Ian Fisher, 9 October 2021
I use SQLite as the database for my personal projects in Python. It is lightweight, reliable, well-documented, and better than the filesystem for persistent storage. I'd like to share a few lessons I have learned on using SQLite effectively in Python.
The official documentation for Python's
sqlite3 module already has a section on "Using
sqlite3 efficiently". It's worth reading that first, as this post covers different topics.
Foreign key constraints are not enforced by default in SQLite. If you want the database to prevent you from inserting invalid foreign keys, then you must run
PRAGMA foreign_keys = 1 to turn enforcement on. Note that this pragma command must be run outside of a transaction; if you run it while a transaction is active, it will silently do nothing.
Since I prefer for my database to detect invalid foreign keys for me, and since (as we'll see below) the Python's
sqlite3 module will sometimes open transactions implicitly, I run
PRAGMA foreign_keys = 1 right after I open the connection to the database.
By default, the underlying SQLite library operates in
autocommit mode, in which changes are committed immediately unless a transaction has been opened with
SAVEPOINT. You can verify this by opening the same database file with the
sqlite3 command-line shell in two different terminals at the same time, and observing that, e.g., a row inserted in one terminal will be returned by a
SELECT statement run in the other. Once you open a transaction with
BEGIN, however, subsequent changes will not be visible to the other terminal until you commit the transaction with
sqlite3 module does not operate in
autocommit mode. Instead, it will start a transaction before data manipulation language (DML) statements1 such as
UPDATE, and, until Python 3.6, data definition language (DDL) statements such as
Opening a transaction in SQLite has several implications:
These consequences can come as a surprise when
sqlite3 has silently opened a transaction without your knowledge. Even worse, the
Connection.close method will not commit an open transaction, so you have to manually commit the transaction that
sqlite3 automatically opened.
I prefer to manage my transactions explicitly. To do so, pass
isolation_level=None as an argument to
sqlite3.connect, which will leave the database in the default
autocommit mode and allow you to issue
ROLLBACK statements yourself.
sqlite3 module allows you to register adapters to convert Python objects to SQLite values, and converters to convert SQLite values to Python objects (based on the type of the column).
sqlite3 automatically registers converters for
TIMESTAMP columns, and corresponding adapters for Python
datetime objects. Adapters are enabled by default, while converters must be explicitly enabled with the
detect_types parameter to
In addition to the default converters, I register my own for
TIME columns, to convert them to
datetime.time values, respectively.
TIMESTAMP converter ignores UTC offsets in the database row and always returns a naive datetime object. If your
TIMESTAMP rows contain UTC offsets, you can register your own converter to return aware datetime objects:2
import datetime import sqlite3 sqlite3.register_converter("TIMESTAMP", datetime.datetime.fromisoformat)
Keep in mind that it is generally considered better practice to store time zone information as a string identifier from the IANA time zone database in a separate column, rather than use UTC offsets, which change often (e.g., due to daylight saving time).
Adapters and converters are registered globally, not per-database. Be warned that some Python libraries, like Django, register their own adapters and converters which will apply even if you use the raw
sqlite3 interface instead of, e.g., Django's ORM.
SQLite lets you declare columns with any type that you want (or none at all). This can work nicely with Python's converters and adapters; for example, in one of my projects, I had columns of type
CSV and used a converter and an adapter to transparently convert them to Python lists and back.
Although SQLite is flexible with typing, ultimately it must choose a storage class for data, either
BLOB. Columns have a "type affinity" which determines the preferred storage class for a column through a somewhat arbitrary set of rules. This ensures that inserting a string into an
INT column will convert the string to an integer, for compatibility with other, rigidly-typed database engines.
A corollary of SQLite's flexible typing is that different values in the same column can have different type affinities:
In SQLite, the datatype of a value is associated with the value itself, not with its container.
This can cause problems. I once wanted to copy some rows from one table to another. My rows had
TIMESTAMP columns, and since, as we saw, Python will silently drop UTC offsets, I replaced Python's
TIMESTAMP converter with one that simply returns the bytes object unchanged:
sqlite3.register_converter("TIMESTAMP", lambda b: b)
Unfortunately, this converter resulted in the new
TIMESTAMP columns having
BLOB affinity instead of
TEXT. This was a problem, because some SQL operations are sensitive to the affinities of their operands. One of them is
LIKE, which does not work on blob values:
sqlite> SELECT 'a' LIKE 'a'; 1 sqlite> SELECT X'61'; -- 0x61 is the hexadecimal value of ASCII 'a' a sqlite> SELECT X'61' LIKE 'a'; 0
Consequently, the query
SELECT * FROM table WHERE date LIKE '2019%' did not return any of the inserted rows because they all had
BLOB affinity and the
LIKE comparison always failed. Only when I ran
SELECT typeof(date) FROM table did I discover that some of the values in the same column had different affinities.
The correct procedure would have been to register the converter as
lambda b: b.decode() so that Python would insert string values with
Because I use SQLite in Python so often, I wrote my own library, isqlite, that handles most of these issues for me, and also provides a more convenient Python API and many other useful features. You will be able to read about isqlite in next week's blog post. ∎
datetime.fromisoformat was added in Python 3.7, so if you are using an older version of Python you will have to write the converter function yourself. You can take a look at how the
sqlite3 module implements the naive datetime converter, and adapt it to also read the UTC offset if present. Or you can copy the implementation of
You might reasonably wonder why I had enabled converters in the first place if I knew that they were not going to work for my
TIMESTAMP columns. In this case, I was using a library that wrapped
sqlite3.connect and enabled converters for me. ↩