Python中的Spark数据框-使用UDF时执行停滞 [英] Spark dataframe in Python - execution stuck when using UDFs

问题描述

我有一个用Python编写的Spark作业，该作业正在使用DataBricks CSV阅读器从CSV文件读取数据.

I have a spark job written in Python which is reading data from the CSV files using DataBricks CSV reader.

我想通过应用udf函数将某些列从字符串转换为双精度，而udf函数实际上也在更改浮点分隔符.

I want to convert some columns from string to double by applying an udf function which actually is also changing the floating point separator.

convert_udf = F.udf(
    lambda decimal_str: _to_float(decimal_separator, decimal_str), 
    returnType=FloatType())

for name in columns:
     df = df.withColumn(name, convert_udf(df[name]))

def _to_float(decimal_separator, decimal_str):
    if isinstance(decimal_str, str) or isinstance(decimal_str, unicode):
        return (None if len(decimal_str.strip()) == 0 
               else float(decimal_str.replace(decimal_separator, '.')))
    else:
        return decimal_str

调用udf函数时，Spark作业被卡住.我试图从_to_float函数返回固定的double值，但没有成功.使用SQL上下文的udf和数据帧之间似乎有问题.

The Spark job is getting stuck when the udf function is called. I tried to return a fixed double value from the _to_float function without success. It looks like there is something wrong between the udf and data frame using SQL context.

推荐答案

长话短说，除非有必要，否则请不要使用Python UDF(通常是UDF):

Long story short don't use Python UDFs (and UDFs in general) unless it is necessary:

• 由于通过Python解释器的完整往返，因此效率低下
• 无法通过Catalyst优化
如果反复使用，
• 会创建长血统

• it is inefficient due to full round-trip through Python interpreter
• cannot be optimized by Catalyst
• creates long lineages if used iteratively

对于像这样的简单操作，只需使用内置函数即可:

For simple operations like this one just use built-in functions:

from pyspark.sql.functions import regexp_replace

decimal_separator = ","
exprs = [
    regexp_replace(c, decimal_separator, ".").cast("float").alias(c) 
    if c in columns else c 
    for c in df.columns
]

df.select(*exprs)

这篇关于Python中的Spark数据框-使用UDF时执行停滞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！