User guide

Installation

The preferred installation method is by using pip:

$ pip install apscheduler

If you don’t have pip installed, you need to install that first.

Code examples

The source distribution contains the examples directory where you can find many working examples for using APScheduler in different ways. The examples can also be browsed online.

Introduction

The core concept of APScheduler is to give the user the ability to queue Python code to be executed, either as soon as possible, later at a given time, or on a recurring schedule.

The scheduler is the user-facing interface of the system. When it’s running, it does two things concurrently. The first is processing schedules. From its data store, it fetches schedules due to be run. For each such schedule, it then uses the schedule’s trigger to calculate run times up to the present. The scheduler then creates one or more jobs (controllable by configuration) based on these run times and adds them to the data store.

The second role of the scheduler is running jobs The scheduler asks the data store for jobs, and then starts running those jobs. If the data store signals that it has new jobs, the scheduler will try to acquire those jobs if it is capable of accommodating more. When a scheduler completes a job, it will then also ask the data store for as many more jobs as it can handle.

By default, schedulers operate in both of these roles, but can be configured to only process schedules or run jobs if deemed necessary. It may even be desirable to use the scheduler only as an interface to an external data store while leaving schedule and job processing to other scheduler instances running elsewhere.

Basic concepts / glossary

These are the basic components and concepts of APScheduler which will be referenced later in this guide.

A callable is any object that returns True from callable(). These are:

  • A free function (def something(...): ...)

  • An instance method (class Foo: ... def something(self, ...): ...)

  • A class method (class Foo: ... @classmethod ... def something(cls, ...): ...)

  • A static method (class Foo: ... @staticmethod ... def something(...): ...)

  • A lambda (lambda a, b: a + b)

  • An instance of a class that contains a method named __call__)

A task encapsulates a callable and a number of configuration parameters. They are often implicitly defined as a side effect of the user creating a new schedule against a callable, but can also be explicitly defined beforehand.

A trigger contains the logic and state used to calculate when a scheduled task should be run.

A schedule combines a task with a trigger, plus a number of configuration parameters.

A job is request for a task to be run. It can be created automatically from a schedule when a scheduler processes it, or it can be directly created by the user if they directly request a task to be run.

A data store is used to store schedules and jobs, and to keep track of tasks.

A job executor runs the job, by calling the function associated with the job’s task. An executor could directly call the callable, or do it in another thread, subprocess or even some external service.

An event broker delivers published events to all interested parties. It facilitates the cooperation between schedulers by notifying them of new or updated schedules and jobs.

A scheduler is the main interface of this library. It houses both a data store and an event broker, plus one or more job executors. It contains methods users can use to work with tasks, schedules and jobs. Behind the scenes, it also processes due schedules, spawning jobs and updating the next run times. It also processes available jobs, making the appropriate job executors to run them, and then sending back the results to the data store.

Running the scheduler

The scheduler comes in two flavors: synchronous and asynchronous. The synchronous scheduler actually runs an asynchronous scheduler behind the scenes in a dedicated thread, so if your app runs on asyncio or Trio_, you should prefer the asynchronous scheduler.

The scheduler can run either in the foreground, blocking on a call to run_until_stopped(), or in the background where it does its work while letting the rest of the program run.

If the only intent of your program is to run scheduled tasks, then you should start the scheduler with run_until_stopped(). But if you need to do other things too, then you should call start_in_background() before running the rest of the program.

In almost all cases, the scheduler should be used as a context manager. This initializes the underlying data store and event broker, allowing you to use the scheduler for manipulating tasks, schedules and jobs prior to starting the processing of schedules and jobs. Exiting the context manager will shut down the scheduler and its underlying services. This mode of operation is mandatory for the asynchronous scheduler when running it in the background, but it is preferred for the synchronous scheduler too.

As a special consideration (for use with WSGI based web frameworks), the synchronous scheduler can be run in the background without being used as a context manager. In this scenario, the scheduler adds an atexit hook that will perform an orderly shutdown of the scheduler before the process terminates.

Warning

If you start the scheduler in the background and let the script finish execution, the scheduler will automatically shut down as well.

from apscheduler import Scheduler

with Scheduler() as scheduler:
    # Add schedules, configure tasks here
    scheduler.run_until_stopped()

Configuring tasks

In order to add schedules or jobs to the data store, you need to have a task that defines which callable will be called when each job is run.

In most cases, you don’t need to go through this step, and instead have a task implicitly created for you by the methods that add `schedules or jobs.

Explicitly configuring a task is generally only necessary in the following cases:

  • You need to have more than one task with the same callable

  • You need to set any of the task settings to non-default values

  • You need to add schedules/jobs targeting lambdas, nested functions or instances of unserializable classes

Scheduling tasks

To create a schedule for running a task, you need, at the minimum:

If you’ve configured a task (as per the previous section), you can pass the task object or its ID to Scheduler.add_schedule(). As a shortcut, you can pass a callable instead, in which case a task will be automatically created for you if necessary.

If the callable you’re trying to schedule is either a lambda or a nested function, then you need to explicitly create a task beforehand, as it is not possible to create a reference (package.module:varname) to these types of callables.

The trigger determines the scheduling logic for your schedule. In other words, it is used to calculate the datetimes on which the task will be run. APScheduler comes with a number of built-in trigger classes:

  • DateTrigger: use when you want to run the task just once at a certain point of time

  • IntervalTrigger: use when you want to run the task at fixed intervals of time

  • CronTrigger: use when you want to run the task periodically at certain time(s) of day

  • CalendarIntervalTrigger: use when you want to run the task on calendar-based intervals, at a specific time of day

Combining multiple triggers

Occasionally, you may find yourself in a situation where your scheduling needs are too complex to be handled with any of the built-in triggers directly.

One examples of such a need would be when you want the task to run at 10:00 from Monday to Friday, but also at 11:00 from Saturday to Sunday. A single CronTrigger would not be able to handle this case, but an OrTrigger containing two cron triggers can:

from apscheduler.triggers.combining import OrTrigger
from apscheduler.triggers.cron import CronTrigger

trigger = OrTrigger(
    CronTrigger(day_of_week="mon-fri", hour=10),
    CronTrigger(day_of_week="sat-sun", hour=11),
)

On the first run, OrTrigger generates the next run times from both cron triggers and saves them internally. It then returns the earliest one. On the next run, it generates a new run time from the trigger that produced the earliest run time on the previous run, and then again returns the earliest of the two run times. This goes on until all the triggers have been exhausted, if ever.

Another example would be a case where you want the task to be run every 2 months at 10:00, but not on weekends (Saturday or Sunday):

from apscheduler.triggers.calendarinterval import CalendarIntervalTrigger
from apscheduler.triggers.combining import AndTrigger
from apscheduler.triggers.cron import CronTrigger

trigger = AndTrigger(
    CalendarIntervalTrigger(months=2, hour=10),
    CronTrigger(day_of_week="mon-fri", hour=10),
)

On the first run, AndTrigger generates the next run times from both the CalendarIntervalTrigger and CronTrigger. If the run times coincide, it will return that run time. Otherwise, it will calculate a new run time from the trigger that produced the earliest run time. It will keep doing this until a match is found, one of the triggers has been exhausted or the maximum number of iterations (1000 by default) is reached.

If this trigger is created on 2022-06-07 at 09:00:00, its first run times would be:

  • 2022-06-07 10:00:00

  • 2022-10-07 10:00:00

  • 2022-12-07 10:00:00

Notably, 2022-08-07 is skipped because it falls on a Sunday.

Running tasks without scheduling

In some cases, you want to run tasks directly, without involving schedules:

  • You’re only interested in using the scheduler system as a job queue

  • You’re interested in the job’s return value

To queue a job and wait for its completion and get the result, the easiest way is to use run_job(). If you prefer to just launch a job and not wait for its result, use:meth:~Scheduler.add_job instead. If you want to get the results later, you need to pass an appropriate result_expiration_time parameter to add_job() so that the result is saved. Then, you can call get_job_result() with the job ID you got from add_job() to retrieve the result.

Removing schedules

To remove a previously added schedule, call remove_schedule(). Pass the identifier of the schedule you want to remove as an argument. This is the ID you got from add_schedule().

Note that removing a schedule does not cancel any jobs derived from it, but does prevent further jobs from being created from that schedule.

Limiting the number of concurrently executing instances of a job

It is possible to control the maximum number of concurrently running jobs for a particular task. By default, only one job is allowed to be run for every task. This means that if the job is about to be run but there is another job for the same task still running, the later job is terminated with the outcome of missed_start_deadline.

To allow more jobs to be concurrently running for a task, pass the desired maximum number as the max_running_jobs keyword argument to add_schedule().

Controlling how much a job can be started late

Some tasks are time sensitive, and should not be run at all if they fail to be started on time (like, for example, if the scheduler(s) were down while they were supposed to be running the scheduled jobs). You can control this time limit with the misfire_grace_time option passed to add_schedule(). A scheduler that acquires the job then checks if the current time is later than the deadline (run time + misfire grace time) and if it is, it skips the execution of the job and releases it with the outcome of missed_start_deadline.

Controlling how jobs are queued from schedules

In most cases, when a scheduler processes a schedule, it queues a new job using the run time currently marked for the schedule. Then it updates the next run time using the schedule’s trigger and releases the schedule back to the data store. But sometimes a situation occurs where the schedule did not get processed often or quickly enough, and one or more next run times produced by the trigger are actually in the past.

In a situation like that, the scheduler needs to decide what to do: to queue a job for every run time produced, or to coalesce them all into a single job, effectively just kicking off a single job. To control this, pass the coalesce argument to add_schedule().

The possible values are:

  • latest: queue exactly one job, using the latest run time as the designated run time

  • earliest: queue exactly one job, using the earliest run time as the designated run time

  • all: queue one job for each of the calculated run times

The biggest difference between the first two options is how the designated run time, and by extension, the starting deadline for the job is selected. With the first option, the job is less likely to be skipped due to being started late since the latest of all the collected run times is used for the deadline calculation.

As explained in the previous section, the starting deadline is misfire grace time affects the newly queued job.

Context variables

Schedulers provide certain context variables available to the tasks being run:

Here’s an example:

from apscheduler import current_job

def my_task_function():
    job_info = current_job.get().id
    print(
        f"This is job {job_info.id} and was spawned from schedule "
        f"{job_info.schedule_id}"
    )

Subscribing to events

Schedulers have the ability to notify listeners when some event occurs in the scheduler system. Examples of such events would be schedulers or workers starting up or shutting down, or schedules or jobs being created or removed from the data store.

To listen to events, you need a callable that takes a single positional argument which is the event object. Then, you need to decide which events you’re interested in:

from apscheduler import Event, JobAcquired, JobReleased

def listener(event: Event) -> None:
    print(f"Received {event.__class__.__name__}")

scheduler.subscribe(listener, {JobAcquired, JobReleased})

This example subscribes to the JobAcquired and JobReleased event types. The callback will receive an event of either type, and prints the name of the class of the received event.

Asynchronous schedulers and workers support both synchronous and asynchronous callbacks, but their synchronous counterparts only support synchronous callbacks.

When distributed event brokers (that is, other than the default one) are being used, events other than the ones relating to the life cycles of schedulers and workers, will be sent to all schedulers and workers connected to that event broker.

Clean-up of expired jobs and schedules

Expired job results and finished schedules are, by default, automatically cleaned up by each running scheduler on 15 minute intervals (counting from the scheduler’s start time). This can be adjusted (or disabled entirely) through the cleanup_interval configuration option.

Deployment

Using persistent data stores

The default data store, MemoryDataStore, stores data only in memory so all the schedules and jobs that were added to it will be erased if the process crashes.

When you need your schedules and jobs to survive the application shutting down, you need to use a persistent data store. Such data stores do have additional considerations, compared to the memory data store:

  • Task arguments must be serializable

  • You must either trust the data store, or use an alternate serializer

  • A conflict policy and an explicit identifier must be defined for schedules that are added at application startup

These requirements warrant some explanation. The first point means that since persisting data means saving it externally, either in a file or sending to a database server, all the objects involved are converted to bytestrings. This process is called serialization. By default, this is done using pickle, which guarantees the best compatibility but is notorious for being vulnerable to simple injection attacks. This brings us to the second point. If you cannot be sure that nobody can maliciously alter the externally stored serialized data, it would be best to use another serializer. The built-in alternatives are:

The former requires the cbor2 library, but supports a wider variety of types natively. The latter has no dependencies but has very limited support for different types.

The third point relates to situations where you’re essentially adding the same schedule to the data store over and over again. If you don’t specify a static identifier for the schedules added at the start of the application, you will end up with an increasing number of redundant schedules doing the same thing, which is probably not what you want. To that end, you will need to come up with some identifying name which will ensure that the same schedule will not be added over and over again (as data stores are required to enforce the uniqueness of schedule identifiers). You’ll also need to decide what to do if the schedule already exists in the data store (that is, when the application is started the second time) by passing the conflict_policy argument. Usually you want the replace option, which replaces the existing schedule with the new one.

See also

You can find practical examples of persistent data stores in the examples/standalone directory (async_postgres.py and async_mysql.py).

Using multiple schedulers

There are several situations in which you would want to run several schedulers against the same data store at once:

  • Running a server application (usually a web app) with multiple worker processes

  • You need fault tolerance (scheduling will continue even if a node or process running a scheduler goes down)

When you have multiple schedulers running at once, they need to be able to coordinate their efforts so that the schedules don’t get processed more than once and the schedulers know when to wake up even if another scheduler added the next due schedule to the data store. To this end, a shared event broker must be configured.

See also

You can find practical examples of data store sharing in the examples/web directory.

Using a scheduler without running it

Some deployment scenarios may warrant the use of a scheduler for only interfacing with an external data store, for things like configuring tasks, adding schedules or queuing jobs. One such practical use case is a web application that needs to run heavy computations elsewhere so they don’t cause performance issues with the web application itself.

You can then run one or more schedulers against the same data store and event broker elsewhere where they don’t disturb the web application. These schedulers will do all the heavy lifting like processing schedules and running jobs.

See also

A practical example of this separation of concerns can be found in the examples/separate_worker directory.

Explicitly assigning an identity to the scheduler

If you’re running one or more schedulers against a persistent data store in a production setting, it’d be wise to assign each scheduler a custom identity. The reason for this is twofold:

  1. It helps you figure out which jobs are being run where

  2. It allows crashed jobs to cleared out quicker, as other schedulers aren’t allowed to clean them up until the jobs’ timeouts expire

The best choice would be something that the environment guarantees to be unique among all the scheduler instances but stays the same when the scheduler instance is restarted. For example, on Kubernetes, this would be the name of the pod where the scheduler is running, assuming of course that there is only one scheduler running in each pod against the same data store.

Of course, if you’re only ever running one scheduler against a persistent data store, you can just use a static scheduler ID.

If no ID is explicitly given, the scheduler generates an ID by concatenating the following:

  • the current host name

  • the current process ID

  • the ID of the scheduler instance

Troubleshooting

If something isn’t working as expected, it will be helpful to increase the logging level of the apscheduler logger to the DEBUG level.

If you do not yet have logging enabled in the first place, you can do this:

import logging

logging.basicConfig()
logging.getLogger('apscheduler').setLevel(logging.DEBUG)

This should provide lots of useful information about what’s going on inside the scheduler and/or worker.

Also make sure that you check the Frequently Asked Questions section to see if your problem already has a solution.

Reporting bugs

A bug tracker is provided by GitHub.

Getting help

If you have problems or other questions, you can either: