Tuve el mismo problema y sampleSize soluciona parcialmente este problema, pero no lo resuelve si tiene muchos datos.
Aquí está la solución de cómo puede solucionar esto. Use este enfoque junto con un aumento de sampleSize (en mi caso es 100000):
def fix_schema(schema: StructType) -> StructType:
"""Fix spark schema due to inconsistent MongoDB schema collection.
It fixes such issues like:
Cannot cast STRING into a NullType
Cannot cast STRING into a StructType
:param schema: a source schema taken from a Spark DataFrame to be fixed
"""
if isinstance(schema, StructType):
return StructType([fix_schema(field) for field in schema.fields])
if isinstance(schema, ArrayType):
return ArrayType(fix_schema(schema.elementType))
if isinstance(schema, StructField) and is_struct_oid_obj(schema):
return StructField(name=schema.name, dataType=StringType(), nullable=schema.nullable)
elif isinstance(schema, StructField):
return StructField(schema.name, fix_schema(schema.dataType), schema.nullable)
if isinstance(schema, NullType):
return StringType()
return schema
def is_struct_oid_obj(struct_field: StructField) -> bool:
"""
Checks that our schema has StructType field with single oid name inside
:param struct_field: a StructField from Spark schema
:return bool
"""
return (isinstance(struct_field.dataType, StructType)
and len(struct_field.dataType.fields) == 1
and struct_field.dataType.fields[0].name == "oid")
