Retrieve top n in each group of a DataFrame in pyspark

Question

There's a DataFrame in pyspark with data as below:

user_id object_id score
user_1  object_1  3
user_1  object_1  1
user_1  object_2  2
user_2  object_1  5
user_2  object_2  2
user_2  object_2  6

What I expect is returning 2 records in each group with the same user_id, which need to have the highest score. Consequently, the result should look as the following:

user_id object_id score
user_1  object_1  3
user_1  object_2  2
user_2  object_2  6
user_2  object_1  5

I'm really new to pyspark, could anyone give me a code snippet or portal to the related documentation of this problem? Great thanks!

Answer 1

I believe you need to use window functions to attain the rank of each row based on user_id and score , and subsequently filter your results to only keep the first two values.

from pyspark.sql.window import Window
from pyspark.sql.functions import rank, col

window = Window.partitionBy(df['user_id']).orderBy(df['score'].desc())

df.select('*', rank().over(window).alias('rank')) 
  .filter(col('rank') <= 2) 
  .show() 
#+-------+---------+-----+----+
#|user_id|object_id|score|rank|
#+-------+---------+-----+----+
#| user_1| object_1|    3|   1|
#| user_1| object_2|    2|   2|
#| user_2| object_2|    6|   1|
#| user_2| object_1|    5|   2|
#+-------+---------+-----+----+

In general, the official programming guide is a good place to start learning Spark.

Data

rdd = sc.parallelize([("user_1",  "object_1",  3), 
                      ("user_1",  "object_2",  2), 
                      ("user_2",  "object_1",  5), 
                      ("user_2",  "object_2",  2), 
                      ("user_2",  "object_2",  6)])
df = sqlContext.createDataFrame(rdd, ["user_id", "object_id", "score"])

Answer 2

Top-n is more accurate if using row_number instead of rank when getting rank equality:

val n = 5
df.select(col('*'), row_number().over(window).alias('row_number')) \
  .where(col('row_number') <= n) \
  .limit(20) \
  .toPandas()

Note limit(20).toPandas() trick instead of show() for Jupyter notebooks for nicer formatting.

Answer 3

I know the question is asked for pyspark and I was looking for the similar answer in Scala ie

Retrieve top n values in each group of a DataFrame in Scala

Here is the scala version of @mtoto's answer.

import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions.rank
import org.apache.spark.sql.functions.col

val window = Window.partitionBy("user_id").orderBy('score desc)
val rankByScore = rank().over(window)
df1.select('*, rankByScore as 'rank).filter(col("rank") <= 2).show() 
# you can change the value 2 to any number you want. Here 2 represents the top 2 values

More examples can be found here .

Answer 4

Here is another solution without a window function to get the top N records from pySpark DataFrame.

# Import Libraries
from pyspark.sql.functions import col

# Sample Data
rdd = sc.parallelize([("user_1",  "object_1",  3), 
                      ("user_1",  "object_2",  2), 
                      ("user_2",  "object_1",  5), 
                      ("user_2",  "object_2",  2), 
                      ("user_2",  "object_2",  6)])
df = sqlContext.createDataFrame(rdd, ["user_id", "object_id", "score"])

# Get top n records as Row Objects
row_list = df.orderBy(col("score").desc()).head(5)

# Convert row objects to DF
sorted_df = spark.createDataFrame(row_list)

# Display DataFrame
sorted_df.show()

Output

+-------+---------+-----+
|user_id|object_id|score|
+-------+---------+-----+
| user_1| object_2|    2|
| user_2| object_2|    2|
| user_1| object_1|    3|
| user_2| object_1|    5|
| user_2| object_2|    6|
+-------+---------+-----+

If you are interested in more window functions in Spark you can refer to one of my blogs: https://medium.com/expedia-group-tech/deep-dive-into-apache-spark-window-functions-7b4e39ad3c86

Answer 5

with Python 3 and Spark 2.4

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

def get_topN(df, group_by_columns, order_by_column, n=1):
    window_group_by_columns = Window.partitionBy(group_by_columns)
    ordered_df = df.select(df.columns + [
        f.row_number().over(window_group_by_columns.orderBy(order_by_column.desc())).alias('row_rank')])
    topN_df = ordered_df.filter(f"row_rank <= {n}").drop("row_rank")
    return topN_df

top_n_df = get_topN(your_dataframe, [group_by_columns],[order_by_columns], 1) 

Answer 6

To Find Nth highest value in PYSPARK SQLquery using ROW_NUMBER() function:

SELECT * FROM (
    SELECT e.*, 
    ROW_NUMBER() OVER (ORDER BY col_name DESC) rn 
    FROM Employee e
)
WHERE rn = N

N is the nth highest value required from the column

Output:

[Stage 2:>               (0 + 1) / 1]++++++++++++++++
+-----------+
|col_name   |
+-----------+
|1183395    |
+-----------+

query will return N highest value

Answer 7

如何在同一个 spark SQL 查询中获得前 N 和后 N 条记录？

Answer 8

Adding to moto's answer, you can also consider using row_number , because rank sometimes gives same value to identical rows and this might give you different number of values per group . See below:

from pyspark.sql.window import Window
from pyspark.sql.functions import row_number, col

win = Window.partitionBy(sdf["user_id"]).orderBy(sdf["score"].desc())

sdf.select("*", row_number().over(win).alias('row_n'))\
  .filter(col('row_n') <= 2)\
  .show()

#+-------+---------+-----+-----+
#|user_id|object_id|score|row_n|
#+-------+---------+-----+-----+
#| user_1| object_1|    3|    1|
#| user_1| object_2|    2|    2|
#| user_2| object_2|    6|    1|
#| user_2| object_1|    5|    2|
#+-------+---------+-----+-----+