Upgrading to Newer Releases

This section documents migration paths to new releases.

Upgrading to 3.3

In 3.3, fields.Nested may take a callable that returns a schema instance. Use this to resolve order-of-declaration issues when schemas nest each other.

from marshmallow import Schema, fields


# <3.3
class AlbumSchema(Schema):
    title = fields.Str()
    artist = fields.Nested("ArtistSchema", only=("name",))


class ArtistSchema(Schema):
    name = fields.Str()
    albums = fields.List(fields.Nested(AlbumSchema))


# >=3.3
class AlbumSchema(Schema):
    title = fields.Str()
    artist = fields.Nested(lambda: ArtistSchema(only=("name",)))


class ArtistSchema(Schema):
    name = fields.Str()
    albums = fields.List(fields.Nested(AlbumSchema))

A callable should also be used when nesting a schema within itself. Passing "self" is deprecated.

from marshmallow import Schema, fields


# <3.3
class PersonSchema(Schema):
    partner = fields.Nested("self", exclude=("partner",))
    friends = fields.List(fields.Nested("self"))


# >=3.3
class PersonSchema(Schema):
    partner = fields.Nested(lambda: PersonSchema(exclude=("partner")))
    friends = fields.List(fields.Nested(lambda: PersonSchema()))

Upgrading to 3.0

Python compatibility

The marshmallow 3.x series supports Python >= 3.8.

Schemas are always strict

Two major changes were made to (de)serialization behavior:

  • The strict parameter was removed. Schemas are always strict.

  • Schema().load and Schema().dump don’t return a (data, errors) tuple any more. Only data is returned.

If invalid data are passed, a ValidationError is raised. The dictionary of validation errors is accessible from the ValidationError.messages attribute, along with the valid data from the ValidationError.valid_data attribute.

from marshmallow import ValidationError

# 2.x
schema = UserSchema()
data, errors = schema.load({"name": "Monty", "email": "monty@python.org"})
# OR
schema = UserSchema(strict=True)
try:
    data, _ = schema.load({"name": "Monty", "email": "monty@python.org"})
except ValidationError as err:
    errors = err.messages
    valid_data = err.valid_data

# 3.x
schema = UserSchema()
# There is only one right way
try:
    data = schema.load({"name": "Monty", "email": "monty@python.org"})
except ValidationError as err:
    errors = err.messages
    valid_data = err.valid_data

Schema.validate() always returns a dictionary of validation errors (same as 2.x with strict=False).

schema.validate({"email": "invalid"})
# {'email': ['Not a valid email address.']}

Setting the strict option on class Meta has no effect on Schema behavior. Passing strict=True or strict=False to the Schema constructor will raise a TypeError.

# 3.x
UserSchema(strict=True)
# TypeError: __init__() got an unexpected keyword argument 'strict'

See also

See GitHub issues #377 and #598 for the discussions on this change.

Decorated methods and handle_error receive many and partial

Methods decorated with pre_load, post_load, pre_dump, post_dump, and validates_schema receive many as a keyword argument. In addition, pre_load, post_load, and validates_schema receive partial. To account for these additional arguments, add **kwargs to your methods.

# 2.x
class UserSchema(Schema):
    name = fields.Str()
    slug = fields.Str()

    @pre_load
    def slugify_name(self, in_data):
        in_data["slug"] = in_data["slug"].lower().strip().replace(" ", "-")
        return in_data


# 3.x
class UserSchema(Schema):
    name = fields.Str()
    slug = fields.Str()

    @pre_load
    def slugify_name(self, in_data, **kwargs):
        in_data["slug"] = in_data["slug"].lower().strip().replace(" ", "-")
        return in_data

Schema.handle_error also receives many and partial as keyword arguments.

# 2.x
class UserSchema(Schema):
    def handle_error(self, exc, data):
        raise AppError("An error occurred with input: {0}".format(data))


# 3.x
class UserSchema(Schema):
    def handle_error(self, exc, data, **kwargs):
        raise AppError("An error occurred with input: {0}".format(data))

Validation does not occur on serialization

Schema.dump() will no longer validate and collect error messages. You must validate your data before serializing it.

from marshmallow import Schema, fields, ValidationError

invalid_data = dict(created_at="invalid")


class WidgetSchema(Schema):
    created_at = fields.DateTime()


# 2.x
WidgetSchema(strict=True).dump(invalid_data)
# marshmallow.exceptions.ValidationError: {'created_at': ['"invalid" cannot be formatted as a datetime.']}

# 3.x
WidgetSchema().dump(invalid_data)
# AttributeError: 'str' object has no attribute 'isoformat'

# Instead, validate before dumping
schema = WidgetSchema()
try:
    widget = schema.load(invalid_data)
except ValidationError:
    print("handle errors...")
else:
    dumped = schema.dump(widget)

Deserializing invalid types raises a ValidationError

Numbers, booleans, strings, and None are considered invalid input to Schema.load.

# 2.x
# Passes silently
schema.load(None)
schema.load(False)
schema.load("pass")

# 3.x
# marshmallow.exceptions.ValidationError: {'_schema': ['Invalid input type.']}
schema.load(None)
schema.load(False)
schema.load("nope")

When many=True, non-collection types are also considered invalid.

# 2.x
# Passes silently
schema.load(None, many=True)
schema.load({}, many=True)
schema.load("pass", many=True)

# 3.x
# marshmallow.exceptions.ValidationError: {'_schema': ['Invalid input type.']}
schema.load(None, many=True)
schema.load({}, many=True)
schema.load("invalid", many=True)

ValidationError.fields is removed

ValidationError no longer stores a list of Field instances associated with the validation errors.

If you need field instances associated with an error, you can access them from schema.fields.

from marshmallow import Schema, fields, ValidationError


class MySchema(Schema):
    foo = fields.Int()


schema = MySchema()

try:
    schema.load({"foo": "invalid"})
except ValidationError as error:
    field = schema.fields["foo"]
    # ...

ValidationError expects a single field name

ValidationError no longer accepts a list of field names. It expects a single field name. If none is passed, the error refers to the schema.

To return an error for several fields at once, a dict must be used.

from marshmallow import Schema, fields, validates_schema, ValidationError


class NumberSchema(Schema):
    field_a = fields.Integer()
    field_b = fields.Integer()

    # 2.x
    @validates_schema
    def validate_numbers(self, data):
        if data["field_b"] >= data["field_a"]:
            raise ValidationError(
                "field_a must be greater than field_b", ["field_a", "field_b"]
            )

    # 3.x
    @validates_schema
    def validate_numbers(self, data):
        if data["field_b"] >= data["field_a"]:
            raise ValidationError(
                {
                    "field_a": ["field_a must be greater than field_b"],
                    "field_b": ["field_a must be greater than field_b"],
                }
            )

ValidationError error messages are deep-merged

When multiple ValidationError are raised, the error structures are merged in the final ValidationError raised at the end of the process.

When reporting error messages as dict, the keys should refer to subitems of the item the message refers to, and the values should be error messages.

See the “Schema-level Validation” section of Extending Schemas page for an example.

Schemas raise ValidationError when deserializing data with unknown keys

Marshmallow 3.x schemas can deal with unknown keys in three different ways, configurable with the unknown option:

  • EXCLUDE: drop those keys (same as marshmallow 2)

  • INCLUDE: pass those keys/values as is, with no validation performed

  • RAISE (default): raise a ValidationError

The unknown option can be passed as a Meta option, on Schema instantiation, or at load time.

from marshmallow import Schema, fields, EXCLUDE, INCLUDE, RAISE


class MySchema(Schema):
    foo = fields.Int()

    class Meta:
        # Pass EXCLUDE as Meta option to keep marshmallow 2 behavior
        unknown = EXCLUDE


MySchema().load({"foo": 42, "bar": "whatever"})  # => ['foo': 42]

#  Value passed on instantiation overrides Meta option
schema = MySchema(unknown=INCLUDE)
schema.load({"foo": 42, "bar": "whatever"})  # => ['foo': 42, 'bar': 'whatever']

#  Value passed on load overrides instance attribute
schema.load({"foo": 42, "bar": "whatever"}, unknown=RAISE)  # => ValidationError

Overriding get_attribute

If your Schema overrides get_attribute, you will need to update the method’s signature. The positions of the attr and obj arguments were switched for consistency with Python builtins, e.g. getattr.

from marshmallow import Schema


# 2.x
class MySchema(Schema):
    def get_attribute(self, attr, obj, default):
        return getattr(obj, attr, default)


# 3.x
class MySchema(Schema):
    def get_attribute(self, obj, attr, default):
        return getattr(obj, attr, default)

pass_original=True passes individual items when many=True

When pass_original=True is passed to validates_schema, post_load, or post_dump, the original_data argument will be a single item corresponding to the (de)serialized datum.

from marshmallow import Schema, fields, post_load, EXCLUDE


class ShoeSchema(Schema):
    size = fields.Int()

    class Meta:
        unknown = EXCLUDE

    @post_load(pass_original=True)
    def post_load(self, data, original_data, **kwargs):
        # original_data has 'width' but
        # data does not because it's not
        # in the schema
        assert "width" in original_data
        assert "width" not in data
        return data


input_data = [{"size": 10, "width": "M"}, {"size": 6, "width": "W"}]

print(ShoeSchema(many=True).load(input_data))
# [{'size': 10}, {'size': 6}]

utils.get_func_args no longer returns bound arguments

The utils.get_func_args function will no longer return bound arguments, e.g. 'self'.

from marshmallow.utils import get_func_args


class MyCallable:
    def __call__(self, foo, bar):
        return 42


callable_obj = MyCallable()

# 2.x
get_func_args(callable_obj)  # => ['self', 'foo', 'bar']

# 3.x
get_func_args(callable_obj)  # => ['foo', 'bar']

Handling AttributeError in Method and Function fields

The Method and Function fields no longer swallow AttributeErrors. Therefore, your methods and functions are responsible for handling inputs such as None.

from marshmallow import Schema, fields, missing


# 2.x
class ShapeSchema(Schema):
    area = fields.Method("get_area")

    def get_area(self, obj):
        return obj.height * obj.length


schema = ShapeSchema()
# In 2.x, the following would pass without errors
# In 3.x, and AttributeError would be raised
result = schema.dump(None)
result  # => {}


# 3.x
class ShapeSchema(Schema):
    area = fields.Method("get_area")

    def get_area(self, obj):
        if obj is None:
            # 'area' will not appear in serialized output
            return missing
        return obj.height * obj.length


schema = ShapeSchema()
result = schema.dump(None)
result  # => {}

Adding additional data to serialized output

Use a post_dump to add additional data on serialization. The extra argument on Schema was removed.

from marshmallow import Schema, fields, post_dump


# 2.x
class MySchema(Schema):
    x = fields.Int()
    y = fields.Int()


schema = MySchema(extra={"z": 123})
schema.dump({"x": 1, "y": 2})
# => {'z': 123, 'y': 2, 'x': 1}


# 3.x
class MySchema(Schema):
    x = fields.Int()
    y = fields.Int()

    @post_dump
    def add_z(self, output):
        output["z"] = 123
        return output


schema = MySchema()
schema.dump({"x": 1, "y": 2})
# => {'z': 123, 'y': 2, 'x': 1}

Schema-level validators are skipped when field validation fails

By default, schema validator methods decorated by validates_schema won’t execute if any of the field validators fails (including required=True validation).

from marshmallow import Schema, fields, validates_schema, ValidationError


class MySchema(Schema):
    x = fields.Int(required=True)
    y = fields.Int(required=True)

    @validates_schema
    def validate_schema(self, data):
        if data["x"] <= data["y"]:
            raise ValidationError("x must be greater than y")


schema = MySchema()

# 2.x
# A KeyError is raised in validate_schema
schema.load({"x": 2})

# 3.x
# marshmallow.exceptions.ValidationError: {'y': ['Missing data for required field.']}
# validate_schema is not run
schema.load({"x": 2})

If you want a schema validator to run even if a field validator fails, pass skip_on_field_errors=False. Make sure your code handles cases where fields are missing from the deserialized data (due to validation errors).

from marshmallow import Schema, fields, validates_schema, ValidationError


class MySchema(Schema):
    x = fields.Int(required=True)
    y = fields.Int(required=True)

    @validates_schema(skip_on_field_errors=False)
    def validate_schema(self, data):
        if "x" in data and "y" in data:
            if data["x"] <= data["y"]:
                raise ValidationError("x must be greater than y")


schema = MySchema()
schema.load({"x": 2})
# marshmallow.exceptions.ValidationError: {'y': ['Missing data for required field.']}

SchemaOpts constructor receives ordered argument

Subclasses of SchemaOpts receive an additional argument, ordered, which is True if the ordered option is set to True on a Schema or one of its parent classes.

from marshmallow import SchemaOpts


# 2.x
class CustomOpts(SchemaOpts):
    def __init__(self, meta):
        super().__init__(meta)
        self.custom_option = getattr(meta, "meta", False)


# 3.x
class CustomOpts(SchemaOpts):
    def __init__(self, meta, ordered=False):
        super().__init__(meta, ordered)
        self.custom_option = getattr(meta, "meta", False)

ContainsOnly accepts empty and duplicate values

validate.ContainsOnly now accepts duplicate values in the input value.

from marshmallow import validate

validator = validate.ContainsOnly(["red", "blue"])

# in 2.x the following raises a ValidationError
# in 3.x, no error is raised
validator(["red", "red", "blue"])

If you don’t want to accept duplicates, use a custom validator, like the following.

from marshmallow import ValidationError
from marshmallow.validate import ContainsOnly


class ContainsOnlyNoDuplicates(ContainsOnly):
    def __call__(self, value):
        ret = super(ContainsOnlyNoDuplicates, self).__call__(value)
        if len(set(value)) != len(value):
            raise ValidationError("Duplicate values not allowed")
        return ret

Note

If you need to handle unhashable types, you can use the implementation of ContainsOnly from marshmallow 2.x.

validate.ContainsOnly also accepts empty values as valid input.

from marshmallow import validate

validator = validate.ContainsOnly(["red", "blue"])

# in 2.x the following raises a ValidationError
# in 3.x, no error is raised
validator([])

To validate against empty inputs, use validate.Length(min=1).

json_module option is renamed to render_module

The json_module class Meta option is deprecated in favor of render_module.

import ujson


# 2.x
class MySchema(Schema):
    class Meta:
        json_module = ujson


# 3.x
class MySchema(Schema):
    class Meta:
        render_module = ujson

missing and default Field parameters are passed in deserialized form

# 2.x
class UserSchema(Schema):
    id = fields.UUID(missing=lambda: str(uuid.uuid1()))
    birthdate = fields.DateTime(default=lambda: dt.datetime(2017, 9, 19).isoformat())


# 3.x
class UserSchema(Schema):
    id = fields.UUID(missing=uuid.uuid1)
    birthdate = fields.DateTime(default=dt.datetime(2017, 9, 19))

Pass default as a keyword argument

fields.Boolean now receives additional truthy and falsy parameters. Consequently, the default parameter should always be passed as a keyword argument.

# 2.x
fields.Boolean(True)

# 3.x
fields.Boolean(default=True)

Email and URL fields do not validate on serialization

fields.Email and fields.URL only validate input upon deserialization. They do not validate on serialization. This makes them more consistent with the other fields and improves serialization performance.

load_from and dump_to are merged into data_key

The same key is used for serialization and deserialization.

# 2.x
class UserSchema(Schema):
    email = fields.Email(load_from="CamelCasedEmail", dump_to="CamelCasedEmail")


# 3.x
class UserSchema(Schema):
    email = fields.Email(data_key="CamelCasedEmail")

It is not possible to specify a different key for serialization and deserialization on the same field. This use case is covered by using two different Schema.

from marshmallow import Schema, fields


# 2.x
class UserSchema(Schema):
    id = fields.Str()
    email = fields.Email(load_from="CamelCasedEmail", dump_to="snake_case_email")


# 3.x
class BaseUserSchema(Schema):
    id = fields.Str()


class LoadUserSchema(BaseUserSchema):
    email = fields.Email(data_key="CamelCasedEmail")


class DumpUserSchema(BaseUserSchema):
    email = fields.Email(data_key="snake_case_email")

Also, when data_key is specified on a field, only data_key is checked in the input data. In marshmallow 2.x the field name is checked if load_from is missing from the input data.

Pre/Post-processors must return modified data

In marshmallow 2.x, None returned by a pre or post-processor is interpreted as “the data was mutated”. In marshmallow 3.x, the return value is considered as processed data even if it is None.

Processors that mutate the data should be updated to also return it.

# 2.x
class UserSchema(Schema):
    name = fields.Str()
    slug = fields.Str()

    @pre_load
    def slugify_name(self, in_data):
        # In 2.x, implicitly returning None implied that data were mutated
        in_data["slug"] = in_data["slug"].lower().strip().replace(" ", "-")


# 3.x
class UserSchema(Schema):
    name = fields.Str()
    slug = fields.Str()

    @pre_load
    def slugify_name(self, in_data, **kwargs):
        # In 3.x, always return the processed data
        in_data["slug"] = in_data["slug"].lower().strip().replace(" ", "-")
        return in_data

Nested field no longer supports plucking

In marshmallow 2.x, when a string was passed to a Nested field’s `only parameter, the field would be plucked. In marshmallow 3.x, the Pluck field must be used instead.

# 2.x
class UserSchema(Schema):
    name = fields.Str()
    friends = fields.Nested("self", many=True, only="name")


# 3.x
class UserSchema(Schema):
    name = fields.Str()
    friends = fields.Pluck("self", "name", many=True)

Accessing attributes on objects within a list

In order to serialize attributes on inner objects within a list, use the Pluck field.

# 2.x
class FactorySchema(Schema):
    widget_ids = fields.List(fields.Int(attribute="id"))


# 3.x
class FactorySchema(Schema):
    widget_ids = fields.List(fields.Pluck(WidgetSchema, "id"))

List does not wrap single values in a list on serialization

In marshmallow 2.x, List serializes a single object as a list with a single element. In marshmallow 3.x, the object is assumed to be iterable and passing a non-iterable element results in an error.

class UserSchema(Schema):
    numbers = fields.List(fields.Int())


user = {"numbers": 1}
UserSchema().dump(user)

# 2.x
# => {'numbers': [1]}

# 3.x
# => TypeError: 'int' object is not iterable

Float field takes a new allow_nan parameter

In marshmallow 2.x, Float field would serialize and deserialize special values such as nan, inf or -inf. In marshmallow 3, those values trigger a ValidationError unless allow_nan is True. allow_nan defaults to False.

# 2.x
class MySchema(Schema):
    x = fields.Float()


MySchema().load({"x": "nan"})
# => {{'x': nan}}


# 3.x
class MySchema(Schema):
    x = fields.Float()
    y = fields.Float(allow_nan=True)


MySchema().load({"x": 12, "y": "nan"})
# => {{'x': 12.0, 'y': nan}}

MySchema().load({"x": "nan"})
# marshmallow.exceptions.ValidationError: {'x': ['Special numeric values (nan or infinity) are not permitted.']}

DateTime field dateformat Meta option is renamed datetimeformat

The Meta option dateformat used to pass format to DateTime field is renamed as datetimeformat.

Date field gets a new format parameter to specify the format to use for serialization. dateformat Meta option now applies to Date field.

# 2.x
class MySchema(Schema):
    x = fields.DateTime()

    class Meta:
        dateformat = "%Y-%m"


MySchema().dump({"x": dt.datetime(2017, 9, 19)})
# => {{'x': '2017-09'}}


# 3.x
class MySchema(Schema):
    x = fields.DateTime()
    y = fields.Date()

    class Meta:
        datetimeformat = "%Y-%m"
        dateformat = "%m-%d"


MySchema().dump({"x": dt.datetime(2017, 9, 19), "y": dt.date(2017, 9, 19)})
# => {{'x': '2017-09', 'y': '09-19'}}

DateTime leaves timezone information untouched during serialization

DateTime does not convert naive datetimes to UTC on serialization and LocalDateTime is removed.

# 2.x
class MySchema(Schema):
    x = fields.DateTime()
    y = fields.DateTime()
    z = fields.LocalDateTime()


MySchema().dump(
    {
        "x": dt.datetime(2017, 9, 19),
        "y": dt.datetime(2017, 9, 19, tzinfo=dt.timezone(dt.timedelta(hours=2))),
        "z": dt.datetime(2017, 9, 19, tzinfo=dt.timezone(dt.timedelta(hours=2))),
    }
)
# => {{'x': '2017-09-19T00:00:00+00:00', 'y': '2017-09-18T22:00:00+00:00', 'z': '2017-09-19T00:00:00+02:00'}}


# 3.x
class MySchema(Schema):
    x = fields.DateTime()
    y = fields.DateTime()


MySchema().dump(
    {
        "x": dt.datetime(2017, 9, 19),
        "y": dt.datetime(2017, 9, 19, tzinfo=dt.timezone(dt.timedelta(hours=2))),
    }
)
# => {{'x': '2017-09-19T00:00:00', 'y': '2017-09-19T00:00:00+02:00'}}

The prefix Schema parameter is removed

The prefix parameter of Schema is removed. The same feature can be achieved using a post_dump <marshmallow.decorators.post_dump>` method.

# 2.x
class MySchema(Schema):
    f1 = fields.Field()
    f2 = fields.Field()


MySchema(prefix="pre_").dump({"f1": "one", "f2": "two"})
# {'pre_f1': 'one', '_pre_f2': 'two'}


# 3.x
class MySchema(Schema):
    f1 = fields.Field()
    f2 = fields.Field()

    @post_dump
    def prefix_usr(self, data):
        return {"usr_{}".format(k): v for k, v in iteritems(data)}


MySchema().dump({"f1": "one", "f2": "two"})
# {'pre_f1': 'one', '_pre_f2': 'two'}

fields.FormattedString is removed

fields.FormattedString field is removed. Use fields.Function or fields.Method instead.

# 2.x
class MySchema(Schema):
    full_name = fields.FormattedString("{first_name} {last_name}")


# 3.x
class MySchema(Schema):
    full_name = fields.Function(lambda u: f"{u.first_name} {u.last_name}")

attribute or data_key collision triggers an exception

When a Schema is instantiated, a check is performed and a ValueError is triggered if

  • several fields have the same attribute value (or field name if attribute is not passed), excluding dump_only fields, or

  • several fields have the same data_key value (or field name if data_key is not passed), excluding load_only fields

In marshmallow 2, it was possible to have multiple fields with the same attribute. It would work provided the Schema was only used for dumping. When loading, the behaviour was undefined. In marshmallow 3, all but one of those fields must be marked as dump_only. Likewise for data_key (formerly dump_to) for fields that are not load_only.

# 2.x
class MySchema(Schema):
    f1 = fields.Field()
    f2 = fields.Field(attribute="f1")
    f3 = fields.Field(attribute="f5")
    f4 = fields.Field(attribute="f5")


MySchema()
#  No error


# 3.x
class MySchema(Schema):
    f1 = fields.Field()
    f2 = fields.Field(attribute="f1")
    f3 = fields.Field(attribute="f5")
    f4 = fields.Field(attribute="f5")


MySchema()
# ValueError: 'Duplicate attributes: ['f1', 'f5]'


class MySchema(Schema):
    f1 = fields.Field()
    f2 = fields.Field(attribute="f1", dump_only=True)
    f3 = fields.Field(attribute="f5")
    f4 = fields.Field(attribute="f5", dump_only=True)


MySchema()
# No error

Field.fail is deprecated in favor of Field.make_error

Field.fail is deprecated. Use Field.make_error. This allows you to re-raise exceptions using raise ... from ....

from marshmallow import fields, ValidationError
from packaging import version


# 2.x
class Version(fields.Field):
    default_error_messages = {"invalid": "Not a valid version."}

    def _deserialize(self, value, *args, **kwargs):
        try:
            return version.Version(value)
        except version.InvalidVersion:
            self.fail("invalid")


# 3.x
class Version(fields.Field):
    default_error_messages = {"invalid": "Not a valid version."}

    def _deserialize(self, value, *args, **kwargs):
        try:
            return version.Version(value)
        except version.InvalidVersion as error:
            raise self.make_error("invalid") from error

Custom Fields

To make your custom fields compatible with marshmallow 3, _deserialize should accept **kwargs:

from marshmallow import fields, ValidationError
from packaging import version


# 2.x
class MyCustomField(fields.Field):
    def _deserialize(self, value, attr, obj):
        ...


# 3.x
class MyCustomField(fields.Field):
    def _deserialize(self, value, attr, obj, **kwargs):
        ...

Upgrading to 2.3

The func parameter of fields.Function was renamed to serialize.

# YES
lowername = fields.Function(serialize=lambda obj: obj.name.lower())
# or
lowername = fields.Function(lambda obj: obj.name.lower())

# NO
lowername = fields.Function(func=lambda obj: obj.name.lower())

Similarly, the method_name of fields.Method was also renamed to serialize.

# YES
lowername = fields.Method(serialize="lowercase")
# or
lowername = fields.Method("lowercase")

# NO
lowername = fields.Method(method_name="lowercase")

The func parameter is still available for backwards-compatibility. It will be removed in marshmallow 3.0.

Both fields.Function and fields.Method will allow the serialize parameter to not be passed, in this case use the deserialize parameter by name.

lowername = fields.Function(deserialize=lambda name: name.lower())
# or
lowername = fields.Method(deserialize="lowername")

Upgrading to 2.0

Deserializing None

In 2.0, validation/deserialization of None is consistent across field types. If allow_none is False (the default), validation fails when the field’s value is None. If allow_none is True, None is considered valid, and the field deserializes to None.

from marshmallow import fields

# In 1.0, deserialization of None was inconsistent
fields.Int().deserialize(None)  # 0
fields.Str().deserialize(None)  # ''
fields.DateTime().deserialize(None)  # error: Could not deserialize None to a datetime.


# In 2.0, validation/deserialization of None is consistent
fields.Int().deserialize(None)  # error: Field may not be null.
fields.Str().deserialize(None)  # error: Field may not be null.
fields.DateTime().deserialize(None)  # error: Field may not be null.

# allow_none makes None a valid value
fields.Int(allow_none=True).deserialize(None)  # None

Default Values

Before version 2.0, certain fields (including String, List, Nested, and number fields) had implicit default values that would be used if their corresponding input value was None or missing.

In 2.0, these implicit defaults are removed. A Field's default parameter is only used if you explicitly set it. Otherwise, missing inputs will be excluded from the serialized output.

from marshmallow import Schema, fields


class MySchema(Schema):
    str_no_default = fields.Str()
    int_no_default = fields.Int()
    list_no_default = fields.List(fields.Str)


schema = MySchema()

# In 1.0, None was treated as a missing input, so implicit default values were used
schema.dump(
    {"str_no_default": None, "int_no_default": None, "list_no_default": None}
).data
# {'str_no_default': '', 'int_no_default': 0, 'list_no_default': []}

# In 2.0, None serializes to None. No more implicit defaults.
schema.dump(
    {"str_no_default": None, "int_no_default": None, "list_no_default": None}
).data
# {'str_no_default': None, 'int_no_default': None, 'list_no_default': None}
# In 1.0, implicit default values were used for missing inputs
schema.dump({}).data
# {'int_no_default': 0, 'str_no_default': '', 'list_no_default': []}

# In 2.0, missing inputs are excluded from the serialized output
# if no defaults are specified
schema.dump({}).data
# {}

As a consequence of this new behavior, the skip_missing class Meta option has been removed.

Pre-processing and Post-processing Methods

The pre- and post-processing API was significantly improved for better consistency and flexibility. The pre_load, post_load, pre_dump, and post_dump should be used to define processing hooks. Schema.preprocessor and Schema.data_handler are removed.

# 1.0 API
from marshmallow import Schema, fields


class ExampleSchema(Schema):
    field_a = fields.Int()


@ExampleSchema.preprocessor
def increment(schema, data):
    data["field_a"] += 1
    return data


@ExampleSchema.data_handler
def decrement(schema, data, obj):
    data["field_a"] -= 1
    return data


# 2.0 API
from marshmallow import Schema, fields, pre_load, post_dump


class ExampleSchema(Schema):
    field_a = fields.Int()

    @pre_load
    def increment(self, data):
        data["field_a"] += 1
        return data

    @post_dump
    def decrement(self, data):
        data["field_a"] -= 1
        return data

See the Extending Schemas page for more information on the pre_* and post_* decorators.

Schema Validators

Similar to pre-processing and post-processing methods, schema validators are now defined as methods. Decorate schema validators with validates_schema. Schema.validator is removed.

# 1.0 API
from marshmallow import Schema, fields, ValidationError


class MySchema(Schema):
    field_a = fields.Int(required=True)
    field_b = fields.Int(required=True)


@ExampleSchema.validator
def validate_schema(schema, data):
    if data["field_a"] < data["field_b"]:
        raise ValidationError("field_a must be greater than field_b")


# 2.0 API
from marshmallow import Schema, fields, validates_schema, ValidationError


class MySchema(Schema):
    field_a = fields.Int(required=True)
    field_b = fields.Int(required=True)

    @validates_schema
    def validate_schema(self, data):
        if data["field_a"] < data["field_b"]:
            raise ValidationError("field_a must be greater than field_b")

Custom Accessors and Error Handlers

Custom accessors and error handlers are now defined as methods. Schema.accessor and Schema.error_handler are deprecated.

from marshmallow import Schema, fields


# 1.0 Deprecated API
class ExampleSchema(Schema):
    field_a = fields.Int()


@ExampleSchema.accessor
def get_from_dict(schema, attr, obj, default=None):
    return obj.get(attr, default)


@ExampleSchema.error_handler
def handle_errors(schema, errors, obj):
    raise CustomError("Something bad happened", messages=errors)


# 2.0 API
class ExampleSchema(Schema):
    field_a = fields.Int()

    def get_attribute(self, attr, obj, default):
        return obj.get(attr, default)

    # handle_error gets passed a ValidationError
    def handle_error(self, exc, data):
        raise CustomError("Something bad happened", messages=exc.messages)

Use post_load instead of make_object

The make_object method was deprecated from the Schema API (see #277 for the rationale). In order to deserialize to an object, use a post_load method.

# 1.0
from marshmallow import Schema, fields, post_load


class UserSchema(Schema):
    name = fields.Str()
    created_at = fields.DateTime()

    def make_object(self, data):
        return User(**data)


# 2.0
from marshmallow import Schema, fields, post_load


class UserSchema(Schema):
    name = fields.Str()
    created_at = fields.DateTime()

    @post_load
    def make_user(self, data):
        return User(**data)

Error Format when many=True

When validating a collection (i.e. when calling load or dump with many=True), the errors dictionary will be keyed on the indices of invalid items.

from marshmallow import Schema, fields


class BandMemberSchema(Schema):
    name = fields.String(required=True)
    email = fields.Email()


user_data = [
    {"email": "mick@stones.com", "name": "Mick"},
    {"email": "invalid", "name": "Invalid"},  # invalid email
    {"email": "keith@stones.com", "name": "Keith"},
    {"email": "charlie@stones.com"},  # missing "name"
]

result = BandMemberSchema(many=True).load(user_data)

# 1.0
result.errors
# {'email': ['"invalid" is not a valid email address.'],
#  'name': ['Missing data for required field.']}

# 2.0
result.errors
# {1: {'email': ['"invalid" is not a valid email address.']},
#  3: {'name': ['Missing data for required field.']}}

You can still get the pre-2.0 behavior by setting index_errors = False in a Schema's class Meta options.

Use ValidationError instead of MarshallingError and UnmarshallingError

The MarshallingError and UnmarshallingError exceptions are deprecated in favor of a single ValidationError. Users who have written custom fields or are using strict mode will need to change their code accordingly.

Handle ValidationError in strict mode

When using strict mode, you should handle ValidationErrors when calling Schema.dump and Schema.load.

from marshmallow import exceptions as exc

schema = BandMemberSchema(strict=True)

# 1.0
try:
    schema.load({"email": "invalid-email"})
except exc.UnmarshallingError as err:
    handle_error(err)

# 2.0
try:
    schema.load({"email": "invalid-email"})
except exc.ValidationError as err:
    handle_error(err)

Accessing error messages in strict mode

In 2.0, strict mode was improved so that you can access all error messages for a schema (rather than failing early) by accessing a ValidationError's messages attribute.

schema = BandMemberSchema(strict=True)

try:
    result = schema.load({"email": "invalid"})
except ValidationMessage as err:
    print(err.messages)
# {
#     'email': ['"invalid" is not a valid email address.'],
#     'name': ['Missing data for required field.']
# }

Custom Fields

Two changes must be made to make your custom fields compatible with version 2.0.

  • The _deserialize method of custom fields now receives attr (the key corresponding to the value to be deserialized) and the raw input data as arguments.

  • Custom fields should raise ValidationError in their _deserialize and _serialize methods when a validation error occurs.

from marshmallow import fields, ValidationError
from marshmallow.exceptions import UnmarshallingError


# In 1.0, an UnmarshallingError was raised
class PasswordField(fields.Field):
    def _deserialize(self, val):
        if not len(val) >= 6:
            raise UnmarshallingError("Password too short.")
        return val


# In 2.0, _deserialize receives attr and data,
# and a ValidationError is raised
class PasswordField(fields.Field):
    def _deserialize(self, val, attr, data):
        if not len(val) >= 6:
            raise ValidationError("Password too short.")
        return val

To make a field compatible with both marshmallow 1.x and 2.x, you can pass *args and **kwargs to the signature.

class PasswordField(fields.Field):
    def _deserialize(self, val, *args, **kwargs):
        if not len(val) >= 6:
            raise ValidationError("Password too short.")
        return val

Custom Error Messages

Error messages can be customized at the Field class or instance level.

# 1.0
field = fields.Number(error="You passed a bad number")

# 2.0
# Instance-level
field = fields.Number(error_messages={"invalid": "You passed a bad number."})


# Class-level
class MyNumberField(fields.Number):
    default_error_messages = {"invalid": "You passed a bad number."}

Passing a string to required is deprecated.

# 1.0
field = fields.Str(required="Missing required argument.")

# 2.0
field = fields.Str(error_messages={"required": "Missing required argument."})

Use OneOf instead of fields.Select

The fields.Select field is deprecated in favor of the newly-added OneOf validator.

from marshmallow import fields
from marshmallow.validate import OneOf

# 1.0
fields.Select(["red", "blue"])

# 2.0
fields.Str(validate=OneOf(["red", "blue"]))

Accessing Context from Method fields

Use self.context to access a schema’s context within a Method field.

class UserSchema(Schema):
    name = fields.String()
    likes_bikes = fields.Method("writes_about_bikes")

    def writes_about_bikes(self, user):
        return "bicycle" in self.context["blog"].title.lower()

Validation Error Messages

The default error messages for many fields and validators have been changed for better consistency.

from marshmallow import Schema, fields, validate


class ValidatingSchema(Schema):
    foo = fields.Str()
    bar = fields.Bool()
    baz = fields.Int()
    qux = fields.Float()
    spam = fields.Decimal(2, 2)
    eggs = fields.DateTime()
    email = fields.Str(validate=validate.Email())
    homepage = fields.Str(validate=validate.URL())
    nums = fields.List(fields.Int())


schema = ValidatingSchema()
invalid_data = {
    "foo": 42,
    "bar": 24,
    "baz": "invalid-integer",
    "qux": "invalid-float",
    "spam": "invalid-decimal",
    "eggs": "invalid-datetime",
    "email": "invalid-email",
    "homepage": "invalid-url",
    "nums": "invalid-list",
}
errors = schema.validate(invalid_data)
# {
#     'foo': ['Not a valid string.'],
#     'bar': ['Not a valid boolean.'],
#     'baz': ['Not a valid integer.'],
#     'qux': ['Not a valid number.'],
#     'spam': ['Not a valid number.']
#     'eggs': ['Not a valid datetime.'],
#     'email': ['Not a valid email address.'],
#     'homepage': ['Not a valid URL.'],
#     'nums': ['Not a valid list.'],
# }

More

For a full list of changes in 2.0, see the Changelog.

Upgrading to 1.2

Validators

Validators were rewritten as class-based callables, making them easier to use when declaring fields.

from marshmallow import fields

# 1.2
from marshmallow.validate import Range

age = fields.Int(validate=[Range(min=0, max=999)])

# Pre-1.2
from marshmallow.validate import ranging

age = fields.Int(validate=[lambda val: ranging(val, min=0, max=999)])

The validator functions from 1.1 are deprecated and will be removed in 2.0.

Deserializing the Empty String

In version 1.2, deserialization of the empty string ('') with DateTime, Date, Time, or TimeDelta fields results in consistent error messages, regardless of whether or not python-dateutil is installed.

from marshmallow import fields

fields.Date().deserialize("")
# UnmarshallingError: Could not deserialize '' to a date object.

Decimal

The Decimal field was added to support serialization/deserialization of decimal.Decimal numbers. You should use this field when dealing with numbers where precision is critical. The Fixed, Price, and Arbitrary fields are deprecated in favor the Decimal field.

Upgrading to 1.0

Version 1.0 marks the first major release of marshmallow. Many big changes were made from the pre-1.0 releases in order to provide a cleaner API, support object deserialization, and improve field validation.

Perhaps the largest change is in how objects get serialized. Serialization occurs by invoking the Schema.dump() method rather than passing the object to the constructor. Because only configuration options (e.g. the many, strict, and only parameters) are passed to the constructor, you can more easily reuse serializer instances. The dump method also forms a nice symmetry with the Schema.load() method, which is used for deserialization.

from marshmallow import Schema, fields


class UserSchema(Schema):
    email = fields.Email()
    name = fields.String()


user = User(email="monty@python.org", name="Monty Python")

# 1.0
serializer = UserSchema()
data, errors = serializer.dump(user)
# OR
result = serializer.dump(user)
result.data  # => serialized result
result.errors  # => errors

# Pre-1.0
serialized = UserSchema(user)
data = serialized.data
errors = serialized.errors

Note

Some crucial parts of the pre-1.0 API have been retained to ease the transition. You can still pass an object to a Schema constructor and access the Schema.data and Schema.errors properties. The is_valid method, however, has been completely removed. It is recommended that you migrate to the new API to prevent future releases from breaking your code.

The Fields interface was also reworked in 1.0 to make it easier to define custom fields with their own serialization and deserialization behavior. Custom fields now implement Field._serialize() and Field._deserialize().

from marshmallow import fields, MarshallingError


class PasswordField(fields.Field):
    def _serialize(self, value, attr, obj):
        if not value or len(value) < 6:
            raise MarshallingError("Password must be greater than 6 characters.")
        return str(value).strip()

    # Similarly, you can override the _deserialize method

Another major change in 1.0 is that multiple validation errors can be stored for a single field. The errors dictionary returned by Schema.dump() and Schema.load() is a list of error messages keyed by field name.

from marshmallow import Schema, fields, ValidationError


def must_have_number(val):
    if not any(ch.isdigit() for ch in val):
        raise ValidationError("Value must have an number.")


def validate_length(val):
    if len(val) < 8:
        raise ValidationError("Value must have 8 or more characters.")


class ValidatingSchema(Schema):
    password = fields.String(validate=[must_have_number, validate_length])


result, errors = ValidatingSchema().load({"password": "secure"})
print(errors)
# {'password': ['Value must have an number.',
#               'Value must have 8 or more characters.']}

Other notable changes:

  • Serialized output is no longer an OrderedDict by default. You must explicitly set the ordered class Meta option to True .

  • Serializer has been renamed to Schema, but you can still import marshmallow.Serializer (which is aliased to Schema).

  • datetime objects serialize to ISO8601-formatted strings by default (instead of RFC821 format).

  • The fields.validated decorator was removed, as it is no longer necessary given the new Fields interface.

  • Schema.factory class method was removed.

See also

See the Changelog for a more complete listing of added features, bugfixes and breaking changes.