Litewriter

pip install litewriter

Source: https://github.com/Bobowski/litewriter

The index name is litewriter. The import is litewriter. The class is LiteWriter. It is SQLite only.

One SQLite file. One writer thread. Writes that are waiting share one COMMIT. Each OS thread reads on its own connection. A read does not wait on that writer. A watch reads again when a commit changes a column it reads.

from litewriter import LiteWriter

A query

A query is a statement. The constructor takes the clauses. col is a name. lit is a string literal. param is a parameter. sql(q) returns one line of SQL. query returns a list of tuples.

from litewriter import Select, col, param, sql

page = Select(
    col("m.id"),
    col("m.body"),
    ("as", ("lower", "u.name"), "author"),
    from_=col("messages").as_("m"),
    join=("left", col("users").as_("u"), col("m.author") == col("u.id")),
    where=col("m.room") == param("room"),
    order_by=col("m.at").desc(),
    limit=50,
)
print(sql(page))

mine = page.where(col("m.author") == param("me")).limit(20)
print(sql(mine))
print(sql(page))

listed = Select(
    "id",
    from_="t",
    where=[col("room") == param("room"), col("at") > 0],
)
print(sql(listed))

inner = Select("id", from_="t", where=col("ok") == 1)
print(sql(Select(col("s.id"), from_=inner.as_("s"))))

SELECT m.id, m.body, lower(u.name) AS author FROM messages AS m LEFT JOIN users AS u ON m.author = u.id WHERE m.room = :room ORDER BY m.at DESC LIMIT 50
SELECT m.id, m.body, lower(u.name) AS author FROM messages AS m LEFT JOIN users AS u ON m.author = u.id WHERE m.room = :room AND m.author = :me ORDER BY m.at DESC LIMIT 20
SELECT m.id, m.body, lower(u.name) AS author FROM messages AS m LEFT JOIN users AS u ON m.author = u.id WHERE m.room = :room ORDER BY m.at DESC LIMIT 50
SELECT id FROM t WHERE room = :room AND at > 0
SELECT s.id FROM (SELECT id FROM t WHERE ok = 1) AS s

The third line is the first line again. page.where(...) returns a new statement. page stays as it was. .where(a, b) joins those arguments with AND. A list is one argument. .where([a, b]) does not compile.

In a constructor, a list of conditions is AND. A list that starts with a name is a call. ["=", "room", ":room"] is room = :room. ("count", "*") is count(*). The string count(*) is not a name.

& and | are bitwise. An expression has no truth value. The words and and or raise TypeError. Put conditions in a list in the constructor. That list is AND.

from_ and with_ keep the underscore. from and with are Python keywords.

A Select is a subquery. from_=inner.as_("s") is FROM (SELECT ...) AS s. The same form works in a join, in where, and in the select list.

A tuple is a call. ("lower", "name") is lower(name). ("as", expr, "name") is the same alias as .as_("name"). ("lit", "open") is the same literal as lit("open").

Select has the select methods. Insert has the insert methods. Update has set, from_, and where. Delete has where. A union has order_by and limit. The constructor is the whole statement. A method adds one clause. A dict of clauses still renders.

Values

A value becomes SQL:

A call name that is not an operator is a function. LiteWriter does not keep a list of functions. json_extract, datetime, and a function that you register on the connection all work the same way.

A value from outside goes in as a parameter (:name), never as text.

Calls

The call name is not case sensitive.

Parentheses come out only where SQL needs them. ["and", ["or", a, b], c] is (a OR b) AND c.

Clauses

The key order in the dict does not matter. The output uses SQL order.

• select, group_by, order_by, returning, and columns are lists. A call sits in its own list: "select": [["count", "*"]].
• from is a table, ("as", "table", "t"), a table function (("json_each", ":ids")), or a subquery.
• join is (kind, source, on). A list of those is many joins. kind is inner, left, right, full, or cross. A cross join is ("cross", source).
• with is a dict of name to query.
• values is one row (a dict) or many rows. A list of lists needs columns.
• set and do_update_set are a dict of column to value.
• on_conflict is a list of columns. [] means any conflict. do_update needs those column names.

An unknown key raises WriterError. A key that does not belong on that statement raises WriterError. A bad name or a missing parameter does too. An insert may contain the select keys. That form is INSERT INTO t SELECT ....

The file

This program replaces example.sqlite3 in the current directory. async with starts the writer and closes it at the end. hz=0 commits as soon as the job arrives. The default is 60.

import asyncio
from pathlib import Path

from litewriter import Insert, LiteWriter, Select, Tx, col, param

page = Select(
    col("m.id"),
    col("m.body"),
    ("as", ("lower", "u.name"), "author"),
    from_=col("messages").as_("m"),
    join=("left", col("users").as_("u"), col("m.author") == col("u.id")),
    where=col("m.room") == param("room"),
    order_by=col("m.at").desc(),
    limit=50,
)
mine = page.where(col("m.author") == param("me")).limit(20)

SCHEMA = """
CREATE TABLE users (
    id INTEGER PRIMARY KEY,
    name TEXT NOT NULL
);
CREATE TABLE messages (
    id INTEGER PRIMARY KEY,
    room INTEGER NOT NULL,
    author INTEGER REFERENCES users (id),
    body TEXT NOT NULL,
    at INTEGER NOT NULL
);
"""


def add_user(tx: Tx, name: str) -> int:
    return tx.value(
        Insert("users", values={"name": ":name"}, returning="id"),
        name=name,
    )


def send(tx: Tx, room: int, body: str, author: int, at: int) -> int:
    return tx.value(
        Insert(
            "messages",
            values={
                "room": ":room",
                "author": ":author",
                "body": ":body",
                "at": ":at",
            },
            returning="id",
        ),
        room=room,
        author=author,
        body=body,
        at=at,
    )


async def main() -> None:
    path = Path("example.sqlite3")
    path.unlink(missing_ok=True)
    path.with_name(path.name + "-wal").unlink(missing_ok=True)
    path.with_name(path.name + "-shm").unlink(missing_ok=True)
    async with LiteWriter(path, hz=0) as db:
        db.execute_script(SCHEMA)
        user = await db.call(add_user, "Ada")
        msg = await db.call(send, 8, "hi", user, 1)
        print("msg", msg)
        print("page", db.query(page, room=8))
        print("mine", db.query(mine, room=8, me=user))
        print("empty", db.one(page, room=9))
        print("count", db.value(Select(("count", "*"), from_="messages")))
        row = db.execute(
            "SELECT name FROM users WHERE id = :id",
            {"id": user},
        ).fetchone()
        print("name", tuple(row))
        print("tables", sorted(db.tables(page)))
        with db.reader() as reader:
            print("reader", reader.query(page, room=8))
        async with db.watch(page, room=8) as live:
            found = aiter(live)
            print("now", await anext(found))
            await db.call(send, 8, "next", user, 2)
            print("later", await anext(found))


asyncio.run(main())

msg 1
page [(1, 'hi', 'ada')]
mine [(1, 'hi', 'ada')]
empty None
count 1
name ('Ada',)
tables ['messages', 'users']
reader [(1, 'hi', 'ada')]
now [(1, 'hi', 'ada')]
later [(2, 'next', 'ada'), (1, 'hi', 'ada')]

The writer calls fn(tx, *args, **kwargs) on the writer thread, inside BEGIN IMMEDIATE. You do not COMMIT. You do not keep tx. You do not block. tx.value is the first column of the first row. tx.one is the first row. tx.query is every row. tx.execute runs SQL you already have. A sequence fills ?. A mapping fills :name.

• await db.call(fn, ...) waits on this asyncio loop. The caller thread stays free.
• db.push(fn, ...) puts the write on the queue. It does not wait. Errors go to on_error.
• db.submit(fn, ...) returns a slot. .result() waits on the caller thread. On an asyncio loop, use call.
• The result comes after the batch COMMIT.
• hz=60 is the default. After a commit, the writer sleeps the rest of that 1/60 s. When the inbox is empty, the writer parks. hz=0 commits as fast as jobs arrive. You can change db.hz while it runs.
• isolated=True is the default. It sets a SAVEPOINT for that write. A failure undoes only that write. The error comes back after the rest of the batch commits. isolated(fn) does the same thing.
• isolated=False shares the batch. A failure rolls back the whole batch. The other writes in that batch get WriterRolledBack.
• db.close() drains the inbox, commits the last batch, and stops. A write that arrives too late fails with WriterRuntime.
• submit returns Slot[R], and call returns R, where R is the return type of fn. Pyright checks the arguments against fn.

Reads

q is SQL text or a query. A dict of clauses is still a query. query, one, and value take keyword arguments. execute takes the bindings as one argument.

A read does not enter the writer queue. The writer does not need to be started. The file must exist. A missing file raises WriterRuntime. db.query uses one read-only connection for this OS thread. Another thread opens its own connection. db.reader() opens one more. Use that reader from the thread that opened it. Close it on that thread. db.close() closes a reader opened on this thread. A reader on another thread closes on its next use.

Watch

The first read is the rows now. A later read comes after a commit that changed them. async for rows in live keeps going. The program above stops after the second read. A slow reader gets the newest rows, not a queue of old ones.

The commit does not run the query. SQLite names the columns the query reads, once, when the watch opens. The writer records inserts, deletes, and updated columns only while a watch is open. BEGIN, COMMIT, SAVEPOINT, RELEASE, ROLLBACK, and PRAGMA schema_version are the writer's own SQL. They add no user table.

An insert or a delete wakes every watch of that table. An update wakes a watch when the updated column is one the query reads. count(*) wakes on insert and delete only. After COMMIT, every matching watch on one asyncio loop is set in one wake.

A write to messages does not wake a watch of users. An update of users.seen does not wake SELECT name FROM users. The watch reads again on this thread's read connection. Equal rows do not yield. A rolled-back batch does not wake anyone. A schema change wakes every watch. A view, a CTE, and raw SQL text all work, because SQLite resolves them to base columns.

Install

CPython 3.14. The package is pure Python. pip install litewriter also installs APSW. A v* tag on the package repository publishes it.

pip install litewriter

From a checkout of this package:

uv sync
uv run pytest

Threads

submit, call, and push are safe from any OS thread. The writer owns the write connection. Each OS thread reads on its own connection. A reader() stays on the thread that opened it. Jobs sit on a deque and a Condition. After COMMIT the writer wakes each asyncio loop once per batch. A socketpair mailbox does that when the loop has add_reader. The Windows loop has no add_reader. One callback then carries the batch. Matching watches on that loop are set in that same wake. The mailbox is made on the loop's own thread (the first call or watch there), because add_reader is not thread-safe.

Each connection has a 64 MiB page cache and maps up to 1 GiB of the file (litewriter.connect.CACHE_KIB, MMAP_BYTES).