PySpark create new column with mapping from a dict
Inefficient solution with UDF (version independent):
from pyspark.sql.types import StringType
from pyspark.sql.functions import udf
def translate(mapping):
def translate_(col):
return mapping.get(col)
return udf(translate_, StringType())
df = sc.parallelize([('DS', ), ('G', ), ('INVALID', )]).toDF(['key'])
mapping = {
'A': 'S', 'B': 'S', 'C': 'S', 'DS': 'S', 'DNS': 'S',
'E': 'NS', 'F': 'NS', 'G': 'NS', 'H': 'NS'}
df.withColumn("value", translate(mapping)("key"))
with the result:
+-------+-----+
| key|value|
+-------+-----+
| DS| S|
| G| NS|
|INVALID| null|
+-------+-----+
Much more efficient (Spark >= 2.0, Spark < 3.0) is to create a MapType
literal:
from pyspark.sql.functions import col, create_map, lit
from itertools import chain
mapping_expr = create_map([lit(x) for x in chain(*mapping.items())])
df.withColumn("value", mapping_expr.getItem(col("key")))
with the same result:
+-------+-----+
| key|value|
+-------+-----+
| DS| S|
| G| NS|
|INVALID| null|
+-------+-----+
but more efficient execution plan:
== Physical Plan ==
*Project [key#15, keys: [B,DNS,DS,F,E,H,C,G,A], values: [S,S,S,NS,NS,NS,S,NS,S][key#15] AS value#53]
+- Scan ExistingRDD[key#15]
compared to UDF version:
== Physical Plan ==
*Project [key#15, pythonUDF0#61 AS value#57]
+- BatchEvalPython [translate_(key#15)], [key#15, pythonUDF0#61]
+- Scan ExistingRDD[key#15]
In Spark >= 3.0 getItem
should be replaced with __getitem__
([]
), i.e:
from pyspark.sql.functions import col, create_map, lit
from itertools import chain
mapping_expr = create_map([lit(x) for x in chain(*mapping.items())])
df.withColumn("value", mapping_expr[col("key")])
Sounds like the simplest solution would be to use the replace function: http://spark.apache.org/docs/2.4.0/api/python/pyspark.sql.html#pyspark.sql.DataFrame.replace
mapping= {
'A': '1',
'B': '2'
}
df2 = df.replace(to_replace=mapping, subset=['yourColName'])