pyspark: count distinct over a window

Question

I just tried doing a countDistinct over a window and got this error:

AnalysisException: u'Distinct window functions are not supported: count(distinct color#1926)

Is there a way to do a distinct count over a window in pyspark?

Here's some example code:

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

#function to calculate number of seconds from number of days
days = lambda i: i * 86400

df = spark.createDataFrame([(17, "2017-03-10T15:27:18+00:00", "orange"),
                    (13, "2017-03-15T12:27:18+00:00", "red"),
                    (25, "2017-03-18T11:27:18+00:00", "red")],
                    ["dollars", "timestampGMT", "color"])
                    
df = df.withColumn('timestampGMT', df.timestampGMT.cast('timestamp'))

#create window by casting timestamp to long (number of seconds)
w = (Window.orderBy(F.col("timestampGMT").cast('long')).rangeBetween(-days(7), 0))

df = df.withColumn('distinct_color_count_over_the_last_week', F.countDistinct("color").over(w))

df.show()

This is the output I'd like to see:

+-------+--------------------+------+---------------------------------------+
|dollars|        timestampGMT| color|distinct_color_count_over_the_last_week|
+-------+--------------------+------+---------------------------------------+
|     17|2017-03-10 15:27:...|orange|                                      1|
|     13|2017-03-15 12:27:...|   red|                                      2|
|     25|2017-03-18 11:27:...|   red|                                      1|
+-------+--------------------+------+---------------------------------------+

Solution 1

EDIT:

As noleto mentions in his answer below, there is now an approx_count_distinct function since pyspark 2.1 that works over a window.

---

Original Answer

I figured out that I can use a combination of the collect_set and size functions to mimic the functionality of countDistinct over a window:

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

#function to calculate number of seconds from number of days
days = lambda i: i * 86400

#create some test data
df = spark.createDataFrame([(17, "2017-03-10T15:27:18+00:00", "orange"),
                    (13, "2017-03-15T12:27:18+00:00", "red"),
                    (25, "2017-03-18T11:27:18+00:00", "red")],
                    ["dollars", "timestampGMT", "color"])

#convert string timestamp to timestamp type             
df = df.withColumn('timestampGMT', df.timestampGMT.cast('timestamp'))

#create window by casting timestamp to long (number of seconds)
w = (Window.orderBy(F.col("timestampGMT").cast('long')).rangeBetween(-days(7), 0))

#use collect_set and size functions to perform countDistinct over a window
df = df.withColumn('distinct_color_count_over_the_last_week', F.size(F.collect_set("color").over(w)))

df.show()

This results in the distinct count of color over the previous week of records:

+-------+--------------------+------+---------------------------------------+
|dollars|        timestampGMT| color|distinct_color_count_over_the_last_week|
+-------+--------------------+------+---------------------------------------+
|     17|2017-03-10 15:27:...|orange|                                      1|
|     13|2017-03-15 12:27:...|   red|                                      2|
|     25|2017-03-18 11:27:...|   red|                                      1|
+-------+--------------------+------+---------------------------------------+

Solution 2

@Bob Swain's answer is nice and works! Since then, Spark version 2.1, Spark offers an equivalent to countDistinct function, approx_count_distinct which is more efficient to use and most importantly, supports counting distinct over a window.

Here goes the code to drop in replacement:

#approx_count_distinct supports a window
df = df.withColumn('distinct_color_count_over_the_last_week', F.approx_count_distinct("color").over(w))

For columns with small cardinalities, result is supposed to be the same as "countDistinct". When dataset grows a lot, you should consider adjusting the parameter rsd – maximum estimation error allowed, which allows you to tune the trade-off precision/performance.