I have the following table
id | country_mapping
--------------------
1 | {"GBR/bla": 1,
"USA/bla": 2}
And I want to create a columns that contains the following
id | source_countries
--------------------
1 | ["GBR", "USA"]
And I need this to be done via a pandas udf. I created the following
import pyspark.sql.functions as F
@F.pandas_udf("string")
def func(s):
return s.apply(lambda x: [y.split("/")[0] for y in x])
I thought this would work, because if I run this code in pure pandas it gives what i need.
import pandas as pd
s = pd.Series([["GBR/1", "USA/2"], ["ITA/1", "FRA/2"]])
s.apply(lambda x: [y.split("/")[0] for y in x])
gives
Out[1]: 0 [GBR, USA]
1 [ITA, FRA]
dtype: object
But when I run
df.withColumn('source_countries',
func(F.map_keys(F.col("country_mapping")))).collect()
It fails with the following error when i run the below:
PythonException: An exception was thrown from a UDF: 'pyarrow.lib.ArrowTypeError: Expected bytes, got a 'list' object'
I'm confused as of why - and how to fix my pandas udf.
CodePudding user response:
Instead of pandas_udf
, you can just use udf
in similar way
from pyspark.sql import functions as F
from pyspark.sql import types as T
def func(v):
return [x.split('/')[0] for x in v]
(df
.withColumn('source_countries', F.udf(func, T.ArrayType(T.StringType()))(F.map_keys(F.col('country_mapping'))))
.show(10, False)
)
# --- ---------------------------- ----------------
# |id |country_mapping |source_countries|
# --- ---------------------------- ----------------
# |1 |{USA/bla -> 2, GBR/bla -> 1}|[USA, GBR] |
# --- ---------------------------- ----------------
CodePudding user response:
The answer to this question is that
Currently, all Spark SQL data types are supported by Arrow-based conversion except MapType, ArrayType of TimestampType, and nested StructType.