This page is part of the .

Previous: Working with Transactions and the DBAPI | Next:

Working with Database Metadata

With engines and SQL execution down, we are ready to begin some Alchemy. The central element of both SQLAlchemy Core and ORM is the SQL Expression Language which allows for fluent, composable construction of SQL queries. The foundation for these queries are Python objects that represent database concepts like tables and columns. These objects are known collectively as .

The most common foundational objects for database metadata in SQLAlchemy are known as MetaData, , and Column. The sections below will illustrate how these objects are used in both a Core-oriented style as well as an ORM-oriented style.

ORM readers, stay with us!

As with other sections, Core users can skip the ORM sections, but ORM users would best be familiar with these objects from both perspectives. The object discussed here is declared in a more indirect (and also fully Python-typed) way when using the ORM, however there is still a Table object within the ORM’s configuration.

When we work with a relational database, the basic data-holding structure in the database which we query from is known a table. In SQLAlchemy, the database “table” is ultimately represented by a Python object similarly named Table.

To start using the SQLAlchemy Expression Language, we will want to have objects constructed that represent all of the database tables we are interested in working with. The Table is constructed programmatically, either directly by using the constructor, or indirectly by using ORM Mapped classes (described later at Using ORM Declarative Forms to Define Table Metadata). There is also the option to load some or all table information from an existing database, called .

Whichever kind of approach is used, we always start out with a collection that will be where we place our tables known as the MetaData object. This object is essentially a around a Python dictionary that stores a series of Table objects keyed to their string name. While the ORM provides some options on where to get this collection, we always have the option to simply make one directly, which looks like:

Once we have a object, we can declare some Table objects. This tutorial will start with the classic SQLAlchemy tutorial model, which has a table called user_account that stores, for example, the users of a website, and a related table address, which stores email addresses associated with rows in the user_account table. When not using ORM Declarative models at all, we construct each object directly, typically assigning each to a variable that will be how we will refer to the table in application code:

>>> from sqlalchemy import Table, Column, Integer, String
>>> user_table = Table(
...     "user_account",
...     metadata_obj,
...     Column("id", Integer, primary_key=True),
...     Column("name", String(30)),
...     Column("fullname", String),
... )

With the above example, when we wish to write code that refers to the user_account table in the database, we will use the user_table Python variable to refer to it.

When do I make a MetaData object in my program?

Having a single MetaData object for an entire application is the most common case, represented as a module-level variable in a single place in an application, often in a “models” or “dbschema” type of package. It is also very common that the is accessed via an ORM-centric registry or base class, so that this same MetaData is shared among ORM- and Core-declared objects.

There can be multiple MetaData collections as well; objects can refer to Table objects in other collections without restrictions. However, for groups of objects that are related to each other, it is in practice much more straightforward to have them set up within a single MetaData collection, both from the perspective of declaring them, as well as from the perspective of DDL (i.e. CREATE and DROP) statements being emitted in the correct order.

We can observe that the Table construct as written in Python has a resemblence to a SQL CREATE TABLE statement; starting with the table name, then listing out each column, where each column has a name and a datatype. The objects we use above are:

- represents a database table and assigns itself to a MetaData collection.
- represents a column in a database table, and assigns itself to a Table object. The usually includes a string name and a type object. The collection of Column objects in terms of the parent are typically accessed via an associative array located at Table.c:
```
>>> user_table.c.name
Column('name', String(length=30), table=<user_account>)
>>> user_table.c.keys()
['id', 'name', 'fullname']
```
, String - these classes represent SQL datatypes and can be passed to a with or without necessarily being instantiated. Above, we want to give a length of “30” to the “name” column, so we instantiated String(30). But for “id” and “fullname” we did not specify these, so we can send the class itself.

Declaring Simple Constraints

The first in the example user_table includes the Column.primary_key parameter which is a shorthand technique of indicating that this should be part of the primary key for this table. The primary key itself is normally declared implicitly and is represented by the PrimaryKeyConstraint construct, which we can see on the attribute on the Table object:

>>> user_table.primary_key
PrimaryKeyConstraint(Column('id', Integer(), table=<user_account>, primary_key=True, nullable=False))

The constraint that is most typically declared explicitly is the object that corresponds to a database foreign key constraint. When we declare tables that are related to each other, SQLAlchemy uses the presence of these foreign key constraint declarations not only so that they are emitted within CREATE statements to the database, but also to assist in constructing SQL expressions.

A that involves only a single column on the target table is typically declared using a column-level shorthand notation via the ForeignKey object. Below we declare a second table address that will have a foreign key constraint referring to the user table:

>>> from sqlalchemy import ForeignKey
>>> address_table = Table(
...     "address",
...     metadata_obj,
...     Column("id", Integer, primary_key=True),
...     Column("user_id", ForeignKey("user_account.id"), nullable=False),
...     Column("email_address", String, nullable=False),
... )

The table above also features a third kind of constraint, which in SQL is the “NOT NULL” constraint, indicated above using the parameter.

Tip

When using the ForeignKey object within a definition, we can omit the datatype for that Column; it is automatically inferred from that of the related column, in the above example the datatype of the user_account.id column.

We’ve constructed a an object structure that represents two database tables in a database, starting at the root object, then into two Table objects, each of which hold onto a collection of and Constraint objects. This object structure will be at the center of most operations we perform with both Core and ORM going forward.

The first useful thing we can do with this structure will be to emit CREATE TABLE statements, or , to our SQLite database so that we can insert and query data from them. We have already all the tools needed to do so, by invoking the MetaData.create_all() method on our , sending it the Engine that refers to the target database:

The DDL create process above includes some SQLite-specific PRAGMA statements that test for the existence of each table before emitting a CREATE. The full series of steps are also included within a BEGIN/COMMIT pair to accommodate for transactional DDL.

The create process also takes care of emitting CREATE statements in the correct order; above, the FOREIGN KEY constraint is dependent on the user table existing, so the address table is created second. In more complicated dependency scenarios the FOREIGN KEY constraints may also be applied to tables after the fact using ALTER.

The object also features a MetaData.drop_all() method that will emit DROP statements in the reverse order as it would emit CREATE in order to drop schema elements.

Migration tools are usually appropriate

Overall, the CREATE / DROP feature of is useful for test suites, small and/or new applications, and applications that use short-lived databases. For management of an application database schema over the long term however, a schema management tool such as Alembic, which builds upon SQLAlchemy, is likely a better choice, as it can manage and orchestrate the process of incrementally altering a fixed database schema over time as the design of the application changes.

Another way to make Table objects?

The preceding examples illustrated direct use of the Table object, which underlies how SQLAlchemy ultimately refers to database tables when constructing SQL expressions. As mentioned, the SQLAlchemy ORM provides for a facade around the declaration process referred towards as Declarative Table. The Declarative Table process accomplishes the same goal as we had in the previous section, that of building Table objects, but also within that process gives us something else called an , or just “mapped class”. The mapped class is the most common foundational unit of SQL when using the ORM, and in modern SQLAlchemy can also be used quite effectively with Core-centric use as well.

Some benefits of using Declarative Table include:

A more succinct and Pythonic style of setting up column definitions, where Python types may be used to represent SQL types to be used in the database
The resulting mapped class can be used to form SQL expressions that in many cases maintain PEP 484 typing information that’s picked up by static analysis tools such as Mypy and IDE type checkers

This section will illustrate the same metadata of the previous section(s) being constructed using Declarative Table.

When using the ORM, the process by which we declare Table metadata is usually combined with the process of declaring classes. The mapped class is any Python class we’d like to create, which will then have attributes on it that will be linked to the columns in a database table. While there are a few varieties of how this is achieved, the most common style is known as declarative, and allows us to declare our user-defined classes and metadata at once.

Establishing a Declarative Base

When using the ORM, the collection remains present, however it itself is associated with an ORM-only construct commonly referred towards as the Declarative Base. The most expedient way to acquire a new Declarative Base is to create a new class that subclasses the SQLAlchemy DeclarativeBase class:

>>> from sqlalchemy.orm import DeclarativeBase
>>> class Base(DeclarativeBase):
...     pass

Above, the Base class is what we’ll refer towards as the Declarative Base. When we make new classes that are subclasses of Base, combined with appropriate class-level directives, they will each be established as a new ORM mapped class at class creation time, each one typically (but not exclusively) referring to a particular object.

The Declarative Base refers to a MetaData collection that is created for us automatically, assuming we didn’t provide one from the outside. This collection is accessible via the DeclarativeBase.metadata class-level attribute. As we create new mapped classes, they each will reference a within this MetaData collection:

>>> Base.metadata

The Declarative Base also refers to a collection called , which is the central “mapper configuration” unit in the SQLAlchemy ORM. While seldom accessed directly, this object is central to the mapper configuration process, as a set of ORM mapped classes will coordinate with each other via this registry. As was the case with MetaData, our Declarative Base also created a for us (again with options to pass our own registry), which we can access via the class variable:

>>> Base.registry
<sqlalchemy.orm.decl_api.registry object at 0x...>

Other ways to map with the registry

DeclarativeBase is not the only way to map classes, only the most common. also provides other mapper configurational patterns, including decorator-oriented and imperative ways to map classes. There’s also full support for creating Python dataclasses while mapping. The reference documentation at ORM Mapped Class Configuration has it all.

With the Base class established, we can now define ORM mapped classes for the user_account and address tables in terms of new classes User and Address. We illustrate below the most modern form of Declarative, which is driven from PEP 484 type annotations using a special type , which indicates attributes to be mapped as particular types:

>>> from typing import List
>>> from typing import Optional
>>> from sqlalchemy.orm import Mapped
>>> from sqlalchemy.orm import mapped_column
>>> from sqlalchemy.orm import relationship
>>> class User(Base):
...     __tablename__ = "user_account"
...
...     id: Mapped[int] = mapped_column(primary_key=True)
...     name: Mapped[str] = mapped_column(String(30))
...     fullname: Mapped[Optional[str]]
...
...     addresses: Mapped[List["Address"]] = relationship(back_populates="user")
...
...     def __repr__(self) -> str:
...         return f"User(id={self.id!r}, name={self.name!r}, fullname={self.fullname!r})"
>>> class Address(Base):
...     __tablename__ = "address"
...
...     id: Mapped[int] = mapped_column(primary_key=True)
...     email_address: Mapped[str]
...     user_id = mapped_column(ForeignKey("user_account.id"))
...
...     user: Mapped[User] = relationship(back_populates="addresses")
...
...     def __repr__(self) -> str:
...         return f"Address(id={self.id!r}, email_address={self.email_address!r})"

The two classes above, User and Address, are now referred towards as ORM Mapped Classes, and are available for use in ORM persistence and query operations, which will be described later. Details about these classes include:

Each class refers to a Table object that was generated as part of the declarative mapping process, which is named by assigning a string to the attribute. Once the class is created, this generated Table is available from the attribute.
As mentioned previously, this form is referred towards as Declarative Table Configuration. One of several alternative declaration styles would instead have us build the object directly, and assign it directly to DeclarativeBase.__table__. This style is known as .
For columns with simple datatypes and no other options, we can indicate a Mapped type annotation alone, using simple Python types like int and str to mean and String. Customization of how Python types are interpreted within the Declarative mapping process is very open ended; see the sections and Customizing the Type Map for background.
Use of explicit typing annotations is completely optional. We can also use without annotations. When using this form, we would use more explicit type objects like Integer and as well as nullable=False as needed within each mapped_column() construct.
Two additional attributes, and Address.user, define a different kind of attribute called , which features similar annotation-aware configuration styles as shown. The relationship() construct is discussed more fully at .
The classes are automatically given an __init__() method if we don’t declare one of our own. The default form of this method accepts all attribute names as optional keyword arguments:

To automatically generate a full-featured __init__() method which provides for positional arguments as well as arguments with default keyword values, the dataclasses feature introduced at Declarative Dataclass Mapping may be used. It’s of course always an option to use an explicit __init__() method as well.
The __repr__() methods are added so that we get a readable string output; there’s no requirement for these methods to be here. As is the case with __init__(), a __repr__() method can be generated automatically by using the feature.

Where’d the old Declarative go?

Users of SQLAlchemy 1.4 or previous will note that the above mapping uses a dramatically different form than before; not only does it use mapped_column() instead of in the Declarative mapping, it also uses Python type annotations to derive column information.

To provide context for users of the “old” way, Declarative mappings can still be made using Column objects (as well as using the function to create the base class) as before, and these forms will continue to be supported with no plans to remove support. The reason these two facilities are superseded by new constructs is first and foremost to integrate smoothly with PEP 484 tools, including IDEs such as VSCode and type checkers such as Mypy and Pyright, without the need for plugins. Secondly, deriving the declarations from type annotations is part of SQLAlchemy’s integration with Python dataclasses, which can now be from mappings.

For users who like the “old” way, but still desire their IDEs to not mistakenly report typing errors for their declarative mappings, the mapped_column() construct is a drop-in replacement for in an ORM Declarative mapping (note that mapped_column() is for ORM Declarative mappings only; it can’t be used within a construct), and the type annotations are optional. Our mapping above can be written without annotations as:

class User(Base):
    __tablename__ = "user_account"
    id = mapped_column(Integer, primary_key=True)
    name = mapped_column(String(30), nullable=False)
    fullname = mapped_column(String)
    addresses = relationship("Address", back_populates="user")
    # ... definition continues

The above class has an advantage over one that uses Column directly, in that the User class as well as instances of User will indicate the correct typing information to typing tools, without the use of plugins. also allows for additional ORM-specific parameters to configure behaviors such as deferred column loading, which previously needed a separate deferred() function to be used with .

There’s also an example of converting an old-style Declarative class to the new style, which can be seen at ORM Declarative Models in the guide.

Emitting DDL to the database from an ORM mapping

As our ORM mapped classes refer to objects contained within a MetaData collection, emitting DDL given the Declarative Base uses the same process as that described previously at . In our case, we have already generated the user and address tables in our SQLite database. If we had not done so already, we would be free to make use of the MetaData associated with our ORM Declarative Base class in order to do so, by accessing the collection from the attribute and then using MetaData.create_all() as before. In this case, PRAGMA statements are run, but no new tables are generated since they are found to be present already:

>>> Base.metadata.create_all(engine)
BEGIN (implicit)
PRAGMA main.table_...info("user_account")
...
PRAGMA main.table_...info("address")
...
COMMIT

Optional Section

This section is just a brief introduction to the related subject of table reflection, or how to generate Table objects automatically from an existing database. Tutorial readers who want to get on with writing queries can feel free to skip this section.

To round out the section on working with table metadata, we will illustrate another operation that was mentioned at the beginning of the section, that of table reflection. Table reflection refers to the process of generating and related objects by reading the current state of a database. Whereas in the previous sections we’ve been declaring Table objects in Python and then emitting DDL to the database, the reflection process does it in reverse.

As an example of reflection, we will create a new object which represents the some_table object we created manually in the earlier sections of this document. There are again some varieties of how this is performed, however the most basic is to construct a Table object, given the name of the table and a collection to which it will belong, then instead of indicating individual Column and objects, pass it the target Engine using the parameter:

>>> some_table = Table("some_table", metadata_obj, autoload_with=engine)
BEGIN (implicit)
PRAGMA main.table_...info("some_table")
[raw sql] ()
SELECT sql FROM  (SELECT * FROM sqlite_master UNION ALL   SELECT * FROM sqlite_temp_master) WHERE name = ? AND type in ('table', 'view')
[raw sql] ('some_table',)
PRAGMA main.foreign_key_list("some_table")
...
PRAGMA main.index_list("some_table")
...
ROLLBACK

At the end of the process, the some_table object now contains the information about the Column objects present in the table, and the object is usable in exactly the same way as a that we declared explicitly:

>>> some_table
Table('some_table', MetaData(),
    Column('x', INTEGER(), table=<some_table>),
    Column('y', INTEGER(), table=<some_table>),