9

I created a PySpark dataframe using the following code

testlist = [
             {"category":"A","name":"A1"}, 
             {"category":"A","name":"A2"}, 
             {"category":"B","name":"B1"},
             {"category":"B","name":"B2"}
]

spark_df = spark.createDataFrame(testlist)

Result:

category    name
A           A1
A           A2
B           B1
B           B2

I want to make it appear as follows:

category    name
A           A1, A2
B           B1, B2

I tried the following code which does not work

spark_df.groupby('category').agg('name', lambda x:x + ', ')

Can anyone help identify what I am doing wrong and the best way to make this happen ?

pault
  • 41,343
  • 15
  • 107
  • 149
Harry
  • 93
  • 1
  • 1
  • 7

3 Answers3

23

One option is to use pyspark.sql.functions.collect_list() as the aggregate function. 

from pyspark.sql.functions import collect_list
grouped_df = spark_df.groupby('category').agg(collect_list('name').alias("name"))

This will collect the values for name into a list and the resultant output will look like:

grouped_df.show()
#+---------+---------+
#|category |name     |
#+---------+---------+
#|A        |[A1, A2] |
#|B        |[B1, B2] |
#+---------+---------+

Update 2019-06-10: If you wanted your output as a concatenated string, you can use pyspark.sql.functions.concat_ws to concatenate the values of the collected list, which will be better than using a udf:

from pyspark.sql.functions import concat_ws

grouped_df.withColumn("name", concat_ws(", ", "name")).show()
#+---------+-------+
#|category |name   |
#+---------+-------+
#|A        |A1, A2 |
#|B        |B1, B2 |
#+---------+-------+

Original Answer: If you wanted your output as a concatenated string, you'd have to can use a udf. For example, you can first do the groupBy() as above and the apply a udf to join the collected list:

from pyspark.sql.functions import udf
concat_list = udf(lambda lst: ", ".join(lst), StringType())

grouped_df.withColumn("name", concat_list("name")).show()
#+---------+-------+
#|category |name   |
#+---------+-------+
#|A        |A1, A2 |
#|B        |B1, B2 |
#+---------+-------+
pault
  • 41,343
  • 15
  • 107
  • 149
  • Hi pault, thanks for your help. I am wondering how you got the answer. I tried exactly the same and I get the error 'NoneType does not have attribute withColumn'. Please let me know if I should correct something. Thanks ! – Harry Mar 24 '18 at 23:26
  • Are you sure you didn't set `grouped_df` equal to that output of `.show()`? That error means the df for which you are calling `withColumn()` is actually `None` – pault Mar 24 '18 at 23:33
3

UNIQUE values

If you want unique values then use collect_set instead of collect_list

from pyspark.sql.functions import collect_set
grouped_df = sdf.groupby('category').agg(collect_set('name').alias("unique_name"))
sdf.show(5)
s510
  • 2,271
  • 11
  • 18
1

Another option is this

>>> df.rdd.reduceByKey(lambda x,y: x+','+y).toDF().show()
+---+-----+
| _1|   _2|
+---+-----+
|  A|A1,A2|
|  B|B1,B2|
+---+-----+
Bala
  • 11,068
  • 19
  • 67
  • 120
  • Hi Bala, Thanks for your help. I tried your script on my data which has more than instances of each key. It gives me the error: ValueError: too many values to unpack (expected 2) – Harry Mar 24 '18 at 22:58
  • `reduceByKey` works with more than one instance of key. Can you give us sample data having exact pattern you are working with. – Bala Mar 25 '18 at 07:17