Deferrable Operators & Triggers¶

Standard Operators and Sensors take up a full worker slot for the entire time they are running, even if they are idle; for example, if you only have 100 worker slots available to run Tasks, and you have 100 DAGs waiting on a Sensor that's currently running but idle, then you cannot run anything else - even though your entire Airflow cluster is essentially idle. reschedule mode for Sensors solves some of this, allowing Sensors to only run at fixed intervals, but it is inflexible and only allows using time as the reason to resume, not anything else.

This is where Deferrable Operators come in. A deferrable operator is one that is written with the ability to suspend itself and free up the worker when it knows it has to wait, and hand off the job of resuming it to something called a Trigger. As a result, while it is suspended (deferred), it is not taking up a worker slot and your cluster will have a lot less resources wasted on idle Operators or Sensors.

Triggers are small, asynchronous pieces of Python code designed to be run all together in a single Python process; because they are asynchronous, they are able to all co-exist efficiently. As an overview of how this process works:

A task instance (running operator) gets to a point where it has to wait, and defers itself with a trigger tied to the event that should resume it. This frees up the worker to run something else.
The new Trigger instance is registered inside Airflow, and picked up by a triggerer process
The trigger is run until it fires, at which point its source task is re-scheduled
The scheduler queues the task to resume on a worker node

Using deferrable operators as a DAG author is almost transparent; writing them, however, takes a bit more work.

Note

Deferrable Operators & Triggers rely on more recent asyncio features, and as a result only work on Python 3.7 or higher.

Using Deferrable Operators¶

If all you wish to do is use pre-written Deferrable Operators (such as TimeSensorAsync, which comes with Airflow), then there are only two steps you need:

Ensure your Airflow installation is running at least one triggerer process, as well as the normal scheduler
Use deferrable operators/sensors in your DAGs

That's it; everything else will be automatically handled for you. If you're upgrading existing DAGs, we even provide some API-compatible sensor variants (e.g. TimeSensorAsync for TimeSensor) that you can swap into your DAG with no other changes required.

Note that you cannot yet use the deferral ability from inside custom PythonOperator/TaskFlow Python functions; it is only available to traditional, class-based Operators at the moment.

Writing Deferrable Operators¶

Writing a deferrable operator takes a bit more work. There are some main points to consider:

Your Operator must defer itself based on a Trigger. If there is a Trigger in core Airflow you can use, great; otherwise, you will have to write one.
Your Operator will be stopped and removed from its worker while deferred, and no state will persist automatically. You can persist state by asking Airflow to resume you at a certain method or pass certain kwargs, but that's it.
You can defer multiple times, and you can defer before/after your Operator does significant work, or only defer if certain conditions are met (e.g. a system does not have an immediate answer). Deferral is entirely under your control.
Any Operator can defer; no special marking on its class is needed, and it's not limited to Sensors.

Triggering Deferral¶

If you want to trigger deferral, at any place in your Operator you can call self.defer(trigger, method_name, kwargs, timeout), which will raise a special exception that Airflow will catch. The arguments are:

trigger: An instance of a Trigger that you wish to defer on. It will be serialized into the database.
method_name: The method name on your Operator you want Airflow to call when it resumes, other than execute.
kwargs: Additional keyword arguments to pass to the method when it is called. Optional, defaults to {}.
timeout: A timedelta that specifies a timeout after which this deferral will fail, and fail the task instance. Optional, defaults to None, meaning no timeout.

When you opt to defer, your Operator will stop executing at that point and be removed from its current worker. No state - such as local variables, or attributes set on self - will persist, and when your Operator is resumed it will be a brand new instance of it. The only way you can pass state from the old instance of the Operator to the new one is via method_name and kwargs.

When your Operator is resumed, you will find an event item added to the kwargs passed to it, which contains the payload from the trigger event that resumed your Operator. Depending on the trigger, this may be useful to your operator (e.g. it's a status code or URL to fetch results), or it may not be important (it's just a datetime). Your method_name method, however, must accept event as a keyword argument.

If your Operator returns from either its first execute() method when it's new, or a subsequent method specified by method_name, it will be considered complete and will finish executing.

You are free to set method_name to execute if you want your Operator to have one entrypoint, but it, too, will have to accept event as an optional keyword argument.

Here's a basic example of how a sensor might trigger deferral:

class WaitOneHourSensor(BaseSensorOperator):
    def execute(self, context):
        self.defer(trigger=TimeDeltaTrigger(timedelta(hours=1), method_name="execute_complete")

    def execute_complete(self, context, event=None):
        # We have no more work to do here. Mark as complete.
        return

This Sensor is literally just a thin wrapper around the Trigger, so all it does is defer to the trigger, and specify a different method to come back to when the trigger fires - which, as it returns immediately, marks the Sensor as successful.

Under the hood, self.defer raises the TaskDeferred exception, so it will work anywhere inside your Operator's code, even buried many nested calls deep inside execute(). You are free to raise TaskDeferred manually if you wish; it takes the same arguments as self.defer.

Note that execution_timeout on Operators is considered over the total runtime, not individual executions in-between deferrals - this means that if execution_timeout is set, an Operator may fail while it's deferred or while it's running after a deferral, even if it's only been resumed for a few seconds.

Writing Triggers¶

A Trigger is written as a class that inherits from BaseTrigger, and implements three methods:

__init__, to receive arguments from Operators instantiating it
run, an asynchronous method that runs its logic and yields one or more TriggerEvent instances as an asynchronous generator
serialize, which returns the information needed to re-construct this trigger, as a tuple of the classpath, and keyword arguments to pass to __init__

There's also some design constraints to be aware of:

The run method must be asynchronous (using Python's asyncio), and correctly await whenever it does a blocking operation.
run must yield its TriggerEvents, not return them. If it returns before yielding at least one event, Airflow will consider this an error and fail any Task Instances waiting on it. If it throws an exception, Airflow will also fail any dependent task instances.
A Trigger must be able to run in parallel with other copies of itself. This can happen both when two tasks defer based on the same trigger, and also if a network partition happens and Airflow re-launches a trigger on a separated machine.
When events are emitted, and if your trigger is designed to emit more than one event, they must contain a payload that can be used to deduplicate events if the trigger is being run in multiple places. If you only fire one event, and don't want to pass information in the payload back to the Operator that deferred, you can just set the payload to None.
A trigger may be suddenly removed from one process and started on a new one (if partitions are being changed, or a deployment is happening). You may provide an optional cleanup method that gets called when this happens.

Here's the structure of a basic Trigger:

class DateTimeTrigger(BaseTrigger):

    def __init__(self, moment):
        super().__init__()
        self.moment = moment

    def serialize(self):
        return ("airflow.triggers.temporal.DateTimeTrigger", {"moment": self.moment})

    async def run(self):
        while self.moment > timezone.utcnow():
            await asyncio.sleep(1)
        yield TriggerEvent(self.moment)

This is a very simplified version of Airflow's DateTimeTrigger, and you can see several things here:

__init__ and serialize are written as a pair; the Trigger is instantiated once when it is submitted by the Operator as part of its deferral request, then serialized and re-instantiated on any triggerer process that runs the trigger.
The run method is declared as an async def, as it must be asynchronous, and uses asyncio.sleep rather than the regular time.sleep (as that would block the process).
When it emits its event it packs self.moment in there, so if this trigger is being run redundantly on multiple hosts, the event can be de-duplicated.

Triggers can be as complex or as simple as you like provided you keep inside this contract; they are designed to be run in a highly-available fashion, auto-distributed among hosts running the triggerer. We encourage you to avoid any kind of persistent state in a trigger; they should get everything they need from their __init__, so they can be serialized and moved around freely.

If you are new to writing asynchronous Python, you should be very careful writing your run() method; Python's async model means that any code that does not correctly await when it does a blocking operation will block the entire process. Airflow will attempt to detect this and warn you in the triggerer logs when it happens, but we strongly suggest you set the variable PYTHONASYNCIODEBUG=1 when you are writing your Trigger to enable extra checks from Python to make sure you're writing non-blocking code. Be especially careful when doing filesystem calls, as if the underlying filesystem is network-backed it may be blocking.

Right now, Triggers are only used up to their first event, as they are only used for resuming deferred tasks (which happens on the first event fired). However, we plan to allow DAGs to be launched from triggers in future, which is where multi-event triggers will be more useful.

High Availability¶

Triggers are designed from the ground-up to be highly-available; if you want to run a highly-available setup, simply run multiple copies of triggerer on multiple hosts. Much like scheduler, they will automatically co-exist with correct locking and HA.

Depending on how much work the triggers are doing, you can fit from hundreds to tens of thousands of triggers on a single triggerer host. By default, every triggerer will have a capacity of 1000 triggers it will try to run at once; you can change this with the --capacity argument. If you have more triggers trying to run than you have capacity across all of your triggerer processes, some triggers will be delayed from running until others have completed.

Airflow tries to only run triggers in one place at once, and maintains a heartbeat to all triggerers that are currently running. If a triggerer dies, or becomes partitioned from the network where Airflow's database is running, Airflow will automatically re-schedule triggers that were on that host to run elsewhere (after waiting 30 seconds for the machine to re-appear).

This means it's possible, but unlikely, for triggers to run in multiple places at once; this is designed into the Trigger contract, however, and entirely expected. Airflow will de-duplicate events fired when a trigger is running in multiple places simultaneously, so this process should be transparent to your Operators.

Note that every extra triggerer you run will result in an extra persistent connection to your database.

Smart Sensors¶

Deferrable Operators essentially supersede Smart Sensors, and should be preferred for almost all situations. They do solve fundamentally the same problem; Smart Sensors, however, only work for certain Sensor workload styles, have no redundancy, and require a custom DAG to run at all times.