pyspark:在窗口上统计

Question

我刚尝试在窗口上执行countDistinct并出现此错误:

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

有没有办法在pyspark中对窗口进行明确计数？

这是一些示例代码:

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()

这是我想看到的输出:

+-------+--------------------+------+---------------------------------------+
|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|
+-------+--------------------+------+---------------------------------------+

Answer 1

我发现我可以使用collect_set和size函数的组合来模拟窗口上countDistinct的功能:

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()

这导致前一周记录的颜色数量明显不同:

+-------+--------------------+------+---------------------------------------+
|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|
+-------+--------------------+------+---------------------------------------+

Answer 2

@Bob Swain的回答很好，而且行得通！从那时起，Spark版本2.1开始，Spark提供了等效的countDistinct功能，approx_count_distinct其使用效率更高，最重要的是，它支持在窗口上进行计数。

下面是替换代码：

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

对于基数较小的列，结果应与“ countDistinct”相同。当数据集增长很多时，您应该考虑调整参数rsd-允许的最大估计误差，这使您可以调整权衡的精度/性能。