Skip to content

guillotinaweb/guillotina_amqp

Repository files navigation

Plugin documentation

This plugin integrates aioamqp into guillotina, providing an execution framework for asyncio tasks:

  • Guillotina command to start a worker: amqp-worker
  • Workers consume tasks from rabbit-mq through the aioamqp integration
  • Redis state manager implementation to keep a global view of running tasks
  • Utilities and endpoints for adding new tasks and for task cancellation

Its distributed design - the absence of a central worker manager - makes it more robust. Task cancelation is signaled over the state manager, and workers will be responsible for stopping canceled tasks.

A watchdog on the asyncio loop can be launched with the auto-kill-timeout command argument, which will kill the worker if one of its tasks has captured the loop for too long.

When a task fails, the worker will send it to the delay queue, which has been configured to re-queue tasks to the main queue after a certain TTL. Failed tasks are retried a limited amount of times.

Configuration

Example docs

    {
        "amqp": {
            "host": "localhost",
            "port": 5673,
            "login": "guest",
            "password": "guest",
            "vhost": "/",
            "heartbeat": 800,
            "queue": "guillotina",
            "persistent_manager": "redis",
            "delayed_ttl_ms": 60000,
            "errored_ttl_ms": 604800000,
        }
    }
  • host and port: should point to the rabbit-mq instance
  • login and password: should match the rabbit-mq access credentials
  • queue: main queue where tasks are consumed from
  • persistent_manager: named utility to use to keep tasks state.
  • delay_ttl_ms and errored_ttl_ms: can be used to configure queue delays. Default to 2 minutes and 1 week, correspondingly.
  • max_running_tasks: maximum number of simultaneous asyncio tasks hat workers are allowed to run.
  • max_task_retries: Max number of retries before an errored task is sent to the dead letter queue. If set to None, it will be retried forever.

Dependencies

Python >= 3.7

Installation

This example will use virtualenv::

virtualenv .
./bin/pip install .[test]

Running

Most simple way to get running::

./bin/guillotina

Queue tasks

    from guillotina_amqp import add_task
    await add_task(my_func, 'foobar', kw_arg='blah')

With decorators

from guillotina_amqp import task

    @task
    async def my_func(foo):
        print(foo)

    await my_func('bar')

Run the worker

    g amqp-worker

You can use a couple of additional parameters:

  • --auto-kill-timeout: time of inactivity after which the worker will restart himself assuming it got stuck.
  • --max-running-tasks: max number of simultaneous asyncio tasks in the event loop. Overwrites configuraiton parameter.

API

  • GET /@amqp-tasks - get list of tasks
  • GET /@amqp-tasks/{task_id} - get task info
  • DELETE /@amqp-tasks/{task_id} - delete task