Concepts

The Airflow Platform is a tool for describing, executing, and monitoring workflows.

Core Ideas

DAGs

In Airflow, a DAG – or a Directed Acyclic Graph – is a collection of all the tasks you want to run, organized in a way that reflects their relationships and dependencies.

For example, a simple DAG could consist of three tasks: A, B, and C. It could say that A has to run successfully before B can run, but C can run anytime. It could say that task A times out after 5 minutes, and B can be restarted up to 5 times in case it fails. It might also say that the workflow will run every night at 10pm, but shouldn’t start until a certain date.

In this way, a DAG describes how you want to carry out your workflow; but notice that we haven’t said anything about what we actually want to do! A, B, and C could be anything. Maybe A prepares data for B to analyze while C sends an email. Or perhaps A monitors your location so B can open your garage door while C turns on your house lights. The important thing is that the DAG isn’t concerned with what its constituent tasks do; its job is to make sure that whatever they do happens at the right time, or in the right order, or with the right handling of any unexpected issues.

DAGs are defined in standard Python files that are placed in Airflow’s DAG_FOLDER. Airflow will execute the code in each file to dynamically build the DAG objects. You can have as many DAGs as you want, each describing an arbitrary number of tasks. In general, each one should correspond to a single logical workflow.

Scope

Airflow will load any DAG object it can import from a DAGfile. Critically, that means the DAG must appear in globals(). Consider the following two DAGs. Only dag_1 will be loaded; the other one only appears in a local scope.

dag_1 = DAG('this_dag_will_be_discovered')

def my_function()
    dag_2 = DAG('but_this_dag_will_not')

my_function()

Sometimes this can be put to good use. For example, a common pattern with SubDagOperator is to define the subdag inside a function so that Airflow doesn’t try to load it as a standalone DAG.

Default Arguments

If a dictionary of default_args is passed to a DAG, it will apply them to any of its operators. This makes it easy to apply a common parameter to many operators without having to type it many times.

default_args=dict(
    start_date=datetime(2016, 1, 1),
    owner='Airflow')

dag = DAG('my_dag', default_args=default_args)
op = DummyOperator(task_id='dummy', dag=dag)
print(op.owner) # Airflow

Context Manager

Added in Airflow 1.8

DAGs can be used as context managers to automatically assign new operators to that DAG.

with DAG('my_dag', start_date=datetime(2016, 1, 1)) as dag:
    op = DummyOperator('op')

op.dag is dag # True

Operators

While DAGs describe how to run a workflow, Operators determine what actually gets done.

An operator describes a single task in a workflow. Operators are usually (but not always) atomic, meaning they can stand on their own and don’t need to share resources with any other operators. The DAG will make sure that operators run in the correct certain order; other than those dependencies, operators generally run independently. In fact, they may run on two completely different machines.

This is a subtle but very important point: in general, if two operators need to share information, like a filename or small amount of data, you should consider combining them into a single operator. If it absolutely can’t be avoided, Airflow does have a feature for operator cross-communication called XCom that is described elsewhere in this document.

Airflow provides operators for many common tasks, including:

  • BashOperator - executes a bash command
  • PythonOperator - calls an arbitrary Python function
  • EmailOperator - sends an email
  • HTTPOperator - sends an HTTP request
  • MySqlOperator, SqliteOperator, PostgresOperator, MsSqlOperator, OracleOperator, JdbcOperator, etc. - executes a SQL command
  • Sensor - waits for a certain time, file, database row, S3 key, etc…

In addition to these basic building blocks, there are many more specific operators: DockerOperator, HiveOperator, S3FileTransferOperator, PrestoToMysqlOperator, SlackOperator… you get the idea!

The airflow/contrib/ directory contains yet more operators built by the community. These operators aren’t always as complete or well-tested as those in the main distribution, but allow users to more easily add new functionality to the platform.

Operators are only loaded by Airflow if they are assigned to a DAG.

DAG Assignment

Added in Airflow 1.8

Operators do not have to be assigned to DAGs immediately (previously dag was a required argument). However, once an operator is assigned to a DAG, it can not be transferred or unassigned. DAG assignment can be done explicitly when the operator is created, through deferred assignment, or even inferred from other operators.

dag = DAG('my_dag', start_date=datetime(2016, 1, 1))

# sets the DAG explicitly
explicit_op = DummyOperator(task_id='op1', dag=dag)

# deferred DAG assignment
deferred_op = DummyOperator(task_id='op2')
deferred_op.dag = dag

# inferred DAG assignment (linked operators must be in the same DAG)
inferred_op = DummyOperator(task_id='op3')
inferred_op.set_upstream(deferred_op)

Bitshift Composition

Added in Airflow 1.8

Traditionally, operator relationships are set with the set_upstream() and set_downstream() methods. In Airflow 1.8, this can be done with the Python bitshift operators >> and <<. The following four statements are all functionally equivalent:

op1 >> op2
op1.set_downstream(op2)

op2 << op1
op2.set_upstream(op1)

When using the bitshift to compose operators, the relationship is set in the direction that the bitshift operator points. For example, op1 >> op2 means that op1 runs first and op2 runs second. Multiple operators can be composed – keep in mind the chain is executed left-to-right and the rightmost object is always returned. For example:

op1 >> op2 >> op3 << op4

is equivalent to:

op1.set_downstream(op2)
op2.set_downstream(op3)
op3.set_upstream(op4)

For convenience, the bitshift operators can also be used with DAGs. For example:

dag >> op1 >> op2

is equivalent to:

op1.dag = dag
op1.set_downstream(op2)

We can put this all together to build a simple pipeline:

with DAG('my_dag', start_date=datetime(2016, 1, 1)) as dag:
    (
        DummyOperator(task_id='dummy_1')
        >> BashOperator(
            task_id='bash_1',
            bash_command='echo "HELLO!"')
        >> PythonOperator(
            task_id='python_1',
            python_callable=lambda: print("GOODBYE!"))
    )

Tasks

Once an operator is instantiated, it is referred to as a “task”. The instantiation defines specific values when calling the abstract operator, and the parameterized task becomes a node in a DAG.

Task Instances

A task instance represents a specific run of a task and is characterized as the combination of a dag, a task, and a point in time. Task instances also have an indicative state, which could be “running”, “success”, “failed”, “skipped”, “up for retry”, etc.

Workflows

You’re now familiar with the core building blocks of Airflow. Some of the concepts may sound very similar, but the vocabulary can be conceptualized like this:

  • DAG: a description of the order in which work should take place
  • Operator: a class that acts as a template for carrying out some work
  • Task: a parameterized instance of an operator
  • Task Instance: a task that 1) has been assigned to a DAG and 2) has a state associated with a specific run of the DAG

By combining DAGs and Operators to create TaskInstances, you can build complex workflows.

Additional Functionality

In addition to the core Airflow objects, there are a number of more complex features that enable behaviors like limiting simultaneous access to resources, cross-communication, conditional execution, and more.

Hooks

Hooks are interfaces to external platforms and databases like Hive, S3, MySQL, Postgres, HDFS, and Pig. Hooks implement a common interface when possible, and act as a building block for operators. They also use the airflow.models.Connection model to retrieve hostnames and authentication information. Hooks keep authentication code and information out of pipelines, centralized in the metadata database.

Hooks are also very useful on their own to use in Python scripts, Airflow airflow.operators.PythonOperator, and in interactive environments like iPython or Jupyter Notebook.

Pools

Some systems can get overwhelmed when too many processes hit them at the same time. Airflow pools can be used to limit the execution parallelism on arbitrary sets of tasks. The list of pools is managed in the UI (Menu -> Admin -> Pools) by giving the pools a name and assigning it a number of worker slots. Tasks can then be associated with one of the existing pools by using the pool parameter when creating tasks (i.e., instantiating operators).

aggregate_db_message_job = BashOperator(
    task_id='aggregate_db_message_job',
    execution_timeout=timedelta(hours=3),
    pool='ep_data_pipeline_db_msg_agg',
    bash_command=aggregate_db_message_job_cmd,
    dag=dag)
aggregate_db_message_job.set_upstream(wait_for_empty_queue)

The pool parameter can be used in conjunction with priority_weight to define priorities in the queue, and which tasks get executed first as slots open up in the pool. The default priority_weight is 1, and can be bumped to any number. When sorting the queue to evaluate which task should be executed next, we use the priority_weight, summed up with all of the priority_weight values from tasks downstream from this task. You can use this to bump a specific important task and the whole path to that task gets prioritized accordingly.

Tasks will be scheduled as usual while the slots fill up. Once capacity is reached, runnable tasks get queued and their state will show as such in the UI. As slots free up, queued tasks start running based on the priority_weight (of the task and its descendants).

Note that by default tasks aren’t assigned to any pool and their execution parallelism is only limited to the executor’s setting.

Connections

The connection information to external systems is stored in the Airflow metadata database and managed in the UI (Menu -> Admin -> Connections) A conn_id is defined there and hostname / login / password / schema information attached to it. Airflow pipelines can simply refer to the centrally managed conn_id without having to hard code any of this information anywhere.

Many connections with the same conn_id can be defined and when that is the case, and when the hooks uses the get_connection method from BaseHook, Airflow will choose one connection randomly, allowing for some basic load balancing and fault tolerance when used in conjunction with retries.

Airflow also has the ability to reference connections via environment variables from the operating system. The environment variable needs to be prefixed with AIRFLOW_CONN_ to be considered a connection. When referencing the connection in the Airflow pipeline, the conn_id should be the name of the variable without the prefix. For example, if the conn_id is named postgres_master the environment variable should be named AIRFLOW_CONN_POSTGRES_MASTER (note that the environment variable must be all uppercase). Airflow assumes the value returned from the environment variable to be in a URI format (e.g. postgres://user:password@localhost:5432/master or s3://accesskey:secretkey@S3).

Queues

When using the CeleryExecutor, the celery queues that tasks are sent to can be specified. queue is an attribute of BaseOperator, so any task can be assigned to any queue. The default queue for the environment is defined in the airflow.cfg’s celery -> default_queue. This defines the queue that tasks get assigned to when not specified, as well as which queue Airflow workers listen to when started.

Workers can listen to one or multiple queues of tasks. When a worker is started (using the command airflow worker), a set of comma delimited queue names can be specified (e.g. airflow worker -q spark). This worker will then only pick up tasks wired to the specified queue(s).

This can be useful if you need specialized workers, either from a resource perspective (for say very lightweight tasks where one worker could take thousands of tasks without a problem), or from an environment perspective (you want a worker running from within the Spark cluster itself because it needs a very specific environment and security rights).

XComs

XComs let tasks exchange messages, allowing more nuanced forms of control and shared state. The name is an abbreviation of “cross-communication”. XComs are principally defined by a key, value, and timestamp, but also track attributes like the task/DAG that created the XCom and when it should become visible. Any object that can be pickled can be used as an XCom value, so users should make sure to use objects of appropriate size.

XComs can be “pushed” (sent) or “pulled” (received). When a task pushes an XCom, it makes it generally available to other tasks. Tasks can push XComs at any time by calling the xcom_push() method. In addition, if a task returns a value (either from its Operator’s execute() method, or from a PythonOperator’s python_callable function), then an XCom containing that value is automatically pushed.

Tasks call xcom_pull() to retrieve XComs, optionally applying filters based on criteria like key, source task_ids, and source dag_id. By default, xcom_pull() filters for the keys that are automatically given to XComs when they are pushed by being returned from execute functions (as opposed to XComs that are pushed manually).

If xcom_pull is passed a single string for task_ids, then the most recent XCom value from that task is returned; if a list of task_ids is passed, then a correpsonding list of XCom values is returned.

# inside a PythonOperator called 'pushing_task'
def push_function():
    return value

# inside another PythonOperator where provide_context=True
def pull_function(**context):
    value = context['task_instance'].xcom_pull(task_ids='pushing_task')

It is also possible to pull XCom directly in a template, here’s an example of what this may look like:

SELECT * FROM {{ task_instance.xcom_pull(task_ids='foo', key='table_name') }}

Note that XComs are similar to Variables, but are specifically designed for inter-task communication rather than global settings.

Variables

Variables are a generic way to store and retrieve arbitrary content or settings as a simple key value store within Airflow. Variables can be listed, created, updated and deleted from the UI (Admin -> Variables), code or CLI. While your pipeline code definition and most of your constants and variables should be defined in code and stored in source control, it can be useful to have some variables or configuration items accessible and modifiable through the UI.

from airflow.models import Variable
foo = Variable.get("foo")
bar = Variable.get("bar", deserialize_json=True)

The second call assumes json content and will be deserialized into bar. Note that Variable is a sqlalchemy model and can be used as such.

Branching

Sometimes you need a workflow to branch, or only go down a certain path based on an arbitrary condition which is typically related to something that happened in an upstream task. One way to do this is by using the BranchPythonOperator.

The BranchPythonOperator is much like the PythonOperator except that it expects a python_callable that returns a task_id. The task_id returned is followed, and all of the other paths are skipped. The task_id returned by the Python function has to be referencing a task directly downstream from the BranchPythonOperator task.

Note that using tasks with depends_on_past=True downstream from BranchPythonOperator is logically unsound as skipped status will invariably lead to block tasks that depend on their past successes. skipped states propagates where all directly upstream tasks are skipped.

If you want to skip some tasks, keep in mind that you can’t have an empty path, if so make a dummy task.

like this, the dummy task “branch_false” is skipped

_images/branch_good.png

Not like this, where the join task is skipped

_images/branch_bad.png

SubDAGs

SubDAGs are perfect for repeating patterns. Defining a function that returns a DAG object is a nice design pattern when using Airflow.

Airbnb uses the stage-check-exchange pattern when loading data. Data is staged in a temporary table, after which data quality checks are performed against that table. Once the checks all pass the partition is moved into the production table.

As another example, consider the following DAG:

_images/subdag_before.png

We can combine all of the parallel task-* operators into a single SubDAG, so that the resulting DAG resembles the following:

_images/subdag_after.png

Note that SubDAG operators should contain a factory method that returns a DAG object. This will prevent the SubDAG from being treated like a separate DAG in the main UI. For example:

#dags/subdag.py
from airflow.models import DAG
from airflow.operators.dummy_operator import DummyOperator


# Dag is returned by a factory method
def sub_dag(parent_dag_name, child_dag_name, start_date, schedule_interval):
  dag = DAG(
    '%s.%s' % (parent_dag_name, child_dag_name),
    schedule_interval=schedule_interval,
    start_date=start_date,
  )

  dummy_operator = DummyOperator(
    task_id='dummy_task',
    dag=dag,
  )

  return dag

This SubDAG can then be referenced in your main DAG file:

# main_dag.py
from datetime import datetime, timedelta
from airflow.models import DAG
from airflow.operators.subdag_operator import SubDagOperator
from dags.subdag import sub_dag


PARENT_DAG_NAME = 'parent_dag'
CHILD_DAG_NAME = 'child_dag'

main_dag = DAG(
  dag_id=PARENT_DAG_NAME,
  schedule_interval=timedelta(hours=1),
  start_date=datetime(2016, 1, 1)
)

sub_dag = SubDagOperator(
  subdag=sub_dag(PARENT_DAG_NAME, CHILD_DAG_NAME, main_dag.start_date,
                 main_dag.schedule_interval),
  task_id=CHILD_DAG_NAME,
  dag=main_dag,
)

You can zoom into a SubDagOperator from the graph view of the main DAG to show the tasks contained within the SubDAG:

_images/subdag_zoom.png

Some other tips when using SubDAGs:

  • by convention, a SubDAG’s dag_id should be prefixed by its parent and a dot. As in parent.child
  • share arguments between the main DAG and the SubDAG by passing arguments to the SubDAG operator (as demonstrated above)
  • SubDAGs must have a schedule and be enabled. If the SubDAG’s schedule is set to None or @once, the SubDAG will succeed without having done anything
  • clearing a SubDagOperator also clears the state of the tasks within
  • marking success on a SubDagOperator does not affect the state of the tasks within
  • refrain from using depends_on_past=True in tasks within the SubDAG as this can be confusing
  • it is possible to specify an executor for the SubDAG. It is common to use the SequentialExecutor if you want to run the SubDAG in-process and effectively limit its parallelism to one. Using LocalExecutor can be problematic as it may over-subscribe your worker, running multiple tasks in a single slot

See airflow/example_dags for a demonstration.

SLAs

Service Level Agreements, or time by which a task or DAG should have succeeded, can be set at a task level as a timedelta. If one or many instances have not succeeded by that time, an alert email is sent detailing the list of tasks that missed their SLA. The event is also recorded in the database and made available in the web UI under Browse->Missed SLAs where events can be analyzed and documented.

Trigger Rules

Though the normal workflow behavior is to trigger tasks when all their directly upstream tasks have succeeded, Airflow allows for more complex dependency settings.

All operators have a trigger_rule argument which defines the rule by which the generated task get triggered. The default value for trigger_rule is all_success and can be defined as “trigger this task when all directly upstream tasks have succeeded”. All other rules described here are based on direct parent tasks and are values that can be passed to any operator while creating tasks:

  • all_success: (default) all parents have succeeded
  • all_failed: all parents are in a failed or upstream_failed state
  • all_done: all parents are done with their execution
  • one_failed: fires as soon as at least one parent has failed, it does not wait for all parents to be done
  • one_success: fires as soon as at least one parent succeeds, it does not wait for all parents to be done
  • dummy: dependencies are just for show, trigger at will

Note that these can be used in conjunction with depends_on_past (boolean) that, when set to True, keeps a task from getting triggered if the previous schedule for the task hasn’t succeeded.

Latest Run Only

Standard workflow behavior involves running a series of tasks for a particular date/time range. Some workflows, however, perform tasks that are independent of run time but need to be run on a schedule, much like a standard cron job. In these cases, backfills or running jobs missed during a pause just wastes CPU cycles.

For situations like this, you can use the LatestOnlyOperator to skip tasks that are not being run during the most recent scheduled run for a DAG. The LatestOnlyOperator skips all immediate downstream tasks, and itself, if the time right now is not between its execution_time and the next scheduled execution_time.

One must be aware of the interaction between skipped tasks and trigger rules. Skipped tasks will cascade through trigger rules all_success and all_failed but not all_done, one_failed, one_success, and dummy. If you would like to use the LatestOnlyOperator with trigger rules that do not cascade skips, you will need to ensure that the LatestOnlyOperator is directly upstream of the task you would like to skip.

It is possible, through use of trigger rules to mix tasks that should run in the typical date/time dependent mode and those using the LatestOnlyOperator.

For example, consider the following dag:

#dags/latest_only_with_trigger.py
import datetime as dt

from airflow.models import DAG
from airflow.operators.dummy_operator import DummyOperator
from airflow.operators.latest_only_operator import LatestOnlyOperator
from airflow.utils.trigger_rule import TriggerRule


dag = DAG(
    dag_id='latest_only_with_trigger',
    schedule_interval=dt.timedelta(hours=4),
    start_date=dt.datetime(2016, 9, 20),
)

latest_only = LatestOnlyOperator(task_id='latest_only', dag=dag)

task1 = DummyOperator(task_id='task1', dag=dag)
task1.set_upstream(latest_only)

task2 = DummyOperator(task_id='task2', dag=dag)

task3 = DummyOperator(task_id='task3', dag=dag)
task3.set_upstream([task1, task2])

task4 = DummyOperator(task_id='task4', dag=dag,
                      trigger_rule=TriggerRule.ALL_DONE)
task4.set_upstream([task1, task2])

In the case of this dag, the latest_only task will show up as skipped for all runs except the latest run. task1 is directly downstream of latest_only and will also skip for all runs except the latest. task2 is entirely independent of latest_only and will run in all scheduled periods. task3 is downstream of task1 and task2 and because of the default trigger_rule being all_success will receive a cascaded skip from task1. task4 is downstream of task1 and task2 but since its trigger_rule is set to all_done it will trigger as soon as task1 has been skipped (a valid completion state) and task2 has succeeded.

_images/latest_only_with_trigger.png

Zombies & Undeads

Task instances die all the time, usually as part of their normal life cycle, but sometimes unexpectedly.

Zombie tasks are characterized by the absence of an heartbeat (emitted by the job periodically) and a running status in the database. They can occur when a worker node can’t reach the database, when Airflow processes are killed externally, or when a node gets rebooted for instance. Zombie killing is performed periodically by the scheduler’s process.

Undead processes are characterized by the existence of a process and a matching heartbeat, but Airflow isn’t aware of this task as running in the database. This mismatch typically occurs as the state of the database is altered, most likely by deleting rows in the “Task Instances” view in the UI. Tasks are instructed to verify their state as part of the heartbeat routine, and terminate themselves upon figuring out that they are in this “undead” state.

Cluster Policy

Your local airflow settings file can define a policy function that has the ability to mutate task attributes based on other task or DAG attributes. It receives a single argument as a reference to task objects, and is expected to alter its attributes.

For example, this function could apply a specific queue property when using a specific operator, or enforce a task timeout policy, making sure that no tasks run for more than 48 hours. Here’s an example of what this may look like inside your airflow_settings.py:

def policy(task):
    if task.__class__.__name__ == 'HivePartitionSensor':
        task.queue = "sensor_queue"
    if task.timeout > timedelta(hours=48):
        task.timeout = timedelta(hours=48)

Documentation & Notes

It’s possible to add documentation or notes to your dags & task objects that become visible in the web interface (“Graph View” for dags, “Task Details” for tasks). There are a set of special task attributes that get rendered as rich content if defined:

attribute rendered to
doc monospace
doc_json json
doc_yaml yaml
doc_md markdown
doc_rst reStructuredText

Please note that for dags, dag_md is the only attribute interpreted.

This is especially useful if your tasks are built dynamically from configuration files, it allows you to expose the configuration that led to the related tasks in Airflow.

"""
### My great DAG
"""

dag = DAG('my_dag', default_args=default_args)
dag.doc_md = __doc__

t = BashOperator("foo", dag=dag)
t.doc_md = """\
#Title"
Here's a [url](www.airbnb.com)
"""

This content will get rendered as markdown respectively in the “Graph View” and “Task Details” pages.

Jinja Templating

Airflow leverages the power of Jinja Templating and this can be a powerful tool to use in combination with macros (see the Macros section).

For example, say you want to pass the execution date as an environment variable to a Bash script using the BashOperator.

# The execution date as YYYY-MM-DD
date = "{{ ds }}"
t = BashOperator(
    task_id='test_env',
    bash_command='/tmp/test.sh ',
    dag=dag,
    env={'EXECUTION_DATE': date})

Here, {{ ds }} is a macro, and because the env parameter of the BashOperator is templated with Jinja, the execution date will be available as an environment variable named EXECUTION_DATE in your Bash script.

You can use Jinja templating with every parameter that is marked as “templated” in the documentation.

Packaged dags

While often you will specify dags in a single .py file it might sometimes be required to combine dag and its dependencies. For example, you might want to combine several dags together to version them together or you might want to manage them together or you might need an extra module that is not available by default on the system you are running airflow on. To allow this you can create a zip file that contains the dag(s) in the root of the zip file and have the extra modules unpacked in directories.

For instance you can create a zip file that looks like this:

my_dag1.py
my_dag2.py
package1/__init__.py
package1/functions.py

Airflow will scan the zip file and try to load my_dag1.py and my_dag2.py. It will not go into subdirectories as these are considered to be potential packages.

In case you would like to add module dependencies to your DAG you basically would do the same, but then it is more to use a virtualenv and pip.

virtualenv zip_dag
source zip_dag/bin/activate

mkdir zip_dag_contents
cd zip_dag_contents

pip install --install-option="--install-lib=$PWD" my_useful_package
cp ~/my_dag.py .

zip -r zip_dag.zip *

Note

the zip file will be inserted at the beginning of module search list (sys.path) and as such it will be available to any other code that resides within the same interpreter.

Note

packaged dags cannot be used with pickling turned on.

Note

packaged dags cannot contain dynamic libraries (eg. libz.so) these need to be available on the system if a module needs those. In other words only pure python modules can be packaged.