Background task queue for Python backed by Redis, a super minimal Celery. Read about the motivation behind this project on this blog post and the accompanying Hacker News discussion. WakaQ is currently used in production at WakaTime.com. WakaQ is also available in TypeScript.
- Queue priority
- Delayed tasks (run tasks after a timedelta eta)
- Scheduled periodic tasks
- Tasks can be async or normal synchronous functions
- Broadcast a task to all workers
- Task soft and hard timeout limits
- Optionally retry tasks on soft timeout
- Combat memory leaks with
max_mem_percentormax_tasks_per_worker - Super minimal
Want more features like rate limiting, task deduplication, etc? Too bad, feature PRs are not accepted. Maximal features belong in your app’s worker tasks.
Polyglot by design. WakaQ works with Python and TypeScript, speaking the same Redis queue protocol. Enqueue a task from a Python service and run it on a Node worker, or the reverse. One Redis instance, one set of queues, two languages. No other task queue in the comparison below can do this.
Production-proven. WakaQ has powered background processing at WakaTime for many years (since 2022). It isn't a proof of concept, it runs real workloads every day.
A codebase you can actually read. WakaQ is small and deliberately simple. You can read the entire source in an afternoon, so when something misbehaves you can understand exactly why instead of spelunking through a sprawling framework. Cron schedules, ETA tasks, broadcast tasks, soft/hard timeouts, and retries — all in a surface area small enough to keep in your head.
| # | Library | License | Broker(s) | Notes |
|---|---|---|---|---|
| 1 | Celery | BSD-3 | RabbitMQ, Redis, SQS | Industry standard. Most features, biggest ecosystem; heavy to operate. |
| 2 | arq | MIT | Redis | Best pick for asyncio-native code. Small, well-built. |
| 3 | TaskTiger | MIT | Redis | Reliability-focused: unique tasks, task locks, flexible retries, periodic tasks, subqueues, per-task forking. Production-proven at Close. |
| 4 | RQ | BSD-3 | Redis | Simplest sync mental model. Big community, thinner feature set. |
| 5 | WakaQ | BSD-3 | Redis | Lightweight Celery alternative — cron/scheduled, ETA, broadcast tasks, soft/hard timeouts, retries. Polyglot with wakaq-ts. |
| 6 | Dramatiq | LGPL-3 | RabbitMQ, Redis | Clean API and middleware, but LGPL gets flagged by many corporate compliance reviews. |
| 7 | TaskIQ | MIT | Redis, RabbitMQ, NATS | Async-first, pluggable broker/result backends. Newer, smaller community. |
| 8 | huey | MIT | Redis, SQLite | Minimal; SQLite mode avoids running Redis. |
| 9 | Procrastinate | MIT | Postgres | Pick only if you specifically want a Postgres broker. |
| # | Library | License | Broker(s) | Notes |
|---|---|---|---|---|
| 1 | BullMQ | MIT | Redis | De facto standard. Flows, rate limiting, repeatable jobs. |
| 2 | pg-boss | MIT | Postgres | Best option for transactional enqueue with no Redis. |
| 3 | Graphile Worker | MIT | Postgres | High-throughput, LISTEN/NOTIFY-driven low latency. |
| 4 | WakaQ-TS | BSD-3 | Redis | Lightweight; same cron/ETA/broadcast/timeout feature set as WakaQ. Polyglot with wakaq. |
| 5 | Bee-Queue | MIT | Redis | Simple and fast for short, latency-sensitive jobs. |
| 6 | Agenda | MIT | MongoDB | Reasonable if already on MongoDB; uneven maintenance. |
| 7 | Bull (legacy) | MIT | Redis | Superseded by BullMQ — don't start new projects on it. |
pip install wakaq
import logging
from datetime import timedelta
from wakaq import WakaQ, Queue, CronTask
# use constants to prevent misspelling queue names
Q_HIGH = 'a-high-priority-queue'
Q_MED = 'a-medium-priority-queue'
Q_LOW = 'a-low-priority-queue'
Q_OTHER = 'another-queue'
Q_DEFAULT = 'default-lowest-priority-queue'
wakaq = WakaQ(
# List your queues and their priorities.
# Queues can be defined as Queue instances, tuples, or just a str.
queues=[
(0, Q_HIGH),
(1, Q_MED),
(2, Q_LOW),
Queue(Q_OTHER, priority=3, max_retries=5, soft_timeout=300, hard_timeout=360),
Q_DEFAULT,
],
# Number of worker processes. Must be an int or str which evaluates to an
# int. The variable "cores" is replaced with the number of processors on
# the current machine.
concurrency="cores*4",
# Number of concurrent asyncio tasks per worker process. Must be an int or
# str which evaluates to an int. The variable "cores" is replaced with the
# number of processors on the current machine. Default is zero for no limit.
async_concurrency=0,
# Raise SoftTimeout or asyncio.CancelledError in a task if it runs longer
# than 30 seconds. Can also be set per task or queue. If no soft timeout
# set, tasks can run forever.
soft_timeout=30, # seconds
# SIGKILL a task if it runs longer than 1 minute. Can be set per task or queue.
hard_timeout=timedelta(minutes=1),
# If the task soft timeouts, retry up to 3 times. Max retries comes first
# from the task decorator if set, next from the Queue's max_retries,
# lastly from the option below. If No max_retries is found, the task
# is not retried on a soft timeout.
max_retries=3,
# Combat memory leaks by reloading a worker (the one using the most RAM),
# when the total machine RAM usage is at or greater than 98%.
max_mem_percent=98,
# Combat memory leaks by reloading a worker after it's processed 5000 tasks.
max_tasks_per_worker=5000,
# Schedule two tasks, the first runs every minute, the second once every ten minutes.
# Scheduled tasks can be passed as CronTask instances or tuples. To run scheduled
# tasks you must keep a wakaq scheduler running as a daemon.
schedules=[
# Runs mytask on the queue with priority 1.
CronTask('* * * * *', 'mytask', queue=Q_MED, args=[2, 2], kwargs={}),
# Runs mytask once every 5 minutes.
('*/5 * * * *', 'mytask', [1, 1], {}),
# Runs anothertask on the default lowest priority queue.
('*/10 * * * *', 'anothertask'),
],
)
# timeouts can be customized per task with a timedelta or integer seconds
@wakaq.task(queue=Q_MED, max_retries=7, soft_timeout=420, hard_timeout=480)
def mytask(x, y):
print(x + y)
@wakaq.task
def a_cpu_intensive_task():
print("hello world")
@wakaq.task
async def an_io_intensive_task():
print("hello world")
@wakaq.wrap_tasks_with
async def custom_task_decorator(fn, args, kwargs):
# do something before each task runs, for ex: `with app.app_context():`
if inspect.iscoroutinefunction(fn):
await fn(*args, **kwargs)
else:
fn(*args, **kwargs)
# do something after each task runs
if __name__ == '__main__':
# add 1 plus 1 on a worker somewhere
mytask.delay(1, 1)
# add 1 plus 1 on a worker somewhere, overwriting the task's queue from medium to high
mytask.delay(1, 1, queue=Q_HIGH)
# print hello world on a worker somewhere, running on the default lowest priority queue
anothertask.delay()
# print hello world on a worker somewhere, after 10 seconds from now
anothertask.delay(eta=timedelta(seconds=10))
# print hello world on a worker concurrently, even if you only have 1 worker process
an_io_intensive_task.delay()See the WakaQ init params for a full list of options, like Redis host and Redis socket timeout values.
When using in production, make sure to increase the max open ports allowed for your Redis server process.
When using eta tasks a Redis sorted set is used, so eta tasks are automatically deduped based on task name, args, and kwargs.
If you want multiple pending eta tasks with the same arguments, just add a throwaway random string to the task’s kwargs for ex: str(uuid.uuid1()).
Here’s an example systemd config to run wakaq-worker as a daemon:
[Unit]
Description=WakaQ Worker Service
[Service]
WorkingDirectory=/opt/yourapp
ExecStart=/opt/yourapp/venv/bin/python /opt/yourapp/venv/bin/wakaq-worker --app=yourapp.wakaq
RemainAfterExit=no
Restart=always
RestartSec=30s
KillSignal=SIGINT
LimitNOFILE=99999
[Install]
WantedBy=multi-user.target
Create a file at /etc/systemd/system/wakaqworker.service with the above contents, then run:
systemctl daemon-reload && systemctl enable wakaqworker
In dev and test environments, it’s easier to run tasks synchronously so you don’t need Redis or any worker processes. The recommended way is mocking WakaQ:
class WakaQMock:
def __init__(self):
self.task = TaskMock
def wrap_tasks_with(self, fn):
return fn
class TaskMock(object):
fn = None
name = None
args = ()
kwargs = {}
def __init__(self, *args, **kwargs):
if len(args) == 1 and len(kwargs) == 0:
self.fn = args[0]
self.name = args[0].__name__
else:
self.args = args
self.kwargs = kwargs
def delay(self, *args, **kwargs):
kwargs.pop("queue", None)
kwargs.pop("eta", None)
return self.fn(*args, **kwargs)
def broadcast(self, *args, **kwargs):
return
def __call__(self, *args, **kwargs):
if not self.fn:
task = TaskMock(args[0])
task.args = self.args
task.kwargs = self.kwargs
return task
else:
return self.fn(*args, **kwargs)Then in dev and test environments instead of using wakaq.WakaQ use WakaQMock.