How to pivot Spark DataFrame?

Question

I am starting to use Spark DataFrames and I need to be able to pivot the data to create multiple columns out of 1 column with multiple rows. There is built in functionality

Answer 1

I overcame this by writing a for loop to dynamically create a SQL query. Say I have:

id  tag  value
1   US    50
1   UK    100
1   Can   125
2   US    75
2   UK    150
2   Can   175

and I want:

id  US  UK   Can
1   50  100  125
2   75  150  175

I can create a list with the value I want to pivot and then create a string containing the SQL query I need.

val countries = List("US", "UK", "Can")
val numCountries = countries.length - 1

var query = "select *, "
for (i <- 0 to numCountries-1) {
  query += """case when tag = """" + countries(i) + """" then value else 0 end as """ + countries(i) + ", "
}
query += """case when tag = """" + countries.last + """" then value else 0 end as """ + countries.last + " from myTable"

myDataFrame.registerTempTable("myTable")
val myDF1 = sqlContext.sql(query)

I can create similar query to then do the aggregation. Not a very elegant solution but it works and is flexible for any list of values, which can also be passed in as an argument when your code is called.

Answer 2

As mentioned by David Anderson Spark provides pivot function since version 1.6. General syntax looks as follows:

df
  .groupBy(grouping_columns)
  .pivot(pivot_column, [values]) 
  .agg(aggregate_expressions)

Usage examples using nycflights13 and csv format:

Python:

from pyspark.sql.functions import avg

flights = (sqlContext
    .read
    .format("csv")
    .options(inferSchema="true", header="true")
    .load("flights.csv")
    .na.drop())

flights.registerTempTable("flights")
sqlContext.cacheTable("flights")

gexprs = ("origin", "dest", "carrier")
aggexpr = avg("arr_delay")

flights.count()
## 336776

%timeit -n10 flights.groupBy(*gexprs ).pivot("hour").agg(aggexpr).count()
## 10 loops, best of 3: 1.03 s per loop

Scala:

val flights = sqlContext
  .read
  .format("csv")
  .options(Map("inferSchema" -> "true", "header" -> "true"))
  .load("flights.csv")

flights
  .groupBy($"origin", $"dest", $"carrier")
  .pivot("hour")
  .agg(avg($"arr_delay"))

Java:

import static org.apache.spark.sql.functions.*;
import org.apache.spark.sql.*;

Dataset<Row> df = spark.read().format("csv")
        .option("inferSchema", "true")
        .option("header", "true")
        .load("flights.csv");

df.groupBy(col("origin"), col("dest"), col("carrier"))
        .pivot("hour")
        .agg(avg(col("arr_delay")));

R / SparkR:

library(magrittr)

flights <- read.df("flights.csv", source="csv", header=TRUE, inferSchema=TRUE)

flights %>% 
  groupBy("origin", "dest", "carrier") %>% 
  pivot("hour") %>% 
  agg(avg(column("arr_delay")))

R / sparklyr

library(dplyr)

flights <- spark_read_csv(sc, "flights", "flights.csv")

avg.arr.delay <- function(gdf) {
   expr <- invoke_static(
      sc,
      "org.apache.spark.sql.functions",
      "avg",
      "arr_delay"
    )
    gdf %>% invoke("agg", expr, list())
}

flights %>% 
  sdf_pivot(origin + dest + carrier ~  hour, fun.aggregate=avg.arr.delay)

SQL:

Note that PIVOT keyword in Spark SQL is supported starting from version 2.4.

CREATE TEMPORARY VIEW flights 
USING csv 
OPTIONS (header 'true', path 'flights.csv', inferSchema 'true') ;

 SELECT * FROM (
   SELECT origin, dest, carrier, arr_delay, hour FROM flights
 ) PIVOT (
   avg(arr_delay)
   FOR hour IN (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
                13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23)
 );

Example data:

"year","month","day","dep_time","sched_dep_time","dep_delay","arr_time","sched_arr_time","arr_delay","carrier","flight","tailnum","origin","dest","air_time","distance","hour","minute","time_hour"
2013,1,1,517,515,2,830,819,11,"UA",1545,"N14228","EWR","IAH",227,1400,5,15,2013-01-01 05:00:00
2013,1,1,533,529,4,850,830,20,"UA",1714,"N24211","LGA","IAH",227,1416,5,29,2013-01-01 05:00:00
2013,1,1,542,540,2,923,850,33,"AA",1141,"N619AA","JFK","MIA",160,1089,5,40,2013-01-01 05:00:00
2013,1,1,544,545,-1,1004,1022,-18,"B6",725,"N804JB","JFK","BQN",183,1576,5,45,2013-01-01 05:00:00
2013,1,1,554,600,-6,812,837,-25,"DL",461,"N668DN","LGA","ATL",116,762,6,0,2013-01-01 06:00:00
2013,1,1,554,558,-4,740,728,12,"UA",1696,"N39463","EWR","ORD",150,719,5,58,2013-01-01 05:00:00
2013,1,1,555,600,-5,913,854,19,"B6",507,"N516JB","EWR","FLL",158,1065,6,0,2013-01-01 06:00:00
2013,1,1,557,600,-3,709,723,-14,"EV",5708,"N829AS","LGA","IAD",53,229,6,0,2013-01-01 06:00:00
2013,1,1,557,600,-3,838,846,-8,"B6",79,"N593JB","JFK","MCO",140,944,6,0,2013-01-01 06:00:00
2013,1,1,558,600,-2,753,745,8,"AA",301,"N3ALAA","LGA","ORD",138,733,6,0,2013-01-01 06:00:00

Performance considerations:

Generally speaking pivoting is an expensive operation.

• if you can, try to provide values list, as this avoids an extra hit to compute the uniques:

  vs = list(range(25))
  %timeit -n10 flights.groupBy(*gexprs ).pivot("hour", vs).agg(aggexpr).count()
  ## 10 loops, best of 3: 392 ms per loop
  

• in some cases it proved to be beneficial (likely no longer worth the effort in 2.0 or later) to repartition and / or pre-aggregate the data

• for reshaping only, you can use first: Pivot String column on Pyspark Dataframe

Related questions:

• How to melt Spark DataFrame?
• Unpivot in spark-sql/pyspark
• Transpose column to row with Spark

Answer 3

The built-in spark pivot function is inefficient. The bellow implementation works on spark 2.4+ - the idea is to aggregate a map and extract the values as columns. The only limitation is it does not handle aggregate function in the pivoted columns, only column(s).

On a 8M table, those functions applies on 3 secondes, versus 40 minutes in the built-in spark version:

# pass an optional list of string to avoid computation of columns
def pivot(df, group_by, key, aggFunction, levels=[]):
    if not levels:
        levels = [row[key] for row in df.filter(col(key).isNotNull()).groupBy(col(key)).agg(count(key)).select(key).collect()]
    return df.filter(col(key).isin(*levels) == True).groupBy(group_by).agg(map_from_entries(collect_list(struct(key, expr(aggFunction)))).alias("group_map")).select([group_by] + ["group_map." + l for l in levels])

# Usage
pivot(df, "id", "key", "value")
pivot(df, "id", "key", "array(value)")

// pass an optional list of string to avoid computation of columns
  def pivot(df: DataFrame, groupBy: Column, key: Column, aggFunct: String, _levels: List[String] = Nil): DataFrame = {
    val levels =
      if (_levels.isEmpty) df.filter(key.isNotNull).select(key).distinct().collect().map(row => row.getString(0)).toList
      else _levels

    df
      .filter(key.isInCollection(levels))
      .groupBy(groupBy)
      .agg(map_from_entries(collect_list(struct(key, expr(aggFunct)))).alias("group_map"))
      .select(groupBy.toString, levels.map(f => "group_map." + f): _*)
  }

// Usage:
pivot(df, col("id"), col("key"), "value")
pivot(df, col("id"), col("key"), "array(value)")

Answer 4

A pivot operator has been added to the Spark dataframe API, and is part of Spark 1.6.

See https://github.com/apache/spark/pull/7841 for details.

Answer 5

There is simple and elegant solution.

scala> spark.sql("select * from k_tags limit 10").show()
+---------------+-------------+------+
|           imsi|         name| value|
+---------------+-------------+------+
|246021000000000|          age|    37|
|246021000000000|       gender|Female|
|246021000000000|         arpu|    22|
|246021000000000|   DeviceType| Phone|
|246021000000000|DataAllowance|   6GB|
+---------------+-------------+------+

scala> spark.sql("select * from k_tags limit 10").groupBy($"imsi").pivot("name").agg(min($"value")).show()
+---------------+-------------+----------+---+----+------+
|           imsi|DataAllowance|DeviceType|age|arpu|gender|
+---------------+-------------+----------+---+----+------+
|246021000000000|          6GB|     Phone| 37|  22|Female|
|246021000000001|          1GB|     Phone| 72|  10|  Male|
+---------------+-------------+----------+---+----+------+

Answer 6

I have solved a similar problem using dataframes with the following steps:

Create columns for all your countries, with 'value' as the value:

import org.apache.spark.sql.functions._
val countries = List("US", "UK", "Can")
val countryValue = udf{(countryToCheck: String, countryInRow: String, value: Long) =>
  if(countryToCheck == countryInRow) value else 0
}
val countryFuncs = countries.map{country => (dataFrame: DataFrame) => dataFrame.withColumn(country, countryValue(lit(country), df("tag"), df("value"))) }
val dfWithCountries = Function.chain(countryFuncs)(df).drop("tag").drop("value")

Your dataframe 'dfWithCountries' will look like this:

+--+--+---+---+
|id|US| UK|Can|
+--+--+---+---+
| 1|50|  0|  0|
| 1| 0|100|  0|
| 1| 0|  0|125|
| 2|75|  0|  0|
| 2| 0|150|  0|
| 2| 0|  0|175|
+--+--+---+---+

Now you can sum together all the values for your desired result:

dfWithCountries.groupBy("id").sum(countries: _*).show

Result:

+--+-------+-------+--------+
|id|SUM(US)|SUM(UK)|SUM(Can)|
+--+-------+-------+--------+
| 1|     50|    100|     125|
| 2|     75|    150|     175|
+--+-------+-------+--------+

It's not a very elegant solution though. I had to create a chain of functions to add in all the columns. Also if I have lots of countries, I will expand my temporary data set to a very wide set with lots of zeroes.