Working with Transactions and the DBAPI

With the Engine object ready to go, we can dive into the basic operation of an Engine and its primary endpoints, the Connection and Result. We’ll also introduce the ORM’s facade for these objects, known as the Session.

Note to ORM readers

When using the ORM, the Engine is managed by the Session. The Session in modern SQLAlchemy emphasizes a transactional and SQL execution pattern that is largely identical to that of the Connection discussed below, so while this subsection is Core-centric, all of the concepts here are relevant to ORM use as well and is recommended for all ORM learners. The execution pattern used by the Connection will be compared to the Session at the end of this section.

As we have yet to introduce the SQLAlchemy Expression Language that is the primary feature of SQLAlchemy, we’ll use a simple construct within this package called the text() construct, to write SQL statements as textual SQL. Rest assured that textual SQL is the exception rather than the rule in day-to-day SQLAlchemy use, but it’s always available.

Getting a Connection

The purpose of the Engine is to connect to the database by providing a Connection object. When working with the Core directly, the Connection object is how all interaction with the database is done. Because the Connection creates an open resource against the database, we want to limit our use of this object to a specific context. The best way to do that is with a Python context manager, also known as the with statement. Below we use a textual SQL statement to show “Hello World”. Textual SQL is created with a construct called text() which we’ll discuss in more detail later:

>>> from sqlalchemy import text

>>> with engine.connect() as conn:
...     result = conn.execute(text("select 'hello world'"))
...     print(result.all())
BEGIN (implicit) select 'hello world' [...] ()
[('hello world',)]
ROLLBACK

In the example above, the context manager creates a database connection and executes the operation in a transaction. The default behavior of the Python DBAPI is that a transaction is always in progress; when the connection is released, a ROLLBACK is emitted to end the transaction. The transaction is not committed automatically; if we want to commit data we need to call Connection.commit() as we’ll see in the next section.

Tip

“autocommit” mode is available for special cases. The section Setting Transaction Isolation Levels including DBAPI Autocommit discusses this.

The result of our SELECT was returned in an object called Result that will be discussed later. For the moment we’ll add that it’s best to use this object within the “connect” block, and to not use it outside of the scope of our connection.

Committing Changes

We just learned that the DBAPI connection doesn’t commit automatically. What if we want to commit some data? We can change our example above to create a table, insert some data and then commit the transaction using the Connection.commit() method, inside the block where we have the Connection object:

# "commit as you go"
>>> with engine.connect() as conn:
...     conn.execute(text("CREATE TABLE some_table (x int, y int)"))
...     conn.execute(
...         text("INSERT INTO some_table (x, y) VALUES (:x, :y)"),
...         [{"x": 1, "y": 1}, {"x": 2, "y": 4}],
...     )
...     conn.commit()
BEGIN (implicit) CREATE TABLE some_table (x int, y int) [...] () <sqlalchemy.engine.cursor.CursorResult object at 0x...> INSERT INTO some_table (x, y) VALUES (?, ?) [...] ((1, 1), (2, 4)) <sqlalchemy.engine.cursor.CursorResult object at 0x...> COMMIT

Above, we execute two SQL statements, a “CREATE TABLE” statement [1] and an “INSERT” statement that’s parameterized (we discuss the parameterization syntax later in Sending Multiple Parameters). To commit the work we’ve done in our block, we call the Connection.commit() method which commits the transaction. After this, we can continue to run more SQL statements and call Connection.commit() again for those statements. SQLAlchemy refers to this style as commit as you go.

There’s also another style to commit data. We can declare our “connect” block to be a transaction block up front. To do this, we use the Engine.begin() method to get the connection, rather than the Engine.connect() method. This method will manage the scope of the Connection and also enclose everything inside of a transaction with either a COMMIT at the end if the block was successful, or a ROLLBACK if an exception was raised. This style is known as begin once:

# "begin once"
>>> with engine.begin() as conn:
...     conn.execute(
...         text("INSERT INTO some_table (x, y) VALUES (:x, :y)"),
...         [{"x": 6, "y": 8}, {"x": 9, "y": 10}],
...     )
BEGIN (implicit) INSERT INTO some_table (x, y) VALUES (?, ?) [...] ((6, 8), (9, 10)) <sqlalchemy.engine.cursor.CursorResult object at 0x...> COMMIT

You should mostly prefer the “begin once” style because it’s shorter and shows the intention of the entire block up front. However, in this tutorial we’ll use “commit as you go” style as it’s more flexible for demonstration purposes.

Basics of Statement Execution

We have seen a few examples that run SQL statements against a database, making use of a method called Connection.execute(), in conjunction with an object called text(), and returning an object called Result. In this section we’ll illustrate more closely the mechanics and interactions of these components.

Most of the content in this section applies equally well to modern ORM use when using the Session.execute() method, which works very similarly to that of Connection.execute(), including that ORM result rows are delivered using the same Result interface used by Core.

Fetching Rows

We’ll first illustrate the Result object more closely by making use of the rows we’ve inserted previously, running a textual SELECT statement on the table we’ve created:

>>> with engine.connect() as conn:
...     result = conn.execute(text("SELECT x, y FROM some_table"))
...     for row in result:
...         print(f"x: {row.x}  y: {row.y}")
BEGIN (implicit) SELECT x, y FROM some_table [...] ()
x: 1 y: 1 x: 2 y: 4 x: 6 y: 8 x: 9 y: 10
ROLLBACK

Above, the “SELECT” string we executed selected all rows from our table. The object returned is called Result and represents an iterable object of result rows.

Result has lots of methods for fetching and transforming rows, such as the Result.all() method illustrated previously, which returns a list of all Row objects. It also implements the Python iterator interface so that we can iterate over the collection of Row objects directly.

The Row objects themselves are intended to act like Python named tuples. Below we illustrate a variety of ways to access rows.

  • Tuple Assignment - This is the most Python-idiomatic style, which is to assign variables to each row positionally as they are received:

    result = conn.execute(text("select x, y from some_table"))
    
    for x, y in result:
        # ...
  • Integer Index - Tuples are Python sequences, so regular integer access is available too:

    result = conn.execute(text("select x, y from some_table"))
    
      for row in result:
          x = row[0]
  • Attribute Name - As these are Python named tuples, the tuples have dynamic attribute names matching the names of each column. These names are normally the names that the SQL statement assigns to the columns in each row. While they are usually fairly predictable and can also be controlled by labels, in less defined cases they may be subject to database-specific behaviors:

    result = conn.execute(text("select x, y from some_table"))
    
    for row in result:
        y = row.y
    
        # illustrate use with Python f-strings
        print(f"Row: {row.x} {y}")
  • Mapping Access - To receive rows as Python mapping objects, which is essentially a read-only version of Python’s interface to the common dict object, the Result may be transformed into a MappingResult object using the Result.mappings() modifier; this is a result object that yields dictionary-like RowMapping objects rather than Row objects:

    result = conn.execute(text("select x, y from some_table"))
    
    for dict_row in result.mappings():
        x = dict_row["x"]
        y = dict_row["y"]

Sending Parameters

SQL statements are usually accompanied by data that is to be passed with the statement itself, as we saw in the INSERT example previously. The Connection.execute() method therefore also accepts parameters, which are referred towards as bound parameters. A rudimentary example might be if we wanted to limit our SELECT statement only to rows that meet a certain criteria, such as rows where the “y” value were greater than a certain value that is passed in to a function.

In order to achieve this such that the SQL statement can remain fixed and that the driver can properly sanitize the value, we add a WHERE criteria to our statement that names a new parameter called “y”; the text() construct accepts these using a colon format “:y”. The actual value for “:y” is then passed as the second argument to Connection.execute() in the form of a dictionary:

>>> with engine.connect() as conn:
...     result = conn.execute(text("SELECT x, y FROM some_table WHERE y > :y"), {"y": 2})
...     for row in result:
...         print(f"x: {row.x}  y: {row.y}")
BEGIN (implicit) SELECT x, y FROM some_table WHERE y > ? [...] (2,)
x: 2 y: 4 x: 6 y: 8 x: 9 y: 10
ROLLBACK

In the logged SQL output, we can see that the bound parameter :y was converted into a question mark when it was sent to the SQLite database. This is because the SQLite database driver uses a format called “qmark parameter style”, which is one of six different formats allowed by the DBAPI specification. SQLAlchemy abstracts these formats into just one, which is the “named” format using a colon.

Sending Multiple Parameters

In the example at Committing Changes, we executed an INSERT statement where it appeared that we were able to INSERT multiple rows into the database at once. For statements that operate upon data, but do not return result sets, namely DML statements such as “INSERT” which don’t include a phrase like “RETURNING”, we can send multi params to the Connection.execute() method by passing a list of dictionaries instead of a single dictionary, thus allowing the single SQL statement to be invoked against each parameter set individually:

>>> with engine.connect() as conn:
...     conn.execute(
...         text("INSERT INTO some_table (x, y) VALUES (:x, :y)"),
...         [{"x": 11, "y": 12}, {"x": 13, "y": 14}],
...     )
...     conn.commit()
BEGIN (implicit) INSERT INTO some_table (x, y) VALUES (?, ?) [...] ((11, 12), (13, 14)) <sqlalchemy.engine.cursor.CursorResult object at 0x...> COMMIT

Behind the scenes, the Connection objects uses a DBAPI feature known as cursor.executemany(). This method performs the equivalent operation of invoking the given SQL statement against each parameter set individually. The DBAPI may optimize this operation in a variety of ways, by using prepared statements, or by concatenating the parameter sets into a single SQL statement in some cases. Some SQLAlchemy dialects may also use alternate APIs for this case, such as the psycopg2 dialect for PostgreSQL which uses more performant APIs for this use case.

Tip

you may have noticed this section isn’t tagged as an ORM concept. That’s because the “multiple parameters” use case is usually used for INSERT statements, which when using the ORM are invoked in a different way. Multiple parameters also may be used with UPDATE and DELETE statements to emit distinct UPDATE/DELETE operations on a per-row basis, however again when using the ORM, there is a different technique generally used for updating or deleting many individual rows separately.

Executing with an ORM Session

As mentioned previously, most of the patterns and examples above apply to use with the ORM as well, so here we will introduce this usage so that as the tutorial proceeds, we will be able to illustrate each pattern in terms of Core and ORM use together.

The fundamental transactional / database interactive object when using the ORM is called the Session. In modern SQLAlchemy, this object is used in a manner very similar to that of the Connection, and in fact as the Session is used, it refers to a Connection internally which it uses to emit SQL.

When the Session is used with non-ORM constructs, it passes through the SQL statements we give it and does not generally do things much differently from how the Connection does directly, so we can illustrate it here in terms of the simple textual SQL operations we’ve already learned.

The Session has a few different creational patterns, but here we will illustrate the most basic one that tracks exactly with how the Connection is used which is to construct it within a context manager:

>>> from sqlalchemy.orm import Session

>>> stmt = text("SELECT x, y FROM some_table WHERE y > :y ORDER BY x, y")
>>> with Session(engine) as session:
...     result = session.execute(stmt, {"y": 6})
...     for row in result:
...         print(f"x: {row.x}  y: {row.y}")
BEGIN (implicit) SELECT x, y FROM some_table WHERE y > ? ORDER BY x, y [...] (6,)
x: 6 y: 8 x: 9 y: 10 x: 11 y: 12 x: 13 y: 14
ROLLBACK

The example above can be compared to the example in the preceding section in Sending Parameters - we directly replace the call to with engine.connect() as conn with with Session(engine) as session, and then make use of the Session.execute() method just like we do with the Connection.execute() method.

Also, like the Connection, the Session features “commit as you go” behavior using the Session.commit() method, illustrated below using a textual UPDATE statement to alter some of our data:

>>> with Session(engine) as session:
...     result = session.execute(
...         text("UPDATE some_table SET y=:y WHERE x=:x"),
...         [{"x": 9, "y": 11}, {"x": 13, "y": 15}],
...     )
...     session.commit()
BEGIN (implicit) UPDATE some_table SET y=? WHERE x=? [...] ((11, 9), (15, 13)) COMMIT

Above, we invoked an UPDATE statement using the bound-parameter, “executemany” style of execution introduced at Sending Multiple Parameters, ending the block with a “commit as you go” commit.

Tip

The Session doesn’t actually hold onto the Connection object after it ends the transaction. It gets a new Connection from the Engine the next time it needs to execute SQL against the database.

The Session obviously has a lot more tricks up its sleeve than that, however understanding that it has a Session.execute() method that’s used the same way as Connection.execute() will get us started with the examples that follow later.

See also

Basics of Using a Session - presents basic creational and usage patterns with the Session object.