datasette/datasette/background_tasks.py
Alex Garcia 374b194ff5
Add /-/tasks introspection endpoint for supervised background tasks (#2892)
Co-authored-by: Simon Willison <swillison@gmail.com>
2026-09-15 11:56:53 -07:00

227 lines
8.6 KiB
Python

"""
Supervised background-task registration for Datasette core.
Plugins that need long-lived background work (a polling loop, a queue
consumer, a scheduled job runner) register it with
``datasette.add_background_task(func, name=None)`` - typically from a
``startup`` plugin hook - instead of fire-and-forgetting their own
``asyncio.create_task()``. Core owns:
- **references**: every launched ``asyncio.Task`` is kept alive on a
:class:`BackgroundTaskSupervisor`, so it can never be silently garbage
collected the way an unreferenced ``create_task()`` call can be;
- **launch timing**: registered work is buffered until
:meth:`BackgroundTaskSupervisor.launch_all` runs, which core arranges to
happen only after *every* plugin's ``startup`` hook has finished - so
a task that depends on another plugin having registered something first
doesn't need ``tryfirst=True`` ordering tricks;
- **crash surfacing**: an unhandled exception in a background task is
logged with its full traceback to the ``datasette.background_tasks``
logger and recorded on the handle, instead of becoming an "Task
exception was never retrieved" warning nobody sees;
- **cancellation**: :meth:`BackgroundTaskSupervisor.cancel_all` cancels
every task still running and waits (with a grace period) for them to
actually stop.
"""
from __future__ import annotations
import asyncio
import datetime
import functools
import logging
from collections.abc import Awaitable, Callable
logger = logging.getLogger("datasette.background_tasks")
def _utcnow_iso() -> str:
return datetime.datetime.now(datetime.timezone.utc).isoformat()
def _function_path(func: Callable) -> str:
"""Describe the callable without guessing which plugin registered it."""
while isinstance(func, functools.partial):
func = func.func
if not hasattr(func, "__qualname__"):
func = type(func).__call__
return f"{func.__module__}.{func.__qualname__}"
class BackgroundTask:
"""A handle to a single piece of supervised background work.
States: ``registered`` (added but not yet launched) -> ``running`` ->
one of ``completed`` (returned cleanly), ``crashed`` (raised an
exception other than ``CancelledError`` - see ``.exception``), or
``cancelled`` (``.cancel()`` was called, or it was still running at
shutdown).
"""
def __init__(
self,
name: str,
func: Callable[[object], Awaitable[None]],
):
self.name = name
self.state = "registered"
self.task: asyncio.Task | None = None
self.exception: BaseException | None = None
self.started_at: str | None = None
self.function = _function_path(func)
self._func = func
self._supervisor: BackgroundTaskSupervisor | None = None
def cancel(self) -> None:
"""Cancel this task.
If it has already been launched, cancels the underlying
``asyncio.Task`` - its state becomes ``cancelled`` once the
cancellation is observed (asynchronously, via the task's done
callback). If it has not been launched yet, this is a no-op as
far as asyncio is concerned (there's no task to cancel) but it
deregisters the handle from its supervisor so it never runs.
"""
if self.task is not None:
self.task.cancel()
elif self._supervisor is not None:
self._supervisor._deregister(self)
def __repr__(self) -> str:
return f"<BackgroundTask name={self.name!r} state={self.state!r}>"
class BackgroundTaskSupervisor:
"""Owns registration and launch of every :class:`BackgroundTask` for a
single ``Datasette`` instance.
Registration (:meth:`add`) is separate from launch
(:meth:`launch_all`): plugins register work whenever convenient
(typically from a ``startup`` hook, but request handlers can register
dynamic per-job work too), and it either sits buffered until
:meth:`launch_all` runs, or - if :meth:`launch_all` has already run -
starts immediately.
Strong references to every :class:`BackgroundTask` (and its
``asyncio.Task``) are kept for the life of the instance, by design -
that's what makes the enrichments-style "fire-and-forget task gets
garbage collected mid-flight" bug impossible here. There is currently
no pruning of completed/crashed/cancelled tasks, so a plugin that
dynamically registers many short-lived tasks over a long process
lifetime (a per-job registration pattern, e.g. one task per queued
job) will grow this list without bound. That's an accepted v1
trade-off in favour of full introspection (``/-/tasks``); revisit
with a pruning or capping policy if unbounded growth is reported in
practice.
"""
def __init__(self, datasette):
self._datasette = datasette
self._tasks: list[BackgroundTask] = []
self._names = set()
self._launched = False
self._lock = asyncio.Lock()
def add(self, func, name=None) -> BackgroundTask:
base_name = name or getattr(func, "__qualname__", None) or repr(func)
actual_name = self._unique_name(base_name)
handle = BackgroundTask(actual_name, func)
handle._supervisor = self
self._tasks.append(handle)
self._names.add(actual_name)
if self._launched:
self._launch_one(handle)
return handle
def _unique_name(self, base_name: str) -> str:
if base_name not in self._names:
return base_name
n = 2
while f"{base_name}-{n}" in self._names:
n += 1
return f"{base_name}-{n}"
def _deregister(self, handle: BackgroundTask) -> None:
try:
self._tasks.remove(handle)
except ValueError:
pass
self._names.discard(handle.name)
def _launch_one(self, handle: BackgroundTask) -> None:
handle.state = "running"
handle.started_at = _utcnow_iso()
handle.task = asyncio.create_task(
handle._func(self._datasette), name=handle.name
)
handle.task.add_done_callback(functools.partial(_on_task_done, handle))
async def launch_all(self) -> None:
"""Launch every currently-registered task that hasn't launched
yet. Idempotent and safe to call concurrently: subsequent (or
racing) calls are no-ops once the first has set ``self._launched``.
"""
if self._launched:
return
async with self._lock:
if self._launched:
return
self._launched = True
for handle in list(self._tasks):
if handle.task is None:
self._launch_one(handle)
async def cancel_all(self, grace: float = 5.0) -> None:
"""Cancel every task that isn't already done, then wait up to
``grace`` seconds for them to actually finish. Stragglers still
running after that are logged by name (but left to finish or not
on their own - this does not forcibly kill them, asyncio has no
mechanism for that).
"""
handles_by_task = {
handle.task: handle for handle in self._tasks if handle.task is not None
}
pending = [task for task in handles_by_task if not task.done()]
for task in pending:
task.cancel()
if not pending:
return
_done, not_done = await asyncio.wait(pending, timeout=grace)
if not_done:
names = sorted(handles_by_task[task].name for task in not_done)
logger.warning(
"%d background task(s) did not finish within the %.1fs grace "
"period after cancellation: %s",
len(names),
grace,
", ".join(names),
)
def tasks(self) -> list[BackgroundTask]:
"""Return every registered :class:`BackgroundTask`, launched or
not, in registration order. Used by the ``/-/tasks`` debug
endpoint.
"""
return list(self._tasks)
@property
def launched(self) -> bool:
"""Whether :meth:`launch_all` has run yet - lets ``/-/tasks``
distinguish "no tasks registered" from "tasks registered but
nothing has armed the launch yet" without reaching for the
private ``_launched`` attribute.
"""
return self._launched
def _on_task_done(handle: BackgroundTask, task: asyncio.Task) -> None:
if task.cancelled():
handle.state = "cancelled"
return
exc = task.exception()
if exc is not None:
handle.state = "crashed"
handle.exception = exc
logger.error("Background task %r crashed", handle.name, exc_info=exc)
return
handle.state = "completed"