Migrating from previous versions of APScheduler

From v3.x to v4.0

APScheduler 4.0 has undergone a partial rewrite since the 3.x series.

There is currently no way to automatically import schedules from a persistent 3.x job store, but this shortcoming will be rectified before the final v4.0 release.

Terminology and architectural design changes

The concept of a job has been split into Task, Schedule and Job. See the documentation of each class (and read the tutorial) to understand their roles.

Data stores, previously called job stores, have been redesigned to work with multiple running schedulers and workers, both for purposes of scalability and fault tolerance. Many data store implementations were dropped because they were either too burdensome to support, or the backing services were not sophisticated enough to handle the increased requirements.

Event brokers are a new component in v4.0. They relay events between schedulers and workers, enabling them to work together with a shared data store. External (as opposed to local) event broker services are required in multi-node or multi-process deployment scenarios.

Triggers are now stateful. This change was found to be necessary to properly support combining triggers (AndTrigger and OrTrigger), as they needed to keep track of the next run times of all the triggers contained within. This change also enables some more sophisticated custom trigger implementations.

Time zone support has been revamped to use zoneinfo (or backports.zoneinfo on Python versions earlier than 3.9) zones instead of pytz zones. You should not use pytz with APScheduler anymore.

Entry points are no longer used or supported, as they were more trouble than they were worth, particularly with packagers like py2exe or PyInstaller which by default did not package distribution metadata. Thus, triggers and data stores have to be explicitly instantiated.

Scheduler changes

The add_job() method is now add_schedule(). The scheduler still has a method named add_job(), but this is meant for making one-off runs of a task. Previously you would have had to call add_job() with a DateTrigger using the current time as the run time.

The two most commonly used schedulers, BlockingScheduler and BackgroundScheduler, have often caused confusion among users and have thus been combined into Scheduler. This new unified scheduler class has two methods that replace the start() method used previously: run_until_stopped() and start_in_background(). The former should be used if you previously used BlockingScheduler, and the latter if you used BackgroundScheduler.

The asyncio scheduler has been replaced with a more generic AsyncScheduler, which is based on AnyIO and thus also supports Trio in addition to asyncio. The API of the async scheduler differs somewhat from its synchronous counterpart. In particular, it requires itself to be used as an async context manager – whereas with the synchronous scheduler, use as a context manager is recommended but not required.

All other scheduler implementations have been dropped because they were either too burdensome to support, or did not seem necessary anymore. Some of the dropped implementations (particularly Qt) are likely to be re-added before v4.0 final.

Schedulers no longer support multiple data stores. If you need this capability, you should run multiple schedulers instead.

Configuring and running the scheduler has been radically simplified. The configure() method is gone, and all configuration is now passed as keyword arguments to the scheduler class.

Trigger changes

As the scheduler is no longer used to create triggers, any supplied datetimes will be assumed to be in the local time zone. If you wish to change the local time zone, you should set the TZ environment variable to either the name of the desired timezone (e.g. Europe/Helsinki) or to a path of a time zone file. See the tzlocal documentation for more information.

Jitter support has been moved from individual triggers to the schedule level. This not only simplified trigger design, but also enabled the scheduler to provide information about the randomized jitter and the original run time to the user.

CronTrigger was changed to respect the standard order of weekdays, so that Sunday is now 0 and Saturday is 6. If you used numbered weekdays before, you must change your trigger configuration to match. If in doubt, use abbreviated weekday names (e.g. sun, fri) instead.

IntervalTrigger was changed to start immediately, instead of waiting for the first interval to pass. If you have workarounds in place to “fix” the previous behavior, you should remove them.

From v3.0 to v3.2

Prior to v3.1, the scheduler inadvertently exposed the ability to fetch and manipulate jobs before the scheduler had been started. The scheduler now requires you to call scheduler.start() before attempting to access any of the jobs in the job stores. To ensure that no old jobs are mistakenly executed, you can start the scheduler in paused mode (scheduler.start(paused=True)) (introduced in v3.2) to avoid any premature job processing.

From v2.x to v3.0

The 3.0 series is API incompatible with previous releases due to a design overhaul.

Scheduler changes

  • The concept of “standalone mode” is gone. For standalone=True, use BlockingScheduler instead, and for standalone=False, use BackgroundScheduler. BackgroundScheduler matches the old default semantics.

  • Job defaults (like misfire_grace_time and coalesce) must now be passed in a dictionary as the job_defaults option to configure(). When supplying an ini-style configuration as the first argument, they will need a corresponding job_defaults. prefix.

  • The configuration key prefix for job stores was changed from jobstore. to jobstores. to match the dict-style configuration better.

  • The max_runs option has been dropped since the run counter could not be reliably preserved when replacing a job with another one with the same ID. To make up for this, the end_date option was added to cron and interval triggers.

  • The old thread pool is gone, replaced by ThreadPoolExecutor. This means that the old threadpool options are no longer valid. See scheduler-config on how to configure executors.

  • The trigger-specific scheduling methods have been removed entirely from the scheduler. Use the generic add_job() method or the scheduled_job() decorator instead. The signatures of these methods were changed significantly.

  • The shutdown_threadpool and close_jobstores options have been removed from the shutdown() method. Executors and job stores are now always shut down on scheduler shutdown.

  • unschedule_job() and unschedule_func() have been replaced by remove_job(). You can also unschedule a job by using the job handle returned from add_job().

Job store changes

The job store system was completely overhauled for both efficiency and forwards compatibility. Unfortunately, this means that the old data is not compatible with the new job stores. If you need to migrate existing data from APScheduler 2.x to 3.x, contact the APScheduler author.

The Shelve job store had to be dropped because it could not support the new job store design. Use SQLAlchemyJobStore with SQLite instead.

Trigger changes

From 3.0 onwards, triggers now require a pytz timezone. This is normally provided by the scheduler, but if you were instantiating triggers manually before, then one must be supplied as the timezone argument.

The only other backwards incompatible change was that get_next_fire_time() takes two arguments now: the previous fire time and the current datetime.

From v1.x to 2.0

There have been some API changes since the 1.x series. This document explains the changes made to v2.0 that are incompatible with the v1.x API.

API changes

  • The behavior of cron scheduling with regards to default values for omitted fields has been made more intuitive – omitted fields lower than the least significant explicitly defined field will default to their minimum values except for the week number and weekday fields

  • SchedulerShutdownError has been removed – jobs are now added tentatively and scheduled for real when/if the scheduler is restarted

  • Scheduler.is_job_active() has been removed – use job in scheduler.get_jobs() instead

  • dump_jobs() is now print_jobs() and prints directly to the given file or sys.stdout if none is given

  • The repeat parameter was removed from add_interval_job() and interval_schedule() in favor of the universal max_runs option

  • unschedule_func() now raises a KeyError if the given function is not scheduled

  • The semantics of shutdown() have changed – the method no longer accepts a numeric argument, but two booleans

Configuration changes

  • The scheduler can no longer be reconfigured while it’s running