collect_list by preserving order based on another variable

from pyspark.sql import functions as F
from pyspark.sql import Window

w = Window.partitionBy('id').orderBy('date')

sorted_list_df = input_df.withColumn(
            'sorted_list', F.collect_list('value').over(w)
        )\
        .groupBy('id')\
        .agg(F.max('sorted_list').alias('sorted_list'))

Window examples provided by users often don't really explain what is going on so let me dissect it for you.

As you know, using collect_list together with groupBy will result in an unordered list of values. This is because depending on how your data is partitioned, Spark will append values to your list as soon as it finds a row in the group. The order then depends on how Spark plans your aggregation over the executors.

A Window function allows you to control that situation, grouping rows by a certain value so you can perform an operation over each of the resultant groups:

w = Window.partitionBy('id').orderBy('date')

partitionBy - you want groups/partitions of rows with the same id

orderBy - you want each row in the group to be sorted by date

Once you have defined the scope of your Window - "rows with the same id, sorted by date" -, you can use it to perform an operation over it, in this case, a collect_list:

F.collect_list('value').over(w)

At this point you created a new column sorted_list with an ordered list of values, sorted by date, but you still have duplicated rows per id. To trim out the duplicated rows you want to groupBy id and keep the max value in for each group:

.groupBy('id')\
.agg(F.max('sorted_list').alias('sorted_list'))

---

If you collect both dates and values as a list, you can sort the resulting column according to date using and udf, and then keep only the values in the result.

import operator
import pyspark.sql.functions as F

# create list column
grouped_df = input_df.groupby("id") \
               .agg(F.collect_list(F.struct("date", "value")) \
               .alias("list_col"))

# define udf
def sorter(l):
  res = sorted(l, key=operator.itemgetter(0))
  return [item[1] for item in res]

sort_udf = F.udf(sorter)

# test
grouped_df.select("id", sort_udf("list_col") \
  .alias("sorted_list")) \
  .show(truncate = False)
+---+----------------+
|id |sorted_list     |
+---+----------------+
|1  |[10, 5, 15, 20] |
|2  |[100, 500, 1500]|
+---+----------------+

---

You can use sort_array function. If you collect both dates and values as a list, you can sort the resulting column using sort_array and keep only the columns you require.

import operator
import pyspark.sql.functions as F

grouped_df = input_df.groupby("id") \
               .agg(F.sort_array(F.collect_list(F.struct("date", "value"))) \
.alias("collected_list")) \
.withColumn("sorted_list",col("collected_list.value")) \
.drop("collected_list")
.show(truncate=False)

+---+----------------+
|id |sorted_list     |
+---+----------------+
|1  |[10, 5, 15, 20] |
|2  |[100, 500, 1500]|
+---+----------------+ ```````

---

The question was for PySpark but might be helpful to have it also for Scala Spark.

Let's prepare test dataframe:

import org.apache.spark.sql.functions._
import org.apache.spark.sql.{DataFrame, Row, SparkSession}
import org.apache.spark.sql.expressions.{ Window, UserDefinedFunction}

import java.sql.Date
import java.time.LocalDate

val spark: SparkSession = ...

// Out test data set
val data: Seq[(Int, Date, Int)] = Seq(
  (1, Date.valueOf(LocalDate.parse("2014-01-03")), 10),
  (1, Date.valueOf(LocalDate.parse("2014-01-04")), 5),
  (1, Date.valueOf(LocalDate.parse("2014-01-05")), 15),
  (1, Date.valueOf(LocalDate.parse("2014-01-06")), 20),
  (2, Date.valueOf(LocalDate.parse("2014-02-10")), 100),
  (2, Date.valueOf(LocalDate.parse("2014-02-11")), 500),
  (2, Date.valueOf(LocalDate.parse("2014-02-15")), 1500)
)

// Create dataframe
val df: DataFrame = spark.createDataFrame(data)
  .toDF("id", "date", "value")
df.show()
//+---+----------+-----+
//| id|      date|value|
//+---+----------+-----+
//|  1|2014-01-03|   10|
//|  1|2014-01-04|    5|
//|  1|2014-01-05|   15|
//|  1|2014-01-06|   20|
//|  2|2014-02-10|  100|
//|  2|2014-02-11|  500|
//|  2|2014-02-15| 1500|
//+---+----------+-----+

Use UDF

// Group by id and aggregate date and value to new column date_value
val grouped = df.groupBy(col("id"))
  .agg(collect_list(struct("date", "value")) as "date_value")
grouped.show()
grouped.printSchema()
// +---+--------------------+
// | id|          date_value|
// +---+--------------------+
// |  1|[[2014-01-03,10],...|
// |  2|[[2014-02-10,100]...|
// +---+--------------------+

// udf to extract data from Row, sort by needed column (date) and return value
val sortUdf: UserDefinedFunction = udf((rows: Seq[Row]) => {
  rows.map { case Row(date: Date, value: Int) => (date, value) }
    .sortBy { case (date, value) => date }
    .map { case (date, value) => value }
})

// Select id and value_list
val r1 = grouped.select(col("id"), sortUdf(col("date_value")).alias("value_list"))
r1.show()
// +---+----------------+
// | id|      value_list|
// +---+----------------+
// |  1| [10, 5, 15, 20]|
// |  2|[100, 500, 1500]|
// +---+----------------+

Use Window

val window = Window.partitionBy(col("id")).orderBy(col("date"))
val sortedDf = df.withColumn("values_sorted_by_date", collect_list("value").over(window))
sortedDf.show()
//+---+----------+-----+---------------------+
//| id|      date|value|values_sorted_by_date|
//+---+----------+-----+---------------------+
//|  1|2014-01-03|   10|                 [10]|
//|  1|2014-01-04|    5|              [10, 5]|
//|  1|2014-01-05|   15|          [10, 5, 15]|
//|  1|2014-01-06|   20|      [10, 5, 15, 20]|
//|  2|2014-02-10|  100|                [100]|
//|  2|2014-02-11|  500|           [100, 500]|
//|  2|2014-02-15| 1500|     [100, 500, 1500]|
//+---+----------+-----+---------------------+

val r2 = sortedDf.groupBy(col("id"))
  .agg(max("values_sorted_by_date").as("value_list")) 
r2.show()
//+---+----------------+
//| id|      value_list|
//+---+----------------+
//|  1| [10, 5, 15, 20]|
//|  2|[100, 500, 1500]|
//+---+----------------+

---

To make sure the sort is done for each id, we can use sortWithinPartitions:

from pyspark.sql import functions as F
ordered_df = (
    input_df
        .repartition(input_df.id)
        .sortWithinPartitions(['date'])


)
grouped_df = ordered_df.groupby("id").agg(F.collect_list("value"))

---

I tried TMichel approach and didn't work for me. When I performed the max aggregation I wasn't getting back the highest value of the list. So what worked for me is the following:

def max_n_values(df, key, col_name, number):
    '''
    Returns the max n values of a spark dataframe
    partitioned by the key and ranked by the col_name
    '''
    w2 = Window.partitionBy(key).orderBy(f.col(col_name).desc())
    output = df.select('*',
                       f.row_number().over(w2).alias('rank')).filter(
                           f.col('rank') <= number).drop('rank')
    return output

def col_list(df, key, col_to_collect, name, score):
    w = Window.partitionBy(key).orderBy(f.col(score).desc())

    list_df = df.withColumn(name, f.collect_set(col_to_collect).over(w))
    size_df = list_df.withColumn('size', f.size(name))
    output = max_n_values(df=size_df,
                               key=key,
                               col_name='size',
                               number=1)
    return output

---

As of Spark 2.4, the collect_list(ArrayType) created in @mtoto's answer can be post-processed by using SparkSQL's builtin functions transform and array_sort (no need for udf):

from pyspark.sql.functions import collect_list, expr, struct

df.groupby('id') \
  .agg(collect_list(struct('date','value')).alias('value_list')) \
  .withColumn('value_list', expr('transform(array_sort(value_list), x -> x.value)')) \
  .show()
+---+----------------+
| id|      value_list|
+---+----------------+
|  1| [10, 5, 15, 20]|
|  2|[100, 500, 1500]|
+---+----------------+ 

Note: if descending order is required change array_sort(value_list) to sort_array(value_list, False)

Caveat: array_sort() and sort_array() won't work if items(in collect_list) must be sorted by multiple fields(columns) in a mixed order, i.e. orderBy('col1', desc('col2')).

---

In the Spark SQL world the answer to this would be:

SELECT 
browser, max(list)
from (
  SELECT
    id,
    COLLECT_LIST(value) OVER (PARTITION BY id ORDER BY date DESC) as list
  FROM browser_count
  GROUP BYid, value, date) 
Group by browser;

---

if you want to use spark sql here is how you can achieve this. Assuming the table name (or temporary view) is temp_table.

select
t1.id,
collect_list(value) as value_list
(Select * from temp_table order by id,date) t1
group by 1

---

Complementing what ShadyStego said, I've been testing the usage of sortWithinPartitions and GroupBy on Spark, finding out it performs quite better than Window functions or UDF. Still, there is an issue with a missordering once per partition when using this method, but it can be easily solved. I show it here Spark (pySpark) groupBy misordering first element on collect_list.

This method is specially useful on large DataFrames, but a large number of partitions may be needed if you are short on driver memory.

---