# Inheritance Out of various types of ORM models inheritance `ormar` currently supports two of them: * **Mixins** * **Concrete table inheritance** (with parents set to `abstract=True`) ## Types of inheritance The short summary of different types of inheritance: * **Mixins [SUPPORTED]** - don't subclass `ormar.Model`, just define fields that are later used on different models (like `created_date` and `updated_date` on each model), only actual models create tables, but those fields from mixins are added * **Concrete table inheritance [SUPPORTED]** - means that parent is marked as abstract and each child has its own table with columns from a parent and own child columns, kind of similar to Mixins but parent also is a Model * **Single table inheritance [NOT SUPPORTED]** - means that only one table is created with fields that are combination/sum of the parent and all children models but child models use only subset of column in db (all parent and own ones, skipping the other children ones) * **Multi/ Joined table inheritance [NOT SUPPORTED]** - means that part of the columns is saved on parent model and part is saved on child model that are connected to each other by kind of one to one relation and under the hood you operate on two models at once * **Proxy models [NOT SUPPORTED]** - means that only parent has an actual table, children just add methods, modify settings etc. ## Mixins To use Mixins just define a class that is not inheriting from an `ormar.Model` but is defining `ormar.Fields` as class variables. ```python base_ormar_config = ormar.OrmarConfig( database=databases.Database(DATABASE_URL), metadata=sqlalchemy.MetaData(), engine=sqlalchemy.create_engine(DATABASE_URL), ) # a mixin defines the fields but is a normal python class class AuditMixin: created_by: str = ormar.String(max_length=100) updated_by: str = ormar.String(max_length=100, default="Sam") class DateFieldsMixins: created_date: datetime.datetime = ormar.DateTime(default=datetime.datetime.now) updated_date: datetime.datetime = ormar.DateTime(default=datetime.datetime.now) # a models can inherit from one or more mixins class Category(ormar.Model, DateFieldsMixins, AuditMixin): ormar_config = base_ormar_config.copy(tablename="categories") id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=50, unique=True, index=True) code: int = ormar.Integer() ``` !!!tip Note that Mixins are **not** models, so you still need to inherit from `ormar.Model` as well as define `ormar_config` field in the **final** model. A Category class above will have four additional fields: `created_date`, `updated_date`, `created_by` and `updated_by`. There will be only one table created for model `Category` (`categories`), with `Category` class fields combined with all `Mixins` fields. Note that `Mixin` in class name is optional but is a good python practice. ## Concrete table inheritance In concept concrete table inheritance is very similar to Mixins, but uses actual `ormar.Models` as base classes. !!!warning Note that base classes have `abstract=True` set in `ormar_config` object, if you try to inherit from non abstract marked class `ModelDefinitionError` will be raised. Since this abstract Model will never be initialized you can skip `metadata` and `database` in it's `ormar_config` definition. But if you provide it - it will be inherited, that way you do not have to provide `metadata` and `databases` in the final/concrete class Note that you can always overwrite it in child/concrete class if you need to. More over at least one of the classes in inheritance chain have to provide both `database` and `metadata` - otherwise an error will be raised. ```python # note that base classes have abstract=True # since this model will never be initialized you can skip metadata and database class AuditModel(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) created_by: str = ormar.String(max_length=100) updated_by: str = ormar.String(max_length=100, default="Sam") # but if you provide it it will be inherited - DRY (Don't Repeat Yourself) in action class DateFieldsModel(ormar.Model): ormar_config = base_ormar_config.copy( abstract=True, metadata=metadata, database=db, ) created_date: datetime.datetime = ormar.DateTime(default=datetime.datetime.now) updated_date: datetime.datetime = ormar.DateTime(default=datetime.datetime.now) # that way you do not have to provide metadata and databases in concrete class class Category(DateFieldsModel, AuditModel): ormar_config = base_ormar_config.copy(tablename="categories") id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=50, unique=True, index=True) code: int = ormar.Integer() ``` The list of inherited options/settings is as follows: `metadata`, `database` and `constraints`. Of course apart from that all fields from base classes are combined and created in the concrete table of the final Model. !!!tip Note how you don't have to provide `abstarct=False` in the final class - it's the default setting that is not inherited. ## Redefining fields in subclasses Note that you can redefine previously created fields like in normal python class inheritance. Whenever you define a field with same name and new definition it will completely replace the previously defined one. ```python hl_lines="28" # base class class DateFieldsModel(ormar.Model): ormar_config = OrmarConfig( abstract=True, metadata=metadata, database=db, # note that UniqueColumns need sqlalchemy db columns names not the ormar ones constraints=[ormar.UniqueColumns("creation_date", "modification_date")] ) created_date: datetime.datetime = ormar.DateTime( default=datetime.datetime.now, name="creation_date" ) updated_date: datetime.datetime = ormar.DateTime( default=datetime.datetime.now, name="modification_date" ) class RedefinedField(DateFieldsModel): ormar_config = OrmarConfig( tablename="redefines", metadata=metadata, database=db, ) id: int = ormar.Integer(primary_key=True) # here the created_date is replaced by the String field created_date: str = ormar.String(max_length=200, name="creation_date") # you can verify that the final field is correctly declared and created changed_field = RedefinedField.ormar_config.model_fields["created_date"] assert changed_field.default is None assert changed_field.alias == "creation_date" assert any(x.name == "creation_date" for x in RedefinedField.ormar_config.table.columns) assert isinstance( RedefinedField.ormar_config.table.columns["creation_date"].type, sqlalchemy.sql.sqltypes.String, ) ``` !!!warning If you declare `UniqueColumns` constraint with column names, the final model **has to have** a column with the same name declared. Otherwise, the `ModelDefinitionError` will be raised. So in example above if you do not provide `name` for `created_date` in `RedefinedField` model ormar will complain. `created_date: str = ormar.String(max_length=200) # exception` `created_date: str = ormar.String(max_length=200, name="creation_date2") # exception` ## Relations in inheritance You can declare relations in every step of inheritance, so both in parent and child classes. When you define a relation on a child model level it's either overwriting the relation defined in parent model (if the same field name is used), or is accessible only to this child if you define a new relation. When inheriting relations, you always need to be aware of `related_name` parameter, that has to be unique across a related model, when you define multiple child classes that inherit the same relation. If you do not provide `related_name` parameter ormar calculates it for you. This works with inheritance as all child models have to have different class names, which are used to calculate the default `related_name` (class.name.lower()+'s'). But, if you provide a `related_name` this name cannot be reused in all child models as they would overwrite each other on the related model side. Therefore, you have two options: * redefine relation field in child models and manually provide different `related_name` parameters * let this for `ormar` to handle -> auto adjusted related_name are: original related_name + "_" + child model **table** name That might sound complicated but let's look at the following example: ### ForeignKey relations ```python # normal model used in relation class Person(ormar.Model): ormar_config = base_ormar_config.copy() id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=100) # parent model - needs to be abstract class Car(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=50) owner: Person = ormar.ForeignKey(Person) # note that we refer to the Person model again so we **have to** provide related_name co_owner: Person = ormar.ForeignKey(Person, related_name="coowned") created_date: datetime.datetime = ormar.DateTime(default=datetime.datetime.now) class Truck(Car): ormar_config = base_ormar_config.copy() max_capacity: int = ormar.Integer() class Bus(Car): ormar_config = base_ormar_config.copy(tablename="buses") max_persons: int = ormar.Integer() ``` Now when you will inspect the fields on Person model you will get: ```python Person.ormar_config.model_fields """ {'id': , 'name': , 'trucks': , 'coowned_trucks': , 'buss': , 'coowned_buses': } """ ``` Note how you have `trucks` and `buss` fields that leads to Truck and Bus class that this Person owns. There were no `related_name` parameter so default names were used. At the same time the co-owned cars need to be referenced by `coowned_trucks` and `coowned_buses`. Ormar appended `_trucks` and `_buses` suffixes taken from child model table names. Seems fine, but the default name for owned trucks is ok (`trucks`) but the `buss` is ugly, so how can we change it? The solution is pretty simple - just redefine the field in Bus class and provide different `related_name` parameter. ```python # rest of the above example remains the same class Bus(Car): ormar_config = base_ormar_config.copy(tablename="buses") # new field that changes the related_name owner: Person = ormar.ForeignKey(Person, related_name="buses") max_persons: int = ormar.Integer() ``` Now the columns looks much better. ```python Person.ormar_config.model_fields """ {'id': , 'name': , 'trucks': , 'coowned_trucks': , 'buses': , 'coowned_buses': } """ ``` !!!note You could also provide `related_name` for the `owner` field, that way the proper suffixes would be added. `owner: Person = ormar.ForeignKey(Person, related_name="owned")` and model fields for Person owned cars would become `owned_trucks` and `owned_buses`. ### ManyToMany relations Similarly, you can inherit from Models that have ManyToMany relations declared but there is one, but substantial difference - the Through model. Since the Through model will be able to hold additional fields, and now it links only two Tables (`from` and `to` ones), each child that inherits the m2m relation field has to have separate Through model. Of course, you can overwrite the relation in each Child model, but that requires additional code and undermines the point of the whole inheritance. `Ormar` will handle this for you if you agree with default naming convention, which you can always manually overwrite in children if needed. Again, let's look at the example to easier grasp the concepts. We will modify the previous example described above to use m2m relation for co_owners. ```python # person remain the same as above class Person(ormar.Model): ormar_config = base_ormar_config.copy() id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=100) # new through model between Person and Car2 class PersonsCar(ormar.Model): ormar_config = base_ormar_config.copy(tablename="cars_x_persons") # note how co_owners is now ManyToMany relation class Car2(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=50) # note the related_name - needs to be unique across Person # model, regardless of how many different models leads to Person owner: Person = ormar.ForeignKey(Person, related_name="owned") co_owners: List[Person] = ormar.ManyToMany( Person, through=PersonsCar, related_name="coowned" ) created_date: datetime.datetime = ormar.DateTime(default=datetime.datetime.now) # child models define only additional Fields class Truck2(Car2): ormar_config = base_ormar_config.copy(tablename="trucks2") max_capacity: int = ormar.Integer() class Bus2(Car2): ormar_config = base_ormar_config.copy(tablename="buses2") max_persons: int = ormar.Integer() ``` `Ormar` automatically modifies related_name of the fields to include the **table** name of the children models. The default name is original related_name + '_' + child table name. That way for class Truck2 the relation defined in `owner: Person = ormar.ForeignKey(Person, related_name="owned")` becomes `owned_trucks2` You can verify the names by inspecting the list of fields present on `Person` model. ```python Person.ormar_config.model_fields { # note how all relation fields need to be unique on Person # regardless if autogenerated or manually overwritten 'id': , 'name': , # note that we expanded on previous example so all 'old' fields are here 'trucks': , 'coowned_trucks': , 'buses': , 'coowned_buses': , # newly defined related fields 'owned_trucks2': , 'coowned_trucks2': , 'owned_buses2': , 'coowned_buses2': } ``` But that's not all. It's kind of internal to `ormar` but affects the data structure in the database, so let's examine the through models for both `Bus2` and `Truck2` models. ```python Bus2.ormar_config.model_fields['co_owners'].through Bus2.ormar_config.model_fields['co_owners'].through.ormar_config.tablename 'cars_x_persons_buses2' Truck2.ormar_config.model_fields['co_owners'].through Truck2.ormar_config.model_fields['co_owners'].through.ormar_config.tablename 'cars_x_persons_trucks2' ``` As you can see above `ormar` cloned the Through model for each of the Child classes and added Child **class** name at the end, while changing the table names of the cloned fields the name of the **table** from the child is used. Note that original model is not only not used, the table for this model is removed from metadata: ```python Bus2.ormar_config.metadata.tables.keys() dict_keys(['test_date_models', 'categories', 'subjects', 'persons', 'trucks', 'buses', 'cars_x_persons_trucks2', 'trucks2', 'cars_x_persons_buses2', 'buses2']) ``` So be aware that if you introduce inheritance along the way and convert a model into abstract parent model you may lose your data on through table if not careful. !!!note Note that original table name and model name of the Through model is never used. Only the cloned models tables are created and used. !!!warning Note that each subclass of the Model that has `ManyToMany` relation defined generates a new `Through` model, meaning also **new database table**. That means that each time you define a Child model you need to either manually create the table in the database, or run a migration (with alembic). ## exclude_parent_fields Ormar allows you to skip certain fields in inherited model that are coming from a parent model. !!!Note Note that the same behaviour can be achieved by splitting the model into more abstract models and mixins - which is a preferred way in normal circumstances. To skip certain fields from a child model, list all fields that you want to skip in `model.ormar_config.exclude_parent_fields` parameter like follows: ```python base_ormar_config = OrmarConfig( metadata=sa.MetaData(), database=databases.Database(DATABASE_URL), ) class AuditModel(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) created_by: str = ormar.String(max_length=100) updated_by: str = ormar.String(max_length=100, default="Sam") class DateFieldsModel(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) created_date: datetime.datetime = ormar.DateTime( default=datetime.datetime.now, name="creation_date" ) updated_date: datetime.datetime = ormar.DateTime( default=datetime.datetime.now, name="modification_date" ) class Category(DateFieldsModel, AuditModel): ormar_config = base_ormar_config.copy( tablename="categories", # set fields that should be skipped exclude_parent_fields=["updated_by", "updated_date"], ) id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=50, unique=True, index=True) code: int = ormar.Integer() # Note that now the update fields in Category are gone in all places -> ormar fields, pydantic fields and sqlachemy table columns # so full list of available fields in Category is: ["created_by", "created_date", "id", "name", "code"] ``` Note how you simply need to provide field names and it will exclude the parent field regardless of from which parent model the field is coming from. !!!Note Note that if you want to overwrite a field in child model you do not have to exclude it, simply overwrite the field declaration in child model with same field name. !!!Warning Note that this kind of behavior can confuse mypy and static type checkers, yet accessing the non existing fields will fail at runtime. That's why splitting the base classes is preferred. The same effect can be achieved by splitting base classes like: ```python base_ormar_config = OrmarConfig( metadata=sa.MetaData(), database=databases.Database(DATABASE_URL), ) class AuditCreateModel(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) created_by: str = ormar.String(max_length=100) class AuditUpdateModel(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) updated_by: str = ormar.String(max_length=100, default="Sam") class CreateDateFieldsModel(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) created_date: datetime.datetime = ormar.DateTime( default=datetime.datetime.now, name="creation_date" ) class UpdateDateFieldsModel(ormar.Model): ormar_config = base_ormar_config.copy(abstract=True) updated_date: datetime.datetime = ormar.DateTime( default=datetime.datetime.now, name="modification_date" ) class Category(CreateDateFieldsModel, AuditCreateModel): ormar_config = base_ormar_config.copy(tablename="categories") id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=50, unique=True, index=True) code: int = ormar.Integer() ``` That way you can inherit from both create and update classes if needed, and only one of them otherwise.