Serde¶
Serde is a lightweight, general-purpose framework for defining, serializing, deserializing, and validating data structures in Python.
Getting started¶
Installation¶
Serde is available on PyPI, you can install it using
pip install serde
Extended features can be installed with the ext feature.
pip install serde[ext]
Introduction¶
In Serde models are containers for fields. Data structures are defined by
subclassing Model and assigning Field instances as class annotations.
These fields handle serialization, deserialization, normalization, and
validation for the corresponding model attributes.
from datetime import date
from serde import Model, fields
class Artist(Model):
name: fields.Str()
class Album(Model):
title: fields.Str()
release_date: fields.Optional(fields.Date)
artist: fields.Nested(Artist)
album = Album(
title='Dangerously in Love',
release_date=date(2003, 6, 23),
artist=Artist(name='Beyoncé')
)
assert album.to_dict() == {
'title': 'Dangerously in Love',
'release_date': '2003-06-23',
'artist': {
'name': 'Beyoncé'
}
}
album = Album.from_json("""{
"title": "Lemonade",
"artist": {"name": "Beyoncé"}}"
""")
assert album == Album(title='Lemonade', artist=Artist(name='Beyoncé'))
Basic usage¶
Below we create a User model by subclassing Model and adding the
name and email fields.
>>> from datetime import datetime
>>> from serde import Model, fields
>>>
>>> class User(Model):
... name: fields.Str(rename='username')
... email: fields.Email()
The corresponding attribute names are used to instantiate the model object and access the values on the model instance.
>>> user = User(name='Linus Torvalds', email='torvalds@linuxfoundation.org')
>>> user.name
'Linus Torvalds'
>>> user.email
'torvalds@linuxfoundation.org'
Models are validated when they are instantiated and a ValidationError is
raised if you provide invalid values.
>>> User(name='Linus Torvalds', email='not an email')
Traceback (most recent call last):
...
serde.exceptions.ValidationError: {'email': 'invalid email'}
Models are serialized into primitive Python types using the to_dict() method
on the model instance.
>>> user.to_dict()
OrderedDict([('username', 'Linus Torvalds'), ('email', 'torvalds@linuxfoundation.org')])
Or to JSON using the to_json() method.
>>> user.to_json()
'{"username": "Linus Torvalds", "email": "torvalds@linuxfoundation.org"}'
Models are also validated when they are deserialized. Models are deserialized
from primitive Python types using the reciprocal from_dict() class method.
>>> user = User.from_dict({
... 'username': 'Donald Knuth',
... 'email': 'noreply@stanford.edu'
... })
Or from JSON using the from_json() method.
>>> user = User.from_json('''{
... "username": "Donald Knuth",
... "email": "noreply@stanford.edu"
... }''')
Attempting to deserialize invalid data will result in a ValidationError.
>>> User.from_dict({'username': 'Donald Knuth'})
Traceback (most recent call last):
...
serde.exceptions.ValidationError: {'email': "missing data, expected field 'email'"}
Models¶
Models can be nested and used in container-like fields. Below we create a
Blog with an author and a list of subscribers which must all be User
instances.
>>> class Blog(Model):
... title: fields.Str()
... author: fields.Nested(User)
... subscribers: fields.List(User)
When instantiating you have to supply instances of the nested models.
>>> blog = Blog(
... title="sobolevn's personal blog",
... author=User(name='Nikita Sobolev', email='mail@sobolevn.me'),
... subscribers=[
... User(name='Ned Batchelder', email='ned@nedbatchelder.com')
... ]
... )
Serializing a Blog would serialize the entire nested structure.
>>> print(blog.to_json(indent=2))
{
"title": "sobolevn's personal blog",
"author": {
"username": "Nikita Sobolev",
"email": "mail@sobolevn.me"
},
"subscribers": [
{
"username": "Ned Batchelder",
"email": "ned@nedbatchelder.com"
}
]
}
Similiarly deserializing a Blog would deserialize the entire nested
structure, and create instances of all the submodels.
Subclassed models¶
Models can be subclassed. The subclass will have all the fields of the parent
and any additional ones. Consider the case where we define a SuperUser model
which is a subclass of a User. Simply a User that has an extra level
field.
>>> class SuperUser(User):
... # inherits name and email fields from User
... level: fields.Choice(['admin', 'read-only'])
We instantiate a subclassed model as normal by passing in each field value.
>>> superuser = SuperUser(
... name='Linus Torvalds',
... email='torvalds@linuxfoundation.org',
... level='admin'
... )
This is great for many cases, however, a commonly desired paradigm is to be able
to have the User.from_dict() class method be able to deserialize a
SuperUser as well. This can be made possible through model tagging.
Model tagging¶
Model tagging is a way to mark serialized data in order to show that it is a
particular variant of a model. Serde provides three types of model tagging,
but you can also define you own custom Tag. A Tag can be thought of in
the same way as a Field but instead of deserializing data into an attribute
on a model instance, it deserializes data into a model class.
Internally tagged¶
Internally tagged data stores a tag value inside the serialized data.
Let us consider an example where we define a Pet model with a tag. We
can then subclass this model and deserialize arbitrary subclasses using the
tagged model.
>>> from serde import Model, fields, tags
>>>
>>> class Pet(Model):
... name: fields.Str()
...
... class Meta:
... tag = tags.Internal(tag='species')
...
>>> class Dog(Pet):
... hates_cats: fields.Bool()
...
>>> class Cat(Pet):
... hates_dogs: fields.Bool()
We refer to the Dog and Cat subclasses as variants of Pet. When
serializing all parent model tag serialization is done after field
serialization.
>>> Cat(name='Fluffy', hates_dogs=True).to_dict()
OrderedDict([('name', 'Fluffy'), ('hates_dogs', True), ('species', '__main__.Cat')])
When deserializing, tag deserialization is done first to determine which model to use for the deserialization.
>>> milo = Pet.from_dict({
... 'name': 'Milo',
... 'hates_cats': False,
... 'species': '__main__.Dog'
... })
>>> milo.__class__
<class '__main__.Dog'>
>>> milo.name
'Milo'
>>> milo.hates_cats
False
An invalid or missing tag will raise a ValidationError.
>>> Pet.from_dict({'name': 'Milo', 'hates_cats': False})
Traceback (most recent call last):
...
serde.exceptions.ValidationError: missing data, expected tag 'species'
>>>
>>> Pet.from_dict({'name': 'Duke', 'species': '__main__.Horse'})
Traceback (most recent call last):
...
serde.exceptions.ValidationError: no variant found
Externally tagged¶
Externally tagged data uses the tag value as a key and nests the content underneath that key. All other processes behave similarly to the internally tagged example above.
>>> class Pet(Model):
... name: fields.Str()
...
... class Meta:
... tag = tags.External()
...
>>> class Dog(Pet):
... hates_cats: fields.Bool()
...
>>> Dog(name='Max', hates_cats=True).to_dict()
OrderedDict([('__main__.Dog', OrderedDict([('name', 'Max'), ('hates_cats', True)]))])
Adjacently tagged¶
Adjacently tagged data data stores the tag value and the content underneath two separate keys. All other processes behave similarly to the internally tagged example.
>>> class Pet(Model):
... name: fields.Str()
...
... class Meta:
... tag = tags.Adjacent(tag='species', content='data')
...
>>> class Dog(Pet):
... hates_cats: fields.Bool()
...
>>> Dog(name='Max', hates_cats=True).to_dict()
OrderedDict([('species', '__main__.Dog'), ('data', OrderedDict([('name', 'Max'), ('hates_cats', True)]))])
Abstract models¶
By default model tagging still allows deserialization of the base model. It is
common to have this model be abstract. You can do this by setting the
abstract Meta field to True. This will make it uninstantiatable and it
won’t be included in the variant list during deserialization.
>>> class Fruit(Model):
... class Meta:
... abstract = True
...
>>> Fruit()
Traceback (most recent call last):
...
TypeError: unable to instantiate abstract model 'Fruit'
Fields¶
Fields do the work of serializing, deserializing, normalizing, and validating
the input values. Fields are always assigned to a model as instances , and
they support extra serialization, deserialization, normalization, and validation
of values without having to subclass Field. For example
from serde import Model, fields, validators
class Album(Model):
title: fields.Str(normalizers=[str.strip])
released: fields.Date(
rename='release_date',
validators=[validators.Min(datetime.date(1912, 4, 15))]
)
In the above example we define an Album class. The title field is of
type str , and we apply the str.strip normalizer to automatically strip
the input value when instantiating or deserializing the Album. The
released field is of type datetime.date and we apply an extra validator
to only accept dates after 15th April 1912. Note: the rename argument only
applies to the serializing and deserializing of the data, the Album class
would still be instantiated using Album(released=...).
If these methods of creating custom Field classes are not satisfactory, you
can always subclass a Field and override the relevant methods.
>>> class Percent(fields.Float):
... def validate(self, value):
... super().validate(value)
... validators.Between(0.0, 100.0)(value)
Model states and processes¶
In Serde, there are two states that the data can be in:
Serialized data
Model instance
There are five different processes that the data structure can go through when moving between these two states.
Deserialization happens when you create a model instance from a serialized version using
from_dict()or similar.Instantiation happens when you construct a model instance in Python using the
__init__()constructor.Normalization happens after instantiation and after deserialization. This is usually a way to transform things before they are validated. For example: this is where an
Optionalfield sets default values.Validation is where the model and fields values are validated. This happens after normalization.
Serialization is when you serialize a model instance to a supported serialization format using
to_dict()or similar.
The diagram below shows how the stages (uppercase) and processes (lowercase) fit in with each other.
+---------------+
| Instantiation |
+---------------+
|
v
+---------------+ +---------------+
|Deserialization|-->| Normalization |
+---------------+ +---------------+
^ |
| v
| +---------------+
| | Validation |
| +---------------+
| |
| v
+-------+-------+ +---------------+
|SERIALIZED DATA| | MODEL INSTANCE|
+---------------+ +---------------+
^ |
| |
+-------+-------+ |
| Serialization |<----------+
+---------------+
License¶
Serde is licensed under either of
Apache License, Version 2.0 (LICENSE-APACHE or https://www.apache.org/licenses/LICENSE-2.0)
MIT License (LICENSE-MIT or https://opensource.org/licenses/MIT)
at your option.