execute_write_fn(block=False) raises TypeError when num_sql_threads is 0

[HarperZ9]

With num_sql_threads set to 0, await db.execute_write_fn(fn, block=False) raises TypeError: cannot unpack non-iterable NoneType object instead of returning a UUID.

The docs describe block=False as fire-and-forget returning a task id:

If you specify block=False the method becomes fire-and-forget [...] A UUID representing the queued task will be returned.

Cause

In datasette/database.py, execute_write_fn takes a non-threaded branch when there is no executor, so result is whatever fn returned:

if self.ds.executor is None:
    # non-threaded mode
    ...
    result = fn(self._write_connection)
else:
    result = await self._send_to_write_thread(fn, block=block, transaction=transaction)

The block=False path then unpacks result unconditionally:

else:
    # For non-blocking writes, spawn a background task ...
    task_id, reply_future = result

Only _send_to_write_thread returns the (task_id, reply_future) pair. In the non-threaded branch result is the function's return value, which for an ordinary write function is None.

Reproduction

Against e889403d3bbe143854262682161c98a57bdb6594:

import asyncio
from datasette.app import Datasette

async def main():
    ds = Datasette(memory=True, settings={"num_sql_threads": 0})
    await ds.invoke_startup()
    db = ds.add_memory_database("repro")

    def make_table(conn):
        conn.execute("CREATE TABLE IF NOT EXISTS t (id INTEGER PRIMARY KEY, v TEXT)")

    await db.execute_write_fn(make_table, block=True)   # fine

    def w(conn):
        conn.execute("INSERT INTO t (v) VALUES ('a')")  # returns None

    await db.execute_write_fn(w, block=False)           # TypeError

asyncio.run(main())

TypeError: cannot unpack non-iterable NoneType object

block=True works in the same configuration, and both work with the default num_sql_threads.

Scope, and what I am not claiming

• This only affects num_sql_threads = 0. The default configuration is unaffected.
• I have not found an in-core caller that hits this, so I am not claiming a user-facing break in shipped Datasette. It is reachable from the documented plugin API.
• I have not tested this under Pyodide itself, only with the setting that selects the same branch.
• No security or data-loss impact is claimed. The write is applied before the unpack raises, so the failure is in the return path, not the write.

Possible fix

Return a task id in the non-threaded branch too, so the documented contract holds, or skip the unpack when there is no executor. I did not want to guess which shape you would prefer, and I am happy to send a PR with tests once you say which.

---

Disclosure: I used an AI coding assistant while investigating this. The reproduction above was run and confirmed by me on the stated commit.

[HarperZ9]

Correction and follow-up.

While building the fix I found a second, more consequential bug in the same method, and it changes the framing here. _send_to_write_thread() derives the task id from uuid.uuid5(uuid.NAMESPACE_DNS, "datasette.io"), which is a constant, so every non-blocking write returns the same UUID. Filed separately as #2860.

That also corrects something in my report above: I described the num_sql_threads=0 TypeError as though the threaded path was otherwise fine. It returns a UUID, but the same one every time.

PR #2861 fixes both, since they are two lines in the same method and one regression test covers both. Happy to split it if you would rather review them apart.