如何计算 PySpark 数据框中每个键的百分位数？

Question

我有一个 PySpark 数据框，由三列 x、y、z 组成。

X 在此数据框中可能有多行。如何分别计算 x 中每个键的百分位数？

+------+---------+------+
|  Name|     Role|Salary|
+------+---------+------+
|   bob|Developer|125000|
|  mark|Developer|108000|
|  carl|   Tester| 70000|
|  carl|Developer|185000|
|  carl|   Tester| 65000|
| roman|   Tester| 82000|
| simon|Developer| 98000|
|  eric|Developer|144000|
|carlos|   Tester| 75000|
| henry|Developer|110000|
+------+---------+------+

需要的输出：

+------+---------+------+---------+
|  Name|     Role|Salary|      50%|
+------+---------+------+---------+
|   bob|Developer|125000|117500.0 |
|  mark|Developer|108000|117500.0 |
|  carl|   Tester| 70000|72500.0  |
|  carl|Developer|185000|117500.0 |
|  carl|   Tester| 65000|72500.0  |
| roman|   Tester| 82000|72500.0  |
| simon|Developer| 98000|117500.0 |
|  eric|Developer|144000|117500.0 |
|carlos|   Tester| 75000|72500.0  |
| henry|Developer|110000|117500.0 |
+------+---------+------+---------+

Answer 1

尝试

groupby

+

F.expr

：

import pyspark.sql.functions as F

df1 = df.groupby('Role').agg(F.expr('percentile(Salary, array(0.25))')[0].alias('%25'),
                             F.expr('percentile(Salary, array(0.50))')[0].alias('%50'),
                             F.expr('percentile(Salary, array(0.75))')[0].alias('%75'))
df1.show()

输出：

+---------+--------+--------+--------+
|     Role|     %25|     %50|     %75|
+---------+--------+--------+--------+
|   Tester| 68750.0| 72500.0| 76750.0|
|Developer|108500.0|117500.0|139250.0|
+---------+--------+--------+--------+

现在您可以将

df1

与原始数据框连接：

df.join(df1, on='Role', how='left').show()

输出：

+---------+------+------+--------+--------+--------+
|     Role|  Name|Salary|     %25|     %50|     %75|
+---------+------+------+--------+--------+--------+
|   Tester|  carl| 70000| 68750.0| 72500.0| 76750.0|
|   Tester|  carl| 65000| 68750.0| 72500.0| 76750.0|
|   Tester| roman| 82000| 68750.0| 72500.0| 76750.0|
|   Tester|carlos| 75000| 68750.0| 72500.0| 76750.0|
|Developer|   bob|125000|108500.0|117500.0|139250.0|
|Developer|  mark|108000|108500.0|117500.0|139250.0|
|Developer|  carl|185000|108500.0|117500.0|139250.0|
|Developer| simon| 98000|108500.0|117500.0|139250.0|
|Developer|  eric|144000|108500.0|117500.0|139250.0|
|Developer| henry|110000|108500.0|117500.0|139250.0|
+---------+------+------+--------+--------+--------+

Answer 2

array

其实没有必要:

• 星火3.5+

  F.percentile('Salary', 0.5)
  

• 星火2.1+

  F.expr('percentile(Salary, 0.5)')
  

与窗口函数一起完成这项工作：

df = df.withColumn('50%', F.percentile('Salary', 0.5).over(W.partitionBy('Role')))

df.show()
# +------+---------+------+--------+
# |  Name|     Role|Salary|     50%|
# +------+---------+------+--------+
# |   bob|Developer|125000|117500.0|
# |  mark|Developer|108000|117500.0|
# |  carl|Developer|185000|117500.0|
# | simon|Developer| 98000|117500.0|
# |  eric|Developer|144000|117500.0|
# | henry|Developer|110000|117500.0|
# |  carl|   Tester| 70000| 72500.0|
# |  carl|   Tester| 65000| 72500.0|
# | roman|   Tester| 82000| 72500.0|
# |carlos|   Tester| 75000| 72500.0|
# +------+---------+------+--------+

Answer 3

您可以尝试spark中提供的

approxQuantile

功能。

https://spark.apache.org/docs/latest/api/python/reference/pyspark.sql/api/pyspark.sql.DataFrame.approxQuantile.html