Celery, configured in config/celery.py, autodiscovering tasks.py in every
installed app.
With no CELERY_BROKER_URL (or REDIS_URL), CELERY_TASK_ALWAYS_EAGER is on
and .delay() runs the task inline, synchronously, in the calling process.
This is not just convenience. Without a broker .delay() would otherwise
silently queue into nothing, and you would not discover it until production.
Running inline means task code is genuinely exercised in dev, and
CELERY_TASK_EAGER_PROPAGATES makes failures raise instead of being swallowed.
The trade-off is that eager mode does not exercise serialization boundaries or real concurrency. Before shipping anything task-heavy, run a real worker once.
# terminal 1
redis-server
# terminal 2 — note the pool flag on Windows
uv run celery -A config worker --loglevel=info --pool=solo
# terminal 3, if you need periodic tasks
uv run celery -A config beat --loglevel=info
--pool=solo is required on Windows. Celery's default prefork pool relies
on fork() and does not work there; without it the worker starts and then fails
in confusing ways. On Linux in production, drop the flag.
Verify the wiring:
uv run celery -A config call core.ping
from celery import shared_task
@shared_task(name="myapp.do_thing")
def do_thing(thing_id: int) -> str: ...Use @shared_task, not @app.task, so the module does not import the Celery
app and stays easy to test.
Pass identifiers, not objects. Arguments are JSON-serialized
(CELERY_TASK_SERIALIZER = "json"), and a model instance pickled into a queue is
stale by the time it runs.
| Task | Purpose |
|---|---|
core.ping |
Confirms a worker is connected. |
core.send_email |
Sends mail with retry and exponential backoff, up to 3 attempts. |
All allauth mail — verification, password reset — goes through core.send_email
via AccountAdapter.send_mail. Rendering happens in the request (the templates
need request context); only delivery is queued, so a slow SMTP server cannot
stall a signup response.
django-celery-beat is installed, so schedules live in the database and are
editable in the admin under Periodic Tasks. No code deploy to change a
schedule.
Run the worker as its own process, separate from the web server:
celery -A config worker --loglevel=info --concurrency=4
celery -A config beat --loglevel=info
Run exactly one beat process. Two will double-fire every scheduled task.