Marshmallow is a popular package used for data serialization and validation. One defines data schemas in marshmallow containing rules on how input data should be marshalled. Similar to marshmallow, pyspark also comes with its own schema definitions used to process data frames. This package enables users to utilize marshmallow schemas and its powerful data validation capabilities in pyspark applications. Such capabilities can be utilized in data-pipeline ETL jobs where data consistency and quality is of importance.
The package can be install using pip:
$ pip install marshmallow-pysparkData schemas can can define the same way as you would using marshmallow. A quick example is shown below:
from marshmallow_pyspark import Schema
from marshmallow import fields
# Create data schema.
class AlbumSchema(Schema):
title = fields.Str()
release_date = fields.Date()
# Input data frame to validate.
df = spark.createDataFrame([
{"title": "valid_1", "release_date": "2020-1-10"},
{"title": "valid_2", "release_date": "2020-1-11"},
{"title": "invalid_1", "release_date": "2020-31-11"},
{"title": "invalid_2", "release_date": "2020-1-51"},
])
# Get data frames with valid rows and error prone rows
# from input data frame by validating using the schema.
valid_df, errors_df = AlbumSchema().validate_df(df)
# Output of valid data frame
valid_df.show()
# +-------+------------+
# | title|release_date|
# +-------+------------+
# |valid_1| 2020-01-10|
# |valid_2| 2020-01-11|
# +-------+------------+
# Output of errors data frame
errors_df.show()
# +--------------------+
# | _errors|
# +--------------------+
# |{"row": {"release...|
# |{"row": {"release...|
# +--------------------+On top of marshmallow supported options, the Schema class comes with two additional initialization arguments:
-
error_column_name: name of the column to store validation errors. Default value is_errors. -
split_errors: split rows with validation errors as a separate data frame from valid rows. When set toFalsethe rows with errors are returned together with valid rows as a single data frame. The field values of all error rows are set tonull. For user convenience the original field values can be found in therowattribute of the error JSON. Default value isTrue.
An example is shown below:
from marshmallow import EXCLUDE
schema = AlbumSchema(
error_column_name="custom_errors", # Use 'custom_errors' as name for errors column
split_errors=False, # Don't split the input data frame into valid and errors
unkown=EXCLUDE # Marshmallow option to exclude fields not present in schema
)
# Input data frame to validate.
df = spark.createDataFrame([
{"title": "valid_1", "release_date": "2020-1-10", "garbage": "wdacfa"},
{"title": "valid_2", "release_date": "2020-1-11", "garbage": "5wacfa"},
{"title": "invalid_1", "release_date": "2020-31-11", "garbage": "3aqf"},
{"title": "invalid_2", "release_date": "2020-1-51", "garbage": "vda"},
])
valid_df, errors_df = schema.validate_df(df)
# Output of valid data frame. Contains rows with errors as
# the option 'split_errors' was set to False.
valid_df.show()
# +-------+------------+--------------------+
# | title|release_date| _errors|
# +-------+------------+--------------------+
# |valid_1| 2020-01-10| |
# |valid_2| 2020-01-11| |
# | | |{"row": {"release...|
# | | |{"row": {"release...|
# +-------+------------+--------------------+
# The errors data frame will be set to None
assert errors_df is None # TrueLastly, on top of passing marshmallow specific options in the schema, you can also pass them in the validate_df method.
These are options are passed to the marshmallow's load method:
schema = AlbumSchema(
error_column_name="custom_errors", # Use 'custom_errors' as name for errors column
split_errors=False, # Don't split the input data frame into valid and errors
)
valid_df, errors_df = schema.validate_df(df, unkown=EXCLUDE)Marshmallow-pyspark comes with the ability to validate one or more schema fields for duplicate values. This is achieved
by adding the field names to the UNIQUE attribute of the schema as shown:
class AlbumSchema(Schema):
# Unique valued field "title" in the schema
UNIQUE = ["title"]
title = fields.Str()
release_date = fields.Date()
# Input data frame to validate.
df = spark.createDataFrame([
{"title": "title_1", "release_date": "2020-1-10"},
{"title": "title_2", "release_date": "2020-1-11"},
{"title": "title_2", "release_date": "2020-3-11"}, # duplicate title
{"title": "title_3", "release_date": "2020-1-51"},
])
# Validate data frame
valid_df, errors_df = AlbumSchema().validate_df(df)
# List of valid rows
valid_rows = [row.asDict(recursive=True) for row in valid_df.collect()]
#
# [
# {'title': 'title_1', 'release_date': datetime.date(2020, 1, 10)},
# {'title': 'title_2', 'release_date': datetime.date(2020, 1, 11)}
# ]
#
# Rows with errors
error_rows = [row.asDict(recursive=True) for row in errors_df.collect()]
#
# [
# {'_errors': '{"row": {"release_date": "2020-3-11", "title": "title_2", "__count__title": 2}, '
# '"errors": ["duplicate row"]}'},
# {'_errors': '{"row": {"release_date": "2020-1-51", "title": "title_3", "__count__title": 1}, '
# '"errors": {"release_date": ["Not a valid date."]}}'}
# ]
#The technique to drop duplicates but keep first is discussed in this link.
In case there are multiple unique fields in the schema just add them to the UNIQUE, e.g. UNIQUE=["title", "release_date"].
You can even specify uniqueness for combination of fields by grouping them in a list:
class AlbumSchema(Schema):
# Combined values of "title" and "release_date" should be unique
UNIQUE = [["title", "release_date"]]
title = fields.Str()
release_date = fields.Date()
# Input data frame to validate.
df = spark.createDataFrame([
{"title": "title_1", "release_date": "2020-1-10"},
{"title": "title_2", "release_date": "2020-1-11"},
{"title": "title_2", "release_date": "2020-3-11"},
{"title": "title_3", "release_date": "2020-1-21"},
{"title": "title_3", "release_date": "2020-1-21"},
{"title": "title_4", "release_date": "2020-1-51"},
])
# Validate data frame
valid_df, errors_df = AlbumSchema().validate_df(df)
# List of valid rows
valid_rows = [row.asDict(recursive=True) for row in valid_df.collect()]
#
# [
# {'title': 'title_1', 'release_date': datetime.date(2020, 1, 10)},
# {'title': 'title_2', 'release_date': datetime.date(2020, 1, 11)},
# {'title': 'title_3', 'release_date': datetime.date(2020, 1, 21)}
# ]
#
# Rows with errors
error_rows = [row.asDict(recursive=True) for row in errors_df.collect()]
#
# [
# {'_errors': '{"row": {"release_date": "2020-1-21", "title": "title_3", '
# '"__count__title": 2, "__count__release_date": 2}, '
# '"errors": ["duplicate row"]}'},
# {'_errors': '{"row": {"release_date": "2020-1-51", "title": "title_4", '
# '"__count__title": 1, "__count__release_date": 1}, '
# '"errors": {"release_date": ["Not a valid date."]}}'},
# {'_errors': '{"row": {"release_date": "2020-3-11", "title": "title_2", '
# '"__count__title": 2, "__count__release_date": 1}, '
# '"errors": ["duplicate row"]}'}
# ]
#WARNING: Duplicate check requires data shuffle per unique field. Having large number of unique fields will effect
spark job performance. By default UNIQUE is set to an empty list preventing any duplicate checks.
Marshmallow comes with a variety of different fields that can be used to define schemas. Internally marshmallow-pyspark convert these fields into pyspark SQL data types. The following table lists the supported marshmallow fields and their equivalent spark SQL data types:
| Marshmallow | PySpark |
|---|---|
Raw |
user specified |
String |
StringType |
DateTime |
TimestampType |
Date |
DateType |
Boolean |
BooleanType |
Integer |
IntegerType |
Float |
FloatType |
Number |
DoubleType |
List |
ArrayType |
Dict |
MapType |
Nested |
StructType |
By default the StringType data type is used for marshmallow fields not in the above table. The spark_schema property
of your defined schema can be used to check the converted spark SQL schema:
# Gets the spark schema for the Album schema
AlbumSchema().spark_schema
# StructType(List(StructField(title,StringType,true),StructField(release_date,DateType,true),StructField(_errors,StringType,true)))Marshmallow_pyspark comes with support for an additional Raw field. The Raw field does not perform any formatting
and requires the user to specify the spark data type associated with the field. See the following example:
from marshmallow_pyspark import Schema
from marshmallow_pyspark.fields import Raw
from marshmallow import fields
from pyspark.sql.types import DateType
from datetime import date
class AlbumSchema(Schema):
title = fields.Str()
# Takes python datetime.date objects and treats them as pyspark DateType
release_date = Raw(spark_type=DateType())
# Input data frame to validate.
df = spark.createDataFrame([
{"title": "title_1", "release_date": date(2020, 1, 10)},
{"title": "title_2", "release_date": date(2020, 1, 11)},
{"title": "title_3", "release_date": date(2020, 3, 10)},
])
# Validate data frame
valid_df, errors_df = AlbumSchema().validate_df(df)
# List of valid rows
valid_rows = [row.asDict(recursive=True) for row in valid_df.collect()]
#
# [
# {'title': 'title_1', 'release_date': datetime.date(2020, 1, 10)},
# {'title': 'title_2', 'release_date': datetime.date(2020, 1, 11)},
# {'title': 'title_3', 'release_date': datetime.date(2020, 3, 10)}
# ]
#
# Rows with errors
error_rows = [row.asDict(recursive=True) for row in errors_df.collect()]
#
# []
#It is also possible to add support for custom marshmallow fields, or those missing in the above table. In order to do so,
you would need to create a converter for the custom field. The converter can be built using the ConverterABC interface:
from marshmallow_pyspark import ConverterABC
from pyspark.sql.types import StringType
class EmailConverter(ConverterABC):
"""
Converter to convert marshmallow's Email field to a pyspark
SQL data type.
"""
def convert(self, ma_field):
return StringType()The ma_field argument in the convert method is provided to handle nested fields. For an example you can checkout
NestedConverter. Now the final step would be to add the converter to the CONVERTER_MAP attribute of your schema:
from marshmallow_pyspark import Schema
from marshmallow import fields
class User(Schema):
name = fields.String(required=True)
email = fields.Email(required=True)
# Adding email converter to schema.
User.CONVERTER_MAP[fields.Email] = EmailConverter
# You can now use your schema to validate the input data frame.
valid_df, errors_df = User().validate_df(input_df)To hack marshmallow-pyspark locally run:
$ pip install -e .[dev] # to install all dependencies
$ pytest --cov-config .coveragerc --cov=./ # to get coverage report
$ pylint marshmallow_pyspark # to check code quality with PyLintOptionally you can use make to perform development tasks.
The source code is licensed under Apache License Version 2.
Pull requests always welcomed! :)