使用 UDF 时忽略条件 [英] Using UDF ignores condition in when

问题描述

假设您有以下 pyspark DataFrame:

Suppose you had the following pyspark DataFrame:

data= [('foo',), ('123',), (None,), ('bar',)]
df = sqlCtx.createDataFrame(data, ["col"])
df.show()
#+----+
#| col|
#+----+
#| foo|
#| 123|
#|null|
#| bar|
#+----+

接下来的两个代码块应该做同样的事情——也就是说，如果列不是null，则返回该列的大写.但是，第二种方法(使用 udf)会产生错误.

The next two code blocks should do the same thing- that is, return the uppercase of the column if it is not null. However, the second method (using a udf) produces an error.

方法一:使用pyspark.sql.functions.upper()

import pyspark.sql.functions as f
df.withColumn(
    'upper',
    f.when(
        f.isnull(f.col('col')),
        f.col('col')
    ).otherwise(f.upper(f.col('col')))
).show()
#+----+-----+
#| col|upper|
#+----+-----+
#| foo|  FOO|
#| 123|  123|
#|null| null|
#| bar|  BAR|
#+----+-----+

方法 2:在 udf

df.withColumn(
    'upper',
    f.when(
        f.isnull(f.col('col')),
        f.col('col')
    ).otherwise(f.udf(lambda x: x.upper(), StringType())(f.col('col')))
).show()

这给了我 AttributeError: 'NoneType' object has no attribute 'upper'.为什么 f.isnull() 在对 when 的调用中检查似乎被忽略了?

This gives me AttributeError: 'NoneType' object has no attribute 'upper'. Why is the f.isnull() check in the call to when seemingly being ignored?

我知道我可以将我的 udf 更改为 f.udf(lambda x: x.upper() if x else x, StringType()) 以避免这种情况错误，但我想了解它为什么会发生.

I know that I can change my udf to f.udf(lambda x: x.upper() if x else x, StringType()) to avoid this error, but I'd like to understand why it's happening.

完整追溯:

Py4JJavaErrorTraceback (most recent call last)
<ipython-input-38-cbf0ffe73538> in <module>()
      4         f.isnull(f.col('col')),
      5         f.col('col')
----> 6     ).otherwise(f.udf(lambda x: x.upper(), StringType())(f.col('col')))
      7 ).show()

/opt/SPARK2/lib/spark2/python/pyspark/sql/dataframe.py in show(self, n, truncate)
    316         """
    317         if isinstance(truncate, bool) and truncate:
--> 318             print(self._jdf.showString(n, 20))
    319         else:
    320             print(self._jdf.showString(n, int(truncate)))

/opt/SPARK2/lib/spark2/python/lib/py4j-0.10.4-src.zip/py4j/java_gateway.py in __call__(self, *args)
   1131         answer = self.gateway_client.send_command(command)
   1132         return_value = get_return_value(
-> 1133             answer, self.gateway_client, self.target_id, self.name)
   1134 
   1135         for temp_arg in temp_args:

/opt/SPARK2/lib/spark2/python/pyspark/sql/utils.py in deco(*a, **kw)
     61     def deco(*a, **kw):
     62         try:
---> 63             return f(*a, **kw)
     64         except py4j.protocol.Py4JJavaError as e:
     65             s = e.java_exception.toString()

/opt/SPARK2/lib/spark2/python/lib/py4j-0.10.4-src.zip/py4j/protocol.py in get_return_value(answer, gateway_client, target_id, name)
    317                 raise Py4JJavaError(
    318                     "An error occurred while calling {0}{1}{2}.\n".
--> 319                     format(target_id, ".", name), value)
    320             else:
    321                 raise Py4JError(

Py4JJavaError: An error occurred while calling o642.showString.
: org.apache.spark.SparkException: Job aborted due to stage failure: Task 51 in stage 77.0 failed 4 times, most recent failure: Lost task 51.3 in stage 77.0 (TID 5101, someserver.prod.somecompany.net, executor 99): org.apache.spark.api.python.PythonException: Traceback (most recent call last):
  File "/opt/SPARK2/lib/spark2/python/lib/pyspark.zip/pyspark/worker.py", line 174, in main
    process()
  File "/opt/SPARK2/lib/spark2/python/lib/pyspark.zip/pyspark/worker.py", line 169, in process
    serializer.dump_stream(func(split_index, iterator), outfile)
  File "/opt/SPARK2/lib/spark2/python/lib/pyspark.zip/pyspark/worker.py", line 106, in <lambda>
    func = lambda _, it: map(mapper, it)
  File "/opt/SPARK2/lib/spark2/python/lib/pyspark.zip/pyspark/worker.py", line 92, in <lambda>
    mapper = lambda a: udf(*a)
  File "/opt/SPARK2/lib/spark2/python/lib/pyspark.zip/pyspark/worker.py", line 70, in <lambda>
    return lambda *a: f(*a)
  File "<ipython-input-38-cbf0ffe73538>", line 6, in <lambda>
AttributeError: 'NoneType' object has no attribute 'upper'

推荐答案

您必须记住，Spark SQL(与 RDD 不同)不是所见即所得.优化器/规划器可以自由地以任意顺序安排操作，甚至多次重复阶段.

You have to remember that Spark SQL (unlike RDD) is not what-you-see-is-what-you-get. Optimizer / planner is free to schedule operations in the arbitrary order or even repeat stages multiple times.

Python udfs 不是在 Row 基础上应用的，而是使用批处理模式.when 没有那么多被忽略，而是没有用于优化执行计划:

Python udfs are not applied on a Row basis, but using batch mode. when is not so much ignored, but not used to optimize execution plan:

== Physical Plan ==
*Project [col#0, CASE WHEN isnull(col#0) THEN col#0 ELSE pythonUDF0#21 END AS upper#17]
+- BatchEvalPython [<lambda>(col#0)], [col#0, pythonUDF0#21]
   +- Scan ExistingRDD[col#0]

因此与 udf 一起使用的函数必须对 None 输入具有鲁棒性，例如:

Therefore function used with udf has to be robust to None inputs, for example:

df.withColumn(
    'upper',
    f.udf(
        lambda x: x.upper() if x is not None else None, 
        StringType()
    )(f.col('col'))
).show()

这篇关于使用 UDF 时忽略条件的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！