There are also specific tips in each documentation section, and many of the classes, functions, and attributes.

SQLite is different

While SQLite provides a SQL database like many others out there, it is also unique in many ways. Read about the unique features at the SQLite website and quirks.


Using APSW best practice is recommended to get best performance and avoid common mistakes.


SQLite has 5 storage types.





Text (limit 1GB when encoded as bytes)


Integer (Signed 64 bit)


Float (IEEE754 64 bit)


BLOB (binary data, limit 1GB)

bytes and similar such as bytearray and array.array

Dates and times do not have a dedicated storage type, but do have a variety of functions for creating, manipulating, and storing them. JSON does not have a dedicated storage type, but does have a variety of functions for creating, manipulating, and storing JSON.

APSW provides optional type conversion, but the underlying storage will always be one of the 5 storage types.

If a column declaration gives a type then SQLite attempts conversion.

    create table types1(a, b, c, d, e);
    create table types2(a INTEGER, b REAL, c TEXT, d, e BLOB);

data = ("12", 3, 4, 5.5, b"\x03\x72\xf4\x00\x9e")
connection.execute("insert into types1 values(?,?,?,?,?)", data)
connection.execute("insert into types2 values(?,?,?,?,?)", data)

for row in connection.execute("select * from types1"):
    print("types1", repr(row))

for row in connection.execute("select * from types2"):
    print("types2", repr(row))
types1 ('12', 3, 4, 5.5, b'\x03r\xf4\x00\x9e')
types2 (12, 3.0, '4', 5.5, b'\x03r\xf4\x00\x9e')


Transactions are the changes applied to a database file as a whole. They either happen completely, or not at all. SQLite notes all the changes made during a transaction, and at the end when you commit will cause them to permanently end up in the database. If you do not commit, or just exit, then other/new connections will not see the changes and SQLite handles tidying up the work in progress automatically.

Committing a transaction can be quite time consuming. SQLite uses a robust multi-step process that has to handle errors that can occur at any point, and asks the operating system to ensure that data is on storage and would survive a power cycle. This will limit the rate at which you can do transactions.

If you do nothing, then each statement is a single transaction:

# this will be 3 separate transactions
db.execute("INSERT ...")
db.execute("INSERT ...")
db.execute("INSERT ...")

You can use BEGIN/COMMIT to set the transaction boundary:

# this will be one transaction
db.execute("INSERT ...")
db.execute("INSERT ...")
db.execute("INSERT ...")

However that is extra effort, and also requires error handling. For example if the second INSERT failed then you likely want to ROLLBACK the incomplete transaction, so that additional work on the same connection doesn’t see the partial data.

If you use with Connection then the transaction will be automatically started, and committed on success or rolled back if exceptions occur:

# this will be one transaction with automatic commit and rollback
with db:
    db.execute("INSERT ...")
    db.execute("INSERT ...")
    db.execute("INSERT ...")

There are technical details at the SQLite site.


SQLite only calculates each result row as you request it. For example if your query returns 10 million rows, SQLite will not calculate all 10 million up front. Instead the next row will be calculated as you ask for it. You can use Cursor.fetchall() to get all the results.

Cursors on the same Connection are not isolated from each other. Anything done on one cursor is immediately visible to all other cursors on the same connection. This still applies if you start transactions. Connections are isolated from each other.

Connection.execute() and Connection.executemany() automatically obtain cursors from Connection.cursor() which are very cheap. It is best practise to not re-use them, and instead get a new one each time. If you don’t, code refactoring and nested loops can unintentionally use the same cursor object which will not crash but will cause hard to diagnose behaviour in your program.


When issuing a query, always use bindings. String interpolation may seem more convenient but you will encounter difficulties. You may feel that you have complete control over all data accessed but if your code is at all useful then you will find it being used more and more widely. The computer will always be better than you at parsing SQL and the bad guys have years of experience finding and using SQL injection attacks in ways you never even thought possible.

The tour shows why you use bindings, and the different ways you can supply them.


Both SQLite and APSW provide detailed diagnostic information. Errors will be signalled via an exception.

APSW ensures you have detailed information both in the stack trace as well as what data APSW/SQLite was operating on.

SQLite has a warning/error logging facility. Use best practice to forward SQLite log messages to Python’s logging.

Managing and updating your schema

If your program uses SQLite for data then you’ll need to manage and update your schema. The hard way of doing this is to test for the existence of tables and their columns, and doing that maintenance programmatically. The easy way is to use pragma user_version as in this example:

def ensure_schema(db):
  # a new database starts at user_version 0
  if db.pragma("user_version") == 0:
    with db:
        CREATE TABLE foo(x,y,z);
        CREATE TABLE bar(x,y,z);
        PRAGMA user_version = 1;""")

  if db.pragma("user_version") == 1:
    with db:
      CREATE TABLE baz(x,y,z);
      CREATE INDEX ....
      PRAGMA user_version = 2;""")

  if db.pragma("user_version") == 2:
    with db:
      ALTER TABLE .....
      PRAGMA user_version = 3;""")

This approach will automatically upgrade the schema as you expect. You can also use pragma application_id to mark the database as made by your application.

Parsing SQL

Sometimes you want to know what a particular SQL statement does. Use apsw.ext.query_info() which will provide as much detail as you need.

Busy handling

SQLite uses locks to coordinate access to the database by multiple connections (within the same process or in a different process). The general goal is to have the locks be as lax as possible (allowing concurrency) and when using more restrictive locks to keep them for as short a time as possible. See the SQLite documentation for more details.

By default you will get an immediate BusyError if a lock cannot be acquired. Use best practice which sets a short waiting period, as well as enabling WAL which reduces contention between readers and writers.

Database schema

When starting a new database, it can be quite difficult to decide what tables and column to have and how to link them. The technique used to design SQL schemas is called normalization. The page also shows common pitfalls if you do not normalize your schema.

Write Ahead Logging

SQLite has write ahead logging which has several benefits, but also some drawbacks as the page documents. WAL mode is off by default. Use best practice to automatically enable it for all connections.

Note that if wal mode can’t be set (eg the database is in memory or temporary) then the attempt to set wal mode will be ignored. It is also harmless to call functions like Connection.wal_autocheckpoint() on connections that are not in wal mode.

If you write your own VFS, then inheriting from an existing VFS that supports WAL will make your VFS support the extra WAL methods too.

Customizing Connections

apsw.connection_hooks is a list of callbacks for when each Connection is created. They are called in turn, with the new connection as the only parameter.

For example if you wanted to add an executescript method to Connections that is like Connection.execute() but ignores all returned rows:

def executescript(self, sql, bindings=None):
  for _ in self.execute(sql, bindings):

def my_hook(connection):
  connection.executescript = executescript


Customizing Cursors

You can customize the behaviour of cursors. An example would be wanting a rowcount or batching returned rows. (These don’t make any sense with SQLite but the desire may be to make the code source compatible with other database drivers).

Set Connection.cursor_factory to any callable, which will be called with the connection as the only parameter, and return the object to use as a cursor.