Single Table Inheritance in Django
There are currently two forms of inheritance in Django - MTI (model table inheritance) and ABC (abstract base classes).
I wrote a tutorial on what's going on under the hood.
You can also reference the official docs on model inheritance.
Summary
Django's proxy models provide the basis for Single Table Inheritance.
However, some effort is required to make it work.
Skip to the end for a re-usable example.
Background
Martin Fowler describes Single Table Inheritance (STI) as follows:
Single Table Inheritance maps all fields of all classes of an inheritance structure into a single table.
This is precisely what Django's proxy model inheritance does.
Note, that, according to this blog post from 2010, proxy
models have been around since Django 1.1.
A "normal" Django model is a concrete model, i.e. it has a dedicated table in the database. There are two types of Django model that do not have dedicated database tables, viz. abstract models and proxy models:
Abstract models act as superclasses for concrete models. An abstract model can define fields, but it does not have a database table. The fields are only added to the database tables for its concrete subclasses.
Proxy models act as subclasses for concrete models. A proxy model cannot define new fields. Instead, it operates on the database table associated with its concrete superclass. In other words, a Django concrete model and its proxies all share a single table.
Django's proxy models provide the basis for Single Table Inheritance, viz. they allow different models to share a single table, and they allow us to define proxy-specific behavior on the Python side. However, Django's default object-relational mapping (ORM) does not provide all the behavior that would be expected, so a little customization is required. How much, that depends on your needs.
Let's build a minimal example, step by step, based on the simple data-model in the figure below:
Step 1: basic "proxy model inheritance"
Here's the content of models.py
for a basic proxy inheritance implementation:
from django.db import models
class Party(models.Model):
name = models.CharField(max_length=20)
person_attribute = models.CharField(max_length=20)
organization_attribute = models.CharField(max_length=20)
class Person(Party):
class Meta:
proxy = True
class Organization(Party):
class Meta:
proxy = True
Person
and Organization
are two types of parties.
Only the Party
model has a database table, so all the fields are defined on this model, including any fields that are specific either to Person
or to Organization
.
Because Party
, Person
, and Organization
all use the Party
database table, we can define a single ForeignKey
field to Party
, and assign instances of any of the three models to that field, as implied by the inheritance relation in the figure. Note, that, without inheritance, we would need a separate ForeignKey
field for each model.
For example, suppose we define an Address
model as follows:
class Address(models.Model):
party = models.ForeignKey(to=Party, on_delete=models.CASCADE)
We can then initialize an Address
object using e.g. Address(party=person_instance)
or Address(party=organization_instance)
.
So far, so good.
However, if we try to get a list of objects corresponding to a proxy model, using e.g. Person.objects.all()
, we get a list of all Party
objects instead, i.e. both Person
objects and Organization
objects. This is because the proxy models still use the model manager from the superclass (i.e. Party
).
Step 2: add proxy model managers
To make sure that Person.objects.all()
only returns Person
objects, we need to assign a separate model manager that filters the Party
queryset. To enable this filtering, we need a field that indicates which proxy model should be used for the object.
To be clear: creating a Person
object implies adding a row to the Party
table. The same goes for Organization
. To distinguish between the two, we need a column to indicate if a row represents a Person
or an Organization
. For convenience and clarity, we add a field (i.e. column) called proxy_name
, and use that to store the name of the proxy class.
So, enter the ProxyManager
model manager and the proxy_name
field:
from django.db import models
class ProxyManager(models.Manager):
def get_queryset(self):
return super().get_queryset().filter(proxy_name=self.model.__name__)
class Party(models.Model):
proxy_name = models.CharField(max_length=20)
name = models.CharField(max_length=20)
person_attribute = models.CharField(max_length=20)
organization_attribute = models.CharField(max_length=20)
def save(self, *args, **kwargs):
self.proxy_name = type(self).__name__
super().save(*args, **kwargs)
class Person(Party):
class Meta:
proxy = True
objects = ProxyManager()
class Organization(Party):
class Meta:
proxy = True
objects = ProxyManager()
Now the queryset returned by Person.objects.all()
will only contain Person
objects (and the same for Organization
).
However, this does not work in the case of a ForeignKey
relation to Party
, as in Address.party
above, because that will always return a Party
instance, regardless of the value of the proxy_name
field (also see docs). For example, suppose we create an address = Address(party=person_instance)
, then address.party
will return a Party
instance, instead of a Person
instance.
Step 3: extend the Party
constructor
One way to deal with the related-field issue is to extend the Party.__new__
method, so it returns an instance of the class specified in the 'proxy_name' field. The end result looks like this:
class Party(models.Model):
PROXY_FIELD_NAME = 'proxy_name'
proxy_name = models.CharField(max_length=20)
name = models.CharField(max_length=20)
person_attribute = models.CharField(max_length=20)
organization_attribute = models.CharField(max_length=20)
def save(self, *args, **kwargs):
""" automatically store the proxy class name in the database """
self.proxy_name = type(self).__name__
super().save(*args, **kwargs)
def __new__(cls, *args, **kwargs):
party_class = cls
try:
# get proxy name, either from kwargs or from args
proxy_name = kwargs.get(cls.PROXY_FIELD_NAME)
if proxy_name is None:
proxy_name_field_index = cls._meta.fields.index(
cls._meta.get_field(cls.PROXY_FIELD_NAME))
proxy_name = args[proxy_name_field_index]
# get proxy class, by name, from current module
party_class = getattr(sys.modules[__name__], proxy_name)
finally:
return super().__new__(party_class)
Now address.party
will actually return a Person
instance if the proxy_name
field is Person
.
As a last step, we can make the whole thing re-usable:
Step 4: make it re-usable
To make our rudimentary Single-Table Inheritance implementation re-usable, we can use Django's abstract inheritance:
inheritance/models.py
:
import sys
from django.db import models
class ProxySuper(models.Model):
class Meta:
abstract = True
proxy_name = models.CharField(max_length=20)
def save(self, *args, **kwargs):
""" automatically store the proxy class name in the database """
self.proxy_name = type(self).__name__
super().save(*args, **kwargs)
def __new__(cls, *args, **kwargs):
""" create an instance corresponding to the proxy_name """
proxy_class = cls
try:
field_name = ProxySuper._meta.get_fields()[0].name
proxy_name = kwargs.get(field_name)
if proxy_name is None:
proxy_name_field_index = cls._meta.fields.index(
cls._meta.get_field(field_name))
proxy_name = args[proxy_name_field_index]
proxy_class = getattr(sys.modules[cls.__module__], proxy_name)
finally:
return super().__new__(proxy_class)
class ProxyManager(models.Manager):
def get_queryset(self):
""" only include objects in queryset matching current proxy class """
return super().get_queryset().filter(proxy_name=self.model.__name__)
Then we can implement our inheritance structure as follows:
parties/models.py
:
from django.db import models
from inheritance.models import ProxySuper, ProxyManager
class Party(ProxySuper):
name = models.CharField(max_length=20)
person_attribute = models.CharField(max_length=20)
organization_attribute = models.CharField(max_length=20)
class Person(Party):
class Meta:
proxy = True
objects = ProxyManager()
class Organization(Party):
class Meta:
proxy = True
objects = ProxyManager()
class Placement(models.Model):
party = models.ForeignKey(to=Party, on_delete=models.CASCADE)
More work may be required, depending on your needs, but I believe this covers some of the basics.
I think the OP is asking about Single-Table Inheritance as defined here:
Relational databases don't support inheritance, so when mapping from objects to databases we have to consider how to represent our nice inheritance structures in relational tables. When mapping to a relational database, we try to minimize the joins that can quickly mount up when processing an inheritance structure in multiple tables. Single Table Inheritance maps all fields of all classes of an inheritance structure into a single table.
That is, a single database table for a whole hierarchy of entity classes. Django does not support that kind of inheritance.