Retrieve top n in each group of a DataFrame in pyspark

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.

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"])