            _2|
// +---+--------------+
// |[0]|[0,cat26,30.9]|
// |[1]|[1,cat67,28.5]|
// |[2]|[2,cat56,39.6]|
// |[3]| [3,cat8,35.6]|
// +---+--------------+

Spark 2.0 or later:

df.as[Record]
  .groupByKey(_.Hour)
  .reduceGroups((x, y) => if (x.TotalValue > y.TotalValue) x else y)

The last two methods can leverage map side combine and don't require full shuffle so most of the time should exhibit a better performance compared to window functions and joins. These cane be also used with Structured Streaming in completed output mode.

Don't use:

df.orderBy(...).groupBy(...).agg(first(...), ...)

It may seem to work (especially in the local mode) but it is unreliable (see SPARK-16207, credits to Tzach Zohar for linking relevant JIRA issue, and SPARK-30335).

The same note applies to

df.orderBy(...).dropDuplicates(...)

which internally uses equivalent execution plan.