在pyspark -pandas_udf中应用每组的功能(没有名为pyarrow的模块)

Question

我正在尝试将函数应用于pyspark中的每个数据集组.我得到的第一个错误是

Py4JError: An error occurred while calling o62.__getnewargs__. Trace:
py4j.Py4JException: Method __getnewargs__([]) does not exist

为了解决上述问题,我删除了火花功能(我有spark.range()).现在错误已经解决,但现在我得到以下内容:

File "/opt/cloudera/parcels/SPARK2-2.3.0.cloudera2-1.cdh5.13.3.p0.316101/lib/spark2/python/pyspark/serializers.py", line 276, in load_stream
    import pyarrow as pa
ImportError: No module named pyarrow 

但是当我自己尝试它时,它就可以了.

df = pd.DataFrame({"a": [1, 2, 3]})
pa.Table.from_pandas(df)
pyarrow.Table
a: int64
__index_level_0__: int64
metadata
--------
{'pandas': '{"pandas_version": "0.23.0", "index_columns": ["__index_level_0__"], "columns": [{"metadata": null, "field_name": "a", "name": "a", "numpy_type": "int64", "pandas_type": "int64"}, {"metadata": null, "field_name": "__index_level_0__", "name": null, "numpy_type": "int64", "pandas_type": "int64"}], "column_indexes": [{"metadata": null, "field_name": null, "name": null, "numpy_type": "object", "pandas_type": "bytes"}]}'}

失败的例子 - 从这里开始

import pyspark.sql.functions as F
import pandas as pd

cols = ['id', 'val']
vals = [('A', 5), ('A', 3), ('A', 7), ('B', 12), ('B', 15), ('C', 3)]
d1 = sqlContext.createDataFrame(vals, cols)

>>> @F.pandas_udf(d1.schema, F.PandasUDFType.GROUPED_MAP)
... def subtract_mean(pdf):
...     return pdf.assign(v=pdf.v - pdf.v.mean())
...
>>> try1 = d1.groupby('id').apply(subtract_mean)
>>> try1.show()

如果我尝试转换为pandas数据帧,它不起作用,因为它没有该schema属性.

我在这里错过了什么？

Answer 1

pyarrow 必须出现在每个工作节点的路径上.

• 如果在单个节点上运行此代码,请确保PYSPARK_PYTHON(和可选的PYTHONPATH)与用于测试pyarrow代码的解释器相同.
• 如果您使用群集,请确保pyarrow在每个节点上安装了群集,此外还有上面提到的点.

此外,请确保安装的pyarrow版本大于或等于支持的最小版本(今天为0.8) - 尽管这会导致不同的异常.