PRAGMA table_info sets is_pk to the 1-based position of each column
within the PRIMARY KEY, which can differ from table column order.
table.pks previously returned table column order, so an implicit
compound FOREIGN KEY ... REFERENCES other was introspected with its
referenced columns inverted, and transform() baked that inverted order
into the rewritten schema - failing with IntegrityError on valid data,
or silently reversing the constraint with foreign keys off.
table.pks, compound foreign key guessing (create, add_foreign_key) and
transform() now all use declaration order, and transform() no longer
reorders a compound PRIMARY KEY (b, a) into table column order.
Refs https://github.com/simonw/sqlite-utils/issues/769#issuecomment-4900034150
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
db.query() promises that a statement rejected with ValueError is rolled
back and has no effect. PRAGMA statements are the exception: some of them
refuse to run inside a transaction, so they execute outside the savepoint
guard - a row-less PRAGMA such as "PRAGMA user_version = 5" therefore
takes effect despite the ValueError. Documenting this as a known
limitation rather than fixing it, since a fix would need a hardcoded list
of row-returning PRAGMAs.
Refs https://github.com/simonw/sqlite-utils/issues/769#issuecomment-4900034150
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- table.extract() and the sqlite-utils extract command now skip rows
where every extracted column is null: the new foreign key column is
left null and no all-null record is added to the lookup table. Rows
with at least one non-null extracted column are extracted as before.
- The extracts= option to insert() and friends keeps None values as
null instead of creating a lookup record for them - previously each
insert batch added a duplicate null row to the lookup table.
- table.lookup() compares values using IS rather than = so lookup
values containing None match existing rows correctly, instead of
inserting a duplicate row on every call.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Fix transform() corrupting TRUE/FALSE/NULL column defaults
quote_default_value() passed keyword-literal defaults (TRUE, FALSE, NULL)
through self.quote(), wrapping them in quotes. So transform() rebuilt a
column declared "INTEGER DEFAULT TRUE" as "INTEGER DEFAULT 'TRUE'", and a
later default insert stored the text 'TRUE' instead of the integer 1
(likewise 'FALSE' for 0 and 'NULL' for null) - silent value corruption.
Return these keyword literals unquoted, as already done for the
CURRENT_TIME/DATE/TIMESTAMP literals.
PR #764
Column names passed to Python API methods are now resolved against the
table schema case-insensitively, mirroring how SQLite itself compares
identifiers (ASCII-only case folding). Fixes KeyError populating
last_pk from insert()/upsert(), silently ignored transform()
rename/drop/types options, duplicate column errors from
create_table(transform=True), redundant lookup() indexes, and
case-sensitive foreign key validation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Queries returning duplicate column names - e.g. joins between tables
sharing column names - silently lost values because rows were built
with dict(zip(keys, row)), where the last duplicate wins.
Later occurrences are now renamed with a numeric suffix: id, id
becomes id, id_2 - skipping any suffix that would collide with a
real column in the same query.
The new utils.dedupe_keys() helper transforms the key list once per
query, so the per-row dict construction is unchanged and there is no
measurable performance impact.
Applied in Database.query() (including the PRAGMA and RETURNING
paths), Table.rows_where(), Table.search() and the CLI's JSON output.
CSV, TSV and table output keep the original duplicate headers.
Closes#624
table.add_foreign_key() now accepts on_delete= and on_update= to
create foreign keys with ON DELETE/ON UPDATE actions:
table.add_foreign_key("author_id", "authors", "id", on_delete="CASCADE")
Works for compound foreign keys too.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ForeignKey objects passed to db.add_foreign_keys() were flattened to
plain (table, column, other_table, other_column) tuples before being
handed to transform(), silently discarding their on_delete and
on_update actions. The method now carries ForeignKey objects through
to transform() intact - tuple inputs are converted to ForeignKey
objects up front, which also simplifies the validation loop.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The fk_should_be_dropped and fk_with_renamed_columns closures captured
the drop and rename parameters, but type checkers do not narrow
captured variables inside nested functions so ty saw their Optional
declared types. Bind the already-normalized values to fresh names for
the closures to use.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ForeignKey.columns and .other_columns are now tuples rather than lists,
both when introspected and when constructed directly (lists are
normalized in __post_init__). Tuples are now the documented form for
specifying compound foreign keys everywhere - in create_table()
foreign_keys=, add_foreign_key() and drop_foreign_keys=:
foreign_keys=[(("campus_name", "dept_code"), "departments")]
Lists continue to work as input but are no longer documented.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New ForeignKeyColumns type alias - a column name or a list/tuple of
column names - used by add_foreign_key(), add_foreign_keys() (via the
ForeignKeyTuple alias, replacing its Any slots) and the
ForeignKeyIndicator union, which it also simplifies. Tuples now
type-check anywhere lists are accepted.
ForeignKey.__post_init__ normalizes tuple columns/other_columns to
lists so direct construction with tuples compares equal to
introspected foreign keys.
Refs https://github.com/simonw/sqlite-utils/pull/770/changes#r3525703477
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Changelog entries for all the compound foreign key work
- Upgrading guide notes the transform() behavior changes
- ForeignKey added to the API reference
- Updated .foreign_keys introspection example output
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
For foreign keys declared as "REFERENCES other_table" with no explicit
columns, PRAGMA foreign_key_list returns None for the referenced
column. The foreign_keys property now resolves these to the other
table's primary key columns, so other_column=None unambiguously
indicates a compound foreign key.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ForeignKey gains on_delete and on_update fields (default "NO ACTION"),
populated from PRAGMA foreign_key_list. create_table_sql() renders the
corresponding clauses for both inline and compound table-level foreign
keys, which means table.transform() now preserves actions such as
ON DELETE CASCADE - previously they were silently stripped whenever a
table was transformed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
table.add_foreign_key() and db.add_foreign_keys() now accept lists of
column names to create compound foreign keys:
table.add_foreign_key(
["campus_name", "dept_code"], "departments"
)
Omitting the other columns uses the compound primary key of the other
table. Duplicate detection compares the full column lists.
db.index_foreign_keys() now creates a single composite index across the
columns of a compound foreign key, rather than an index per column.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Compound foreign keys are passed through table.transform() intact
instead of being degraded to per-column single foreign keys:
- rename= applies to each member column of a compound key
- drop= of any member column drops the whole constraint, matching
the existing single-column behavior
- drop_foreign_keys= accepts a bare column name (drops any foreign
key that column participates in) or a tuple of columns (drops the
compound key with exactly those columns)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
create_table() and friends now accept compound foreign keys, specified
as lists of column names in the existing tuple forms:
foreign_keys=[
(["campus_name", "dept_code"], "departments"),
(["campus_name", "dept_code"], "departments", ["campus_name", "dept_code"]),
]
The two-item form guesses the compound primary key of the other table.
Compound keys are rendered as table-level FOREIGN KEY constraints,
while single-column keys keep their inline REFERENCES clauses.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
sorted() on a foreign_keys list containing both compound (column=None)
and single-column ForeignKeys raised TypeError because dataclass
ordering compared None against str. column/other_column are now
excluded from comparison - ordering and equality use the always
populated columns/other_columns lists instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
query() and execute() classified statements using
sql.lstrip().upper().startswith(...), which a leading SQL comment
defeated: db.query('/* c */ COMMIT') inside db.atomic() committed
the caller's transaction and masked the ValueError with a 'no such
savepoint' error, comment-prefixed PRAGMAs ran inside the savepoint
guard where journal mode changes are refused, and a comment-prefixed
BEGIN passed to db.execute() was instantly auto-committed.
The new _first_keyword() helper skips leading whitespace and -- or
/* */ comments - the only things SQLite's tokenizer allows before
the first token - and returns that token for exact comparison, so
keyword detection now matches what SQLite itself will see.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
db.query() rejected statements roll back, RETURNING commits immediately.
Fixes two transaction bugs in db.query():
- A non-row-returning statement such as an UPDATE was executed and
auto-committed by execute() before the ValueError was raised, so the
write took effect despite being rejected. query() now runs the
statement inside a savepoint and rolls it back before raising, so a
rejected statement has no effect on the database - in every
connection mode, including autocommit=True.
- INSERT ... RETURNING only committed once the returned generator was
fully exhausted, so calling query() without iterating - or partially
iterating with next() - left the transaction open and the write could
be rolled back on close. The write is now completed and committed at
call time, as the documentation already promised. Plain SELECTs are
still fetched lazily.
Transaction control statements, VACUUM, ATTACH and DETACH never return
rows, so query() now rejects them without executing them. PRAGMA
statements skip the savepoint guard because some of them - such as
pragma journal_mode=wal - refuse to run inside a transaction.
Claude-Session: https://claude.ai/code/session_012U3iRfJoTZ5vd22cBSF2nJ
Raw write statements previously opened an implicit transaction that
stayed open until something committed it - the write was visible on
the same connection, making it look saved, but was silently rolled
back when the connection closed. execute() now commits any implicit
transaction it opens, so raw writes behave like every other write
in the library: committed as soon as they run, unless an explicit
transaction (db.begin() or db.atomic()) is open, in which case they
join it.
Row-returning writes such as INSERT ... RETURNING via db.query()
commit once their rows have been iterated. atomic(), commit() and
rollback() now issue literal COMMIT/ROLLBACK statements, which have
identical semantics in every sqlite3 connection mode. The CLI bulk
command uses db.atomic() instead of 'with db.conn:'.
This simplifies the transaction contract to: everything commits
immediately unless you explicitly opened a transaction. Docs
updated throughout; changelog and upgrading guide cover the
breaking change for code that relied on rolling back uncommitted
execute() writes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
Connections created with sqlite3.connect(autocommit=True) make
commit() a documented no-op, and autocommit=False connections are
permanently inside a transaction - in both modes every write made
by this library appeared to work in-process but was silently
discarded when the connection closed. Database() now raises
TransactionError for these connections instead of losing data.
Tests are skipped on Python versions before 3.12, where the
autocommit parameter does not exist.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
A dedicated sqlite_utils.db.TransactionError exception is clearer
to catch than a generic RuntimeError, and gives future
transaction-state guards a home.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
Manual transaction control previously required mixing raw
db.execute("begin") strings with methods on db.conn. These thin
wrappers give the documentation and callers a single consistent
idiom. Docs updated to use them throughout.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
The method existed only to raise NotImplementedError, since
full-text search is not supported for views, and it showed up in
the generated API reference as a documented View method. Calling
enable_fts() on a View now raises AttributeError like any other
missing method. The sqlite-utils enable-fts command uses db.table()
and shows a clean error when pointed at a view instead of a
traceback.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
User-facing argument validation in db.py previously used bare
assert statements, which vanish entirely under python -O and raise
AssertionError - an exception type callers should not have to
catch for input validation. Fifteen validation sites now raise
ValueError with the same messages. Internal invariants (an
unreachable branch and a post-update rowcount check) remain
asserts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
- insert/upsert into a name that is actually a view now shows a
clean error instead of a traceback
- db.view() on a name that is a table now says so in its error
- Database.__getitem__ docstring documents that views are returned
for view names; hash_id docstring corrected (it is a column name,
not a boolean)
- Registering two migrations with the same name in one set now
raises ValueError at registration time instead of executing both
and failing with IntegrityError at apply time
- build-system requires setuptools>=77, needed for the PEP 639
license expression
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
Changing the journal mode assigns conn.isolation_level, which
commits any open transaction as a side effect - silently breaking
the rollback guarantee of atomic() blocks and of user-managed
transactions. Both methods now raise RuntimeError if a transaction
is open. Calling them when the database is already in the requested
mode remains a no-op.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
A record missing a primary key value - or with None for one - can
never match an existing row, because a NULL primary key never
satisfies ON CONFLICT. Such records were previously inserted as
brand new rows (silently, for multi-record upsert_all calls) or
triggered a KeyError after the insert had already happened. Both
cases, including upserting an empty record, now raise
PrimaryKeyRequired before any SQL is executed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
Both ran their INSERT INTO fts(fts) statements via a bare execute()
with no commit, leaving the connection inside an open implicit
transaction - the FTS operation and all subsequent writes were then
silently rolled back when the connection closed. Both are now
wrapped in db.atomic(), matching delete_where() and the other
write operations.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
Previously query() was a generator function, so nothing - including
the SQL itself - ran until the result was first iterated. A write
statement passed to query() silently did nothing, and SQL errors
surfaced far from the call site. The SQL now executes as soon as
query() is called, while rows are still fetched lazily during
iteration.
Statements that return no rows now raise a ValueError directing
callers to execute() instead. As a side effect, statements like
INSERT ... RETURNING now work naturally with query().
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
Table.delete_where() ran its DELETE via a bare execute() with no
commit, unlike Table.delete() which wraps in db.atomic(). The
connection was left with an open implicit transaction, so the
deletion (and all subsequent writes, including later atomic()
blocks which switched to savepoint mode) was silently rolled back
when the connection closed.
Wrap the DELETE in db.atomic() and remove the now-unnecessary
atomic() wrapper from the delete_where documentation example.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UnLnhsH25Nnv7LHhekUfPd
* fix: defer plugin entrypoint loading until runtime
Fixessimonw/sqlite-utils#713
* Lazy-load plugins when listing them
Keep plugin entrypoints from loading at module import time, but preserve the existing behavior of get_plugins() by loading entrypoints the first time plugins are listed outside tests.
---------
Co-authored-by: Simon Willison <swillison@gmail.com>
* Add comprehensive type annotations
- mypy.ini: expanded configuration with module-specific settings
- hookspecs.py: type annotations for hook functions
- plugins.py: typed get_plugins() return value
- recipes.py: full type annotations for parsedate, parsedatetime, jsonsplit
- utils.py: extensive type annotations including Row type alias,
TypeTracker, ValueTracker, and all utility functions
- db.py: type annotations for Database methods (__exit__,
ensure_autocommit_off, tracer, register_function, etc.) and
Queryable class methods
- tests/test_docs.py: updated to match new signature display format
* Fix type errors caught by ty check
- Add type: ignore comments for external library type stub limitations
(csv.reader, click.progressbar, IOBase.name, Callable.__name__)
- Change Iterable to Sequence for SQL where_args parameters
- Use db.table() instead of db[name] for proper Table return type
- Fix rebuild_fts return type from None to Table
- Update test_tracer to expect fewer queries (optimization side effect)
* Fix mypy type errors
- Add type: ignore comments for runtime-valid patterns mypy can't verify
- Fix new_column_types annotation to Dict[str, Set[type]]
- Add type: ignore for Default sentinel values passed to create_table
* mypy skip tests directory
* Fix CI: exclude typing imports from recipe docs, skip mypy on tests
- Add Callable and Optional to exclusion list in _generate_convert_help()
- Regenerate docs/cli-reference.rst with cog
- Add [mypy-tests.*] ignore_errors = True to skip test type errors
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
* Fix type warning for pipe.stdout possibly being None
Add conditional check before calling .read() on pipe.stdout since
Popen can return None for stdout.
* Use ctx.meta instead of dynamic attribute for database cleanup
Click's Context.meta dictionary is the proper way to store arbitrary
data on the context object, avoiding type checker warnings about
dynamic attribute assignment.
* Add assert for tables.callback before calling
Click's callback attribute is typed as Optional[Callable], so add
assert to satisfy type checker that it's not None.
* Fix type errors in cli.py and db.py
- Add type annotation for Database.conn to fix context manager errors
- Convert exception objects to str() when raising ClickException
- Handle None return from find_spatialite() with proper error message
* Fix remaining type errors in cli.py
- Add typing import and type annotations for dict kwargs
- Use db.table() instead of db[] for extract command
- Fix missing str() conversion for exception
* Fix type errors in db.py
- Add type annotation for Database.conn
- Add type: ignore for optional sqlite_dump import
- Update execute/query parameter types to Sequence|Dict for sqlite3 compatibility
- Use getattr for fn.__name__ access to handle callables without __name__
- Handle None return from find_spatialite() with OSError
- Fix pk_values assignment to use local variable
* Add type: ignore for optional pysqlite3 and sqlean imports
These are alternative sqlite3 implementations that may not be installed.
* Fix type errors in tests and plugins
- Add type: ignore for monkey-patching Database.__init__ in conftest
- Fix CLI test to pass string "2" instead of integer to Click invoke
- Add type: ignore for optional sqlean import
- Fix add_geometry_column test to use "XY" instead of integer 2
- Add type: ignore for click.Context as context manager
- Add type: ignore for enable_fts test that intentionally omits argument
- Add type: ignore for sys._called_from_test dynamic attribute
- Fix rows_from_file test type error for intentional wrong argument
- Handle None from pm.get_hookcallers in plugins.py
* Use db.table() instead of db[] for Table-specific operations
Changes db[table] to db.table(table) in CLI commands where we know
we're working with tables, not views. This resolves most of the
Table | View disambiguation type warnings since db.table() returns
Table directly rather than Table | View.
* Fix remaining type warnings in sqlite_utils package
- Add assert for sniff_buffer not being None
- Handle cursor.fetchone() potentially returning None
- Use db.table() for counts_table and index_foreign_keys
- Add type: ignore for cursor union type in raw mode
* Ran Black
* Run ty in CI
* ty check sqlite_utils
* Skip running ty on Windows
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
