Mishka Queue - A django task queue

Mishka queue is a task queue system for Django backed by postgres.

It was forked from:

• https://github.com/SweetProcess/django-pg-queue, which in turn was forked from
• https://github.com/gavinwahl/django-postgres-queue by Gavin Wahl.

Why postgres?

I thought you were never supposed to use an RDBMS as a queue? Well, postgres has some features that make it not as bad as you might think, it has some compelling advantages.

• Transactional behavior and reliability.

  Adding tasks is atomic with respect to other database work. There is no need to use transaction.on_commit hooks and there is no risk of a transaction being committed but the tasks it queued being lost.

  Processing tasks is atomic with respect to other database work. Database work done by a task will either be committed, or the task will not be marked as processed, no exceptions. If the task only does database work, you achieve true exactly-once message processing.

• Operational simplicity

  By reusing the durable, transactional storage that we're already using anyway, there's no need to configure, monitor, and backup another stateful service. For small teams and light workloads, this is the right trade-off.

• Easy introspection

  Since tasks are stored in a database table, it's easy to query and monitor the state of the queue.

• Safety

  By using postgres transactions, there is no possibility of jobs being left in a locked or ambiguous state if a worker dies. Tasks immediately become available for another worker to pick up. You can even kill -9 a worker and be sure your database and queue will be left in a consistent state.

• Priority queues

  Since ordering is specified explicitly when selecting the next task to work on, it's easy to ensure high-priority tasks are processed first.

• Queues

  Simply implemented by allowing filtering by a queue name in the query.

Disadvantages

• Lower throughput than a dedicated queue server.
• Harder to scale a relational database than a dedicated queue server.
• Thundering herd. Postgres will notify all workers who LISTEN for the same name.
• With at-least-once delivery, a postgres transaction has to be held open for the duration of the task. For long running tasks, this can cause table bloat and performance problems.
• When a task crashes or raises an exception under at-least-once delivery, it immediately becomes eligible to be retried. If you want to implement a retry delay, you must catch exceptions and requeue the task with a delay. If your task crashes without throwing an exception (eg SIGKILL), you could end up in an endless retry loop that prevents other tasks from being processed.

How it works

mishka-queue is able to claim, process, and remove a task in a single (simplified) query.

DELETE FROM pgq_job
WHERE id = (
    SELECT id
    FROM pgq_job
    WHERE execute_at <= now()
    ORDER BY priority DESC, created_at
    FOR UPDATE SKIP LOCKED
    LIMIT 1
)
RETURNING *;

As soon as this query runs, the task is unable to be claimed by other workers. When the transaction commits, the task will be deleted. If the transaction rolls back or the worker crashes, the task will immediately become available for another worker.

To achieve at-least-once delivery, we begin a transaction, process the task, then commit the transaction. For at-most-once, we claim the task and immediately commit the transaction, then process the task. For tasks that don't have any external effects and only do database work, the at-least-once behavior is actually exactly-once (because both the claiming of the job and the database work will commit or rollback together).

Comparison to Celery

mishka queue fills the same role as Celery. You must use postgres as the backend and the library is small enough that you can read and understand all the code.

A note on the use of AtLeastOnceQueue and Django's transaction.on_commit()

A failure in an on_commit() callback will not cause that job to be retried when using an AtLeastOnceQueue (usually a job in an AtLeastOnceQueue queue will remain in the queue if the job fails). This is because on_commit() callbacks are executed after the transaction has been committed and, for mishka-queue, the job is removed from the queue when the transaction commits.

If you require more certainty that the code in an on_commit() callback is executed successfully, you may need to ensure it is idempotent and call it from within the job rather than using on_commit().

Usage

Requirements

mishka-queue is tested against python 3.9+, at least postgres 9.5 and at least Django 3.2.

Installation

Install with pip:

pip install mishka-queue

Then add 'pgq' to your INSTALLED_APPS. Run manage.py migrate to create the jobs table.

Instantiate a queue object. This can go wherever you like and be named whatever you like. For example, someapp/queue.py:

from pgq.queue import AtLeastOnceQueue

queue = AtLeastOnceQueue(
    tasks={
        # ...
    },
    queue='my-queue',
    notify_channel='my-queue',
)

You will need to import this queue instance to queue or process tasks. Use AtLeastOnceQueue for at-least-once delivery, or AtMostOnceQueue for at-most-once delivery.

mishka-queue comes with a management command base class that you can use to consume your tasks. It can be called whatever you like, for example in a someapp/managment/commands/worker.py:

from pgq.commands import Worker

from someapp.queue import queue

class Command(Worker):
    queue = queue

Then you can run manage.py worker to start your worker.

A task function takes two arguments -- the queue instance in use, and the Job instance for this task. The function can be defined anywhere and called whatever you like. Here's an example:

from pgq.decorators import task

from .queues import queue

@task(queue)
def debug_task(queue, job, args, meta):
    print(args)

Instead of using the task decorator, you can manually register it as a task. Add it to your queue instance when it is being created:

queue = AtLeastOnceQueue(tasks={
    'debug_task': debug_task,
}, queue='my-queue')

The key is the task name, used to queue the task. It doesn't have to match the function name.

To queue the task, if you used the task decorator you may:

debug_task.enqueue({'some_args': 0})

To manually queue the task, use the enqueue method on your queue instance:

queue.enqueue('debug_task', {'some_args': 0})

Assuming you have a worker running for this queue, the task will be run immediately. The second argument must be a single json-serializeable value and will be available to the task as job.args.

Tasks registered using the @task decorator will only be available on the queue if the file in which the task is defined has been imported. If your worker doesn't import the file containing the @task decorators somewhere, the tasks will not be available for dispatch. Importing files in the apps.py AppConfig.ready() method will ensure that the tasks are always available on the queue without having to import them in your worker just for the import side effects.

# Contents of someapp/apps.py
from django.apps import AppConfig

class SomeAppAppConfig(AppConfig):
    def ready(self):
        # Tasks registered with @task are defined in this import
        import someapp.tasks

Multiple Queues

You may run multiple queues and workers may each listen to a queue. You can have multiple workers listening to the same queue too. A queue is implemented as a CharField in the database. The queue would simply filter for jobs matching its queue name.

Bulk Enqueue

Many jobs can be efficiently created using bulk_enqueue() which accepts one task name for all the jobs being created and a list of dictionaries containing args for the task to execute with and, optionally, priority and execute_at for that particular job.

queue.bulk_enqueue(
    'debug_task',
    [
        {'args': {'some_args': 0}},
        {
            'args': {'some_args': 10}
            'priority': 10,
            'execute_at': timezone.now() + timedelta(days=1),
        },
    ]
)

Monitoring

Tasks are just database rows stored in the pgq_job table, so you can monitor the system with SQL.

To get a count of current tasks:

SELECT queue, count(*) FROM pgq_job WHERE execute_at <= now() GROUP BY queue

This will include both tasks ready to process and tasks currently being processed. To see tasks currently being processed, we need visibility into postgres row locks. This can be provided by the pgrowlocks extension. Once installed, this query will count currently-running tasks:

SELECT queue, count(*)
FROM pgrowlocks('pgq_job')
WHERE 'For Update' = ANY(modes)
GROUP BY queue;

You could join the results of pgrowlocks with pgq_job to get the full list of tasks in progress if you want.

Logging

mishka-queue logs through Python's logging framework, so can be configured with the LOGGING dict in your Django settings. It will not log anything under the default config, so be sure to configure some form of logging. Everything is logged under the pgq namespace. Here is an example configuration that will log INFO level messages to stdout:

LOGGING = {
    'version': 1,
    'root': {
        'level': 'DEBUG',
        'handlers': ['console'],
    },
    'formatters': {
        'verbose': {
            'format': '%(levelname)s %(asctime)s %(module)s %(process)d %(thread)d %(message)s',
        },
    },
    'handlers': {
        'console': {
            'level': 'INFO',
            'class': 'logging.StreamHandler',
            'formatter': 'verbose',
        },
    },
    'loggers': {
        'pgq': {
            'handlers': ['console'],
            'level': 'INFO',
            'propagate': False,
        },
    }
}

It would also be sensible to log WARNING and higher messages to something like Sentry:

LOGGING = {
    'version': 1,
    'root': {
        'level': 'INFO',
        'handlers': ['sentry', 'console'],
    },
    'formatters': {
        'verbose': {
            'format': '%(levelname)s %(asctime)s %(module)s %(process)d %(thread)d %(message)s',
        },
    },
    'handlers': {
        'console': {
            'level': 'INFO',
            'class': 'logging.StreamHandler',
            'formatter': 'verbose',
        },
        'sentry': {
            'level': 'WARNING',
            'class': 'raven.contrib.django.handlers.SentryHandler',
        },
    },
    'loggers': {
        'pgq': {
            'level': 'INFO',
            'handlers': ['console', 'sentry'],
            'propagate': False,
        },
    },
}

You could also log to a file by using the built-in logging.FileHandler.

Useful Recipes

These recipes aren't officially supported features of mishka-queue. We provide them so that you can mimick some of the common features in other task queues.

QUEUE_ALWAYS_EAGER

The queues in this library allow you to use the QUEUE_ALWAYS_EAGER setting to run a task immediately, without queueing it for a worker. It could be used during tests, and while debugging in a development environment with any workers turned off.

It is similar in behaviour to CELERY_ALWAYS_EAGER setting in Celery.