a simple but robust task queue
Project description
delayed
Delayed is a simple but robust task queue inspired by rq.
Features
- Robust: all the enqueued tasks will run exactly once, even if the worker got killed at any time.
- Clean: finished tasks (including failed) won't take the space of your Redis.
- Distributed: workers as more as needed can run in the same time without further config.
Requirements
- Python 2.7 or later, tested on Python 2.7, 3.3 - 3.9, PyPy and PyPy3.
- UNIX-like systems (with os.fork() implemented), tested on Ubuntu and macOS.
- Redis 2.6.0 or later.
- Keep syncing time among all the machines of each task queue.
Getting started
-
Run a redis server:
$ redis-server
-
Install delayed:
$ pip install delayed
-
Create a task queue:
import redis from delayed.queue import Queue conn = redis.Redis() queue = Queue(name='default', conn=conn)
-
Four ways to enqueue a task:
-
Define a task function and enqueue it:
from delayed.delay import delayed delayed = delayed(queue) @delayed() def delayed_add(a, b): return a + b delayed_add.delay(1, 2) # enqueue delayed_add delayed_add.delay(1, b=2) # same as above delayed_add(1, 2) # call it immediately
-
Directly enqueue a function:
from delayed.delay import delay, delayed delay = delay(queue) delayed = delayed(queue) def add(a, b): return a + b delay(add)(1, 2) delay(add)(1, b=2) # same as above delayed()(add).delay(1, 2) delayed()(add).delay(1, b=2) # same as above
-
Create a task and enqueue it:
from delayed.task import Task def add(a, b): return a + b task = Task.create(func=add, args=(1,), kwargs={'b': 2}) queue.enqueue(task)
-
Enqueue a predefined task function without importing it:
from delayed.task import Task task = Task(id=None, func_path='test.add', args=(1,), kwargs={'b': 2}) queue.enqueue(task)
-
-
Run a task worker (or more) in a separated process:
import redis from delayed.queue import Queue from delayed.worker import ForkedWorker conn = redis.Redis() queue = Queue(name='default', conn=conn) worker = ForkedWorker(queue=queue) worker.run()
-
Run a task sweeper in a separated process to recovery lost tasks (mainly due to the worker got killed):
import redis from delayed.queue import Queue from delayed.sweeper import Sweeper conn = redis.Redis() queue = Queue(name='default', conn=conn) sweeper = Sweeper(queue=queue) sweeper.run()
Examples
See examples.
```bash
$ redis-server &
$ pip install delayed
$ python -m examples.caller &
$ python -m examples.forked_worker # or python -m examples.preforked_worker
```
QA
-
Q: What's the limitation on a task function?
A: A task function should be defined in module level (except the__main__
module). Itsargs
andkwargs
should be picklable. -
Q: What's the
name
param of a queue?
A: It's the key used to store the tasks of the queue. A queue with name "default" will use those keys:- default: list, enqueued tasks.
- default_id: str, the next task id.
- default_noti: list, the same length as enqueued tasks.
- default_enqueued: sorted set, enqueued tasks with their timeouts.
- default_dequeued: sorted set, dequeued tasks with their dequeued timestamps.
-
Q: Why the worker is slow?
A: TheForkedWorker
forks a new process for each new task. So all the tasks are isolated and you won't leak memory.
To reduce the overhead of forking processes and importing modules, if your task function code won't be changed in the worker's lifetime, you can switch toPreforkedWorker
:import redis from delayed.queue import Queue from delayed.worker import PreforkedWorker conn = redis.Redis() queue = Queue(name='default', conn=conn) worker = PreforkedWorker(queue=queue) worker.run()
-
Q: How does a
ForkedWorker
run?
A: It runs such a loop:- It dequeues a task from the queue periodically.
- It forks a child process to run the task.
- It kills the child process if the child runs out of time.
- When the child process exits, it releases the task.
-
Q: How does a
PreforkedWorker
run?
A: It runs such a loop:- It dequeues a task from the queue periodically.
- If it has no child process, it forks a new one.
- It sends the task through a pipe to the child.
- It kills the child process if the child runs out of time.
- When the child process exits or it received result from the pipe, it releases the task.
-
Q: How does the child process of a worker run?
A: The child of aForkedWorker
just runs the task, unmarks the task as dequeued, then exits. The child of aPreforkedWorker
runs such a loop:- It tries to receive a task from the pipe.
- If the pipe has been closed, it exits.
- It runs the task.
- It sends the task result to the pipe.
- It releases the task.
-
Q: What's lost tasks?
A: There are 2 situations a task might get lost:- a worker popped a task notification, then got killed before dequeueing the task.
- a worker dequeued a task, then both the monitor and its child process got killed before they releasing the task.
-
Q: How to recovery lost tasks?
A: Runs a sweeper. It dose two things:- it keeps the task notification length the same as the task queue.
- it moves the timeout dequeued tasks back to the task queue.
-
Q: How to set the timeout of tasks?
A: You can setdefault_timeout
of a queue ortimeout
of a task:from delayed.delay import delay_with_params queue = Queue('default', conn, default_timeout=60) delayed_add.timeout(10)(1, 2) delay_with_params(queue)(timeout=10)(add)(1, 2)
-
Q: How to enqueue a task in front of the queue?
A: You can setprior
of the task toTrue
:task = Task(id=None, func_path='test.add', args=(1, 2), prior=True) queue.enqueue(task)
-
Q: How to handle the failed tasks?
A: Sets theerror_handler
of the task. The handlers would be called in a forked process, except the forked process got killed or the monitor process raised an exception.from delayed.delay import delay_with_params def error_handler(task, kill_signal, exc_info): if kill_signal: logging.error('task %d got killed by signal %d', task.id, kill_signal) else: logging.exception('task %d failed', task.id, exc_info=exc_info) @delayed_with_param(queue)(error_handler=error_handler) def error(): raise Exception def error2(): raise Exception task = Task.create(func_path='test.error2', error_handler=error_handler)
-
Q: Why does sometimes the
error_handler
not be called for a failed task?
A: If both the child process and the monitor process got killed at the same time, there is no chance to call theerror_handler
. -
Q: How to turn on the debug logs?
A: Adds alogging.DEBUG
level handler todelayed.logger.logger
. The simplest way is to calldelayed.logger.setup_logger()
:from delayed.logger import setup_logger setup_logger()
-
Q: Can I enqueue and dequeue tasks in different Python versions?
A:delayed
uses thepickle
module to serialize and deserialize tasks. Ifpickle.HIGHEST_PROTOCOL
is equal among all your Python runtimes, you can use it without any configurations. Otherwise you have to choose the lowestpickle.HIGHEST_PROTOCOL
of all your Python runtime as the pickle protocol. eg: If you want to enqueue a task in Python 3.7 and dequeue it in Python 2.7. Theirpickle.HIGHEST_PROTOCOL
are4
and2
, so you need to set the version to2
:from delayed.task import set_pickle_protocol_version set_pickle_protocol_version(2)
-
Q: Why not use JSON or MessagePack to serialize tasks?
A: These serializations may confuse some types (eg:bytes
/str
,list
/tuple
). -
Q: What will happen if I changed the pipe capacity?
A:delayed
assumes the pipe capacity is 65536 bytes (the default value on Linux and macOS). To reduce syscalls, it won't check whether the pipe is writable if the length of data to be written is less than 65536. If your system has a lower pipe capacity, thePreforkedWorker
may not working well for some large tasks. To fix it, you can set a lower value todelayed.constants.BUF_SIZE
:import delayed.constants delayed.constants.BUF_SIZE = 1024
Release notes
-
0.9:
- Adds
prior
anderror_handler
params todeleyed.delayed()
, removes itstimeout()
method. - Adds examples.
- Adds
-
0.8:
- The
Task
struct has been changed, it's not compatible with older versions.- Removes
module_name
andfunc_name
fromTask
, addsfunc_path
instead. - Adds
error_handler_path
toTask
.
- Removes
- Removes
success_handler
anderror_handler
fromWorker
.
- The
-
0.7:
- Implements prior task.
-
0.6:
- Adds
dequeued_len()
andindex
toQueue
.
- Adds
-
0.5:
- Adds
delayed.task.set_pickle_protocol_version()
.
- Adds
-
0.4:
- Refactories and fixes bugs.
-
0.3:
- Changes param
second
totimeout
fordelayed.delayed()
. - Adds debug log.
- Changes param
-
0.2:
- Adds
timeout()
todelayed.delayed()
.
- Adds
-
0.1:
- Init version.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distribution
Hashes for delayed-0.9.1b1-py2.py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | dff2e31a6a7af7439d3f060105e5ecb3c9eb8775b08b434d922c1a19c6acbe54 |
|
MD5 | cbfcd273cb494658c9accc566a1db96a |
|
BLAKE2b-256 | eeaf9978140d52bbf001ebf6ed491f377b1d489e9acf2124ce8fa4b687a44de4 |