vercel-celery-bundle

This is a version of vercel-celery with third-party dependencies bundled. For normal use, install the unbundled vercel-celery package instead: https://pypi.org/project/vercel-celery/

vercel-celery

Celery integration for Vercel Queue Service and Vercel Runtime Cache.

This package provides a Celery broker transport backed by Vercel Queue Service and a Celery result backend backed by Vercel Runtime Cache.

from vercel.integrations.celery import install_vercel_celery_integration

install_vercel_celery_integration()

By default, named Celery apps publish to app-prefixed Vercel Queue topics. For example, Celery("my_app") maps Celery queue emails to the Vercel Queue topic celery-my__app-emails after Vercel Queue name sanitization. Apps with no name use unprefixed topics. To override this, set broker_transport_options = {"queue_name_prefix": "jobs-"}. Set queue_name_prefix to "" to use Celery queue names as Vercel Queue topics.

The broker's Vercel Queue consumer group defaults to celery. Workers sharing the same Vercel Queue topic and consumer group compete for tasks. Workers using the same topic with different consumer groups receive fan-out copies. To override it, set broker_transport_options = {"consumer_group": "workers"}.

The installer sets Celery's default broker_url to vercel:// when Celery has no broker default configured. Pass set_default_broker=False to opt out. It also sets Celery's default result_backend to vercel-runtime-cache:// when Celery has no result backend configured. Pass set_default_result_backend=False to opt out.

vercel:// uses push delivery when VERCEL is truthy in the environment and poll delivery otherwise. Set broker_url = "vercel-push://" to force push delivery, or broker_url = "vercel-poll://" for forced polling workers, local workers, and background workers.

To deploy a Celery worker on Vercel, declare a module that exposes the Celery app as a queue subscriber in pyproject.toml:

# worker.py
from tasks import app

__all__ = ["app"]

[[tool.vercel.subscribers]]
entrypoint = "worker:app"

The Vercel build introspects the app's queues and compiles the subscriber into a queue-triggered function; no vercel.json trigger configuration is needed. With no topics filter, the subscriber consumes every queue the app declares.

Pass register_queues=False to skip automatic Vercel Queue trigger registration. When using that mode, manually register each Celery app whose queues should receive Vercel Queue push deliveries:

from vercel.integrations.celery import register_celery_app_queues

register_celery_app_queues(app)

To customize Celery task result storage in Vercel Runtime Cache, configure Celery's result backend transport options:

result_backend_transport_options = {
    "namespace": "celery-results",
    "ttl": 3600,
}

By default, named Celery apps use an app-scoped Runtime Cache namespace. For example, Celery("my_app") stores results under celery-results-my_app. Apps with no name use celery-results. If producers and workers use different Celery app names, configure a shared explicit namespace.

The Runtime Cache backend uses vercel.cache, including its normal key hashing, namespacing, and local in-memory fallback when Runtime Cache is unavailable. Celery result backends are not required to retain results indefinitely; results may expire according to Celery's result expiration settings or the backend's own retention behavior. This backend is therefore an expiring, cache-backed result backend rather than durable result storage. Celery treats missing result entries as pending results. By default, stored results use Celery's result_expires value as their Runtime Cache TTL; set result_backend_transport_options["ttl"] to override that retention period.

JSON result serialization works normally. Binary serializers, including pickle, are stored through an internal base64 wrapper so Runtime Cache only receives JSON-compatible values.

Runtime Cache is a single-key cache API, so this backend does not provide native bulk result fetches or increment-backed chord counters. Result availability is limited by Runtime Cache retention and eviction behavior.