Skip to content
Switch branches/tags
Go to file
115 contributors

Users who have contributed to this file

@kaxil @mik-laj @ashb @michalslowikowski00 @turbaszek @potiuk @bolkedebruin @Fokko @feng-tao @zhongjiajie @feluelle @BasPH
3031 lines (2143 sloc) 170 KB

Updating Airflow

This file documents any backwards-incompatible changes in Airflow and assists users migrating to a new version.

Table of contents


Airflow 2.0.1

Permission to view Airflow Configurations has been removed from User and Viewer role

Previously, Users with User or Viewer role were able to get/view configurations using the REST API or in the Webserver. From Airflow 2.0.1, only users with Admin or Op role would be able to get/view Configurations.

To allow users with other roles to view configuration, add can read on Configurations permissions to that role.

Note that if [webserver] expose_config is set to False, the API will throw a 403 response even if the user has role with can read on Configurations permission.

Default [celery] worker_concurrency is changed to 16

The default value for [celery] worker_concurrency was 16 for Airflow <2.0.0. However, it was unintentionally changed to 8 in 2.0.0.

From Airflow 2.0.1, we revert to the old default of 16.

Default [scheduler] min_file_process_interval is changed to 30

The default value for [scheduler] min_file_process_interval was 0, due to which the CPU Usage mostly stayed around 100% as the DAG files are parsed constantly.

From Airflow 2.0.0, the scheduling decisions have been moved from DagFileProcessor to Scheduler, so we can keep the default a bit higher: 30.

Airflow 2.0.0

The 2.0 release of the Airflow is a significant upgrade, and includes substantial major changes, and some of them may be breaking. Existing code written for earlier versions of this project will may require updates to use this version. Sometimes necessary configuration changes are also required. This document describes the changes that have been made, and what you need to do to update your usage.

If you experience issues or have questions, please file an issue.

Major changes

This section describes the major changes that have been made in this release.

The experimental REST API is disabled by default

The experimental REST API is disabled by default. To restore these APIs while migrating to the stable REST API, set enable_experimental_api option in [api] section to True.

Please note that the experimental REST API do not have access control. The authenticated user has full access.

SparkJDBCHook default connection

For SparkJDBCHook default connection was spark-default, and for SparkSubmitHook it was spark_default. Both hooks now use the spark_default which is a common pattern for the connection names used across all providers.

Changes to output argument in commands

From Airflow 2.0, We are replacing tabulate with rich to render commands output. Due to this change, the --output argument will no longer accept formats of tabulate tables. Instead, it now accepts:

  • table - will render the output in predefined table
  • json - will render the output as a json
  • yaml - will render the output as yaml

By doing this we increased consistency and gave users possibility to manipulate the output programmatically (when using json or yaml).

Affected commands:

  • airflow dags list
  • airflow dags report
  • airflow dags list-runs
  • airflow dags list-jobs
  • airflow connections list
  • airflow connections get
  • airflow pools list
  • airflow pools get
  • airflow pools set
  • airflow pools delete
  • airflow pools import
  • airflow pools export
  • airflow role list
  • airflow providers list
  • airflow providers get
  • airflow providers hooks
  • airflow tasks states-for-dag-run
  • airflow users list
  • airflow variables list

Azure Wasb Hook does not work together with Snowflake hook

The WasbHook in Apache Airflow use a legacy version of Azure library. While the conflict is not significant for most of the Azure hooks, it is a problem for Wasb Hook because the blob folders for both libraries overlap. Installing both Snowflake and Azure extra will result in non-importable WasbHook.

Rename all to devel_all extra

The all extras were reduced to include only user-facing dependencies. This means that this extra does not contain development dependencies. If you were relying on all extra then you should use now devel_all or figure out if you need development extras at all.

Context variables prev_execution_date_success and prev_execution_date_success are now pendulum.DateTime

Rename policy to task_policy

Because Airflow introduced DAG level policy (dag_policy) we decided to rename existing policy function to task_policy to make the distinction more profound and avoid any confusion.

Users using cluster policy need to rename their policy functions in to task_policy.

Default value for [celery] operation_timeout has changed to 1.0

From Airflow 2, by default Airflow will retry 3 times to publish task to Celery broker. This is controlled by [celery] task_publish_max_retries. Because of this we can now have a lower Operation timeout that raises AirflowTaskTimeout. This generally occurs during network blips or intermittent DNS issues.

Adding Operators and Sensors via plugins is no longer supported

Operators and Sensors should no longer be registered or imported via Airflow's plugin mechanism -- these types of classes are just treated as plain python classes by Airflow, so there is no need to register them with Airflow.

If you previously had a plugins/ and you used it like this in a DAG:

from airflow.operators.my_plugin import MyOperator

You should instead import it as:

from my_plugin import MyOperator

The name under airflow.operators. was the plugin name, where as in the second example it is the python module name where the operator is defined.

See for more info.

Importing Hooks via plugins is no longer supported

Importing hooks added in plugins via airflow.hooks.<plugin_name> is no longer supported, and hooks should just be imported as regular python modules.

from airflow.hooks.my_plugin import MyHook

You should instead import it as:

from my_plugin import MyHook

It is still possible (but not required) to "register" hooks in plugins. This is to allow future support for dynamically populating the Connections form in the UI.

See for more info.

Adding Operators and Sensors via plugins is no longer supported

The default value for [core] enable_xcom_pickling has been changed to False

The pickle type for XCom messages has been replaced to JSON by default to prevent RCE attacks. Note that JSON serialization is stricter than pickling, so for example if you want to pass raw bytes through XCom you must encode them using an encoding like base64. If you understand the risk and still want to use pickling, set enable_xcom_pickling = True in your Airflow config's core section.

Airflowignore of base path

There was a bug fixed in that the "airflowignore" checked the base path of the dag folder for forbidden dags, not only the relative part. This had the effect that if the base path contained the excluded word the whole dag folder could have been excluded. For example if the airflowignore file contained x, and the dags folder was '/var/x/dags', then all dags in the folder would be excluded. The fix only matches the relative path only now which means that if you previously used full path as ignored, you should change it to relative one. For example if your dag folder was '/var/dags/' and your airflowignore contained '/var/dag/excluded/', you should change it to 'excluded/'.

ExternalTaskSensor provides all task context variables to execution_date_fn as keyword arguments

The old syntax of passing context as a dictionary will continue to work with the caveat that the argument must be named context. The following will break. To fix it, change ctx to context.

def execution_date_fn(execution_date, ctx):

execution_date_fn can take in any number of keyword arguments available in the task context dictionary. The following forms of execution_date_fn are all supported:

def execution_date_fn(dt):

def execution_date_fn(execution_date):

def execution_date_fn(execution_date, ds_nodash):

def execution_date_fn(execution_date, ds_nodash, dag):

The default value for [webserver] cookie_samesite has been changed to Lax

As recommended by Flask, the [webserver] cookie_samesite has been changed to Lax from '' (empty string) .

Changes to import paths

Formerly the core code was maintained by the original creators - Airbnb. The code that was in the contrib package was supported by the community. The project was passed to the Apache community and currently the entire code is maintained by the community, so now the division has no justification, and it is only due to historical reasons. In Airflow 2.0, we want to organize packages and move integrations with third party services to the airflow.providers package.

All changes made are backward compatible, but if you use the old import paths you will see a deprecation warning. The old import paths can be abandoned in the future.

According to AIP-21 _operator suffix has been removed from operators. A deprecation warning has also been raised for paths importing with the suffix.

The following table shows changes in import paths.

Old path New path
airflow.hooks.base_hook.BaseHook airflow.hooks.base.BaseHook
airflow.hooks.dbapi_hook.DbApiHook airflow.hooks.dbapi.DbApiHook
airflow.operators.dummy_operator.DummyOperator airflow.operators.dummy.DummyOperator
airflow.operators.dagrun_operator.TriggerDagRunOperator airflow.operators.trigger_dagrun.TriggerDagRunOperator
airflow.operators.branch_operator.BaseBranchOperator airflow.operators.branch.BaseBranchOperator
airflow.operators.subdag_operator.SubDagOperator airflow.operators.subdag.SubDagOperator
airflow.sensors.base_sensor_operator.BaseSensorOperator airflow.sensors.base.BaseSensorOperator
airflow.sensors.date_time_sensor.DateTimeSensor airflow.sensors.date_time.DateTimeSensor
airflow.sensors.external_task_sensor.ExternalTaskMarker airflow.sensors.external_task.ExternalTaskMarker
airflow.sensors.external_task_sensor.ExternalTaskSensor airflow.sensors.external_task.ExternalTaskSensor
airflow.sensors.sql_sensor.SqlSensor airflow.sensors.sql.SqlSensor
airflow.sensors.time_delta_sensor.TimeDeltaSensor airflow.sensors.time_delta.TimeDeltaSensor
airflow.contrib.sensors.weekday_sensor.DayOfWeekSensor airflow.sensors.weekday.DayOfWeekSensor

Database schema changes

In order to migrate the database, you should use the command airflow db upgrade, but in some cases manual steps are required.

Unique conn_id in connection table

Previously, Airflow allowed users to add more than one connection with the same conn_id and on access it would choose one connection randomly. This acted as a basic load balancing and fault tolerance technique, when used in conjunction with retries.

This behavior caused some confusion for users, and there was no clear evidence if it actually worked well or not.

Now the conn_id will be unique. If you already have duplicates in your metadata database, you will have to manage those duplicate connections before upgrading the database.

Not-nullable conn_type column in connection table

The conn_type column in the connection table must contain content. Previously, this rule was enforced by application logic, but was not enforced by the database schema.

If you made any modifications to the table directly, make sure you don't have null in the conn_type column.

Configuration changes

This release contains many changes that require a change in the configuration of this application or other application that integrate with it.

This section describes the changes that have been made, and what you need to do to.

airflow.contrib.utils.log has been moved

Formerly the core code was maintained by the original creators - Airbnb. The code that was in the contrib package was supported by the community. The project was passed to the Apache community and currently the entire code is maintained by the community, so now the division has no justification, and it is only due to historical reasons. In Airflow 2.0, we want to organize packages and move integrations with third party services to the airflow.providers package.

To clean up, the following packages were moved:

Old package New package
airflow.contrib.utils.log airflow.utils.log
airflow.utils.log.es_task_handler airflow.providers.elasticsearch.log.es_task_handler

You should update the import paths if you are setting log configurations with the logging_config_class option. The old import paths still works but can be abandoned.

SendGrid emailer has been moved

Formerly the core code was maintained by the original creators - Airbnb. The code that was in the contrib package was supported by the community. The project was passed to the Apache community and currently the entire code is maintained by the community, so now the division has no justification, and it is only due to historical reasons.

To clean up, the send_mail function from the airflow.contrib.utils.sendgrid module has been moved.

If your configuration file looks like this:

email_backend = airflow.contrib.utils.sendgrid.send_email

It should look like this now:

email_backend = airflow.providers.sendgrid.utils.emailer.send_email

The old configuration still works but can be abandoned.

Unify hostname_callable option in core section

The previous option used a colon(:) to split the module from function. Now the dot(.) is used.

The change aims to unify the format of all options that refer to objects in the airflow.cfg file.

Custom executors is loaded using full import path

In previous versions of Airflow it was possible to use plugins to load custom executors. It is still possible, but the configuration has changed. Now you don't have to create a plugin to configure a custom executor, but you need to provide the full path to the module in the executor option in the core section. The purpose of this change is to simplify the plugin mechanism and make it easier to configure executor.

If your module was in the path my_acme_company.executors.MyCustomExecutor and the plugin was called my_plugin then your configuration looks like this

executor = my_plugin.MyCustomExecutor

And now it should look like this:

executor = my_acme_company.executors.MyCustomExecutor

The old configuration is still works but can be abandoned at any time.

Drop plugin support for stat_name_handler

In previous version, you could use plugins mechanism to configure stat_name_handler. You should now use the stat_name_handler option in [scheduler] section to achieve the same effect.

If your plugin looked like this and was available through the test_plugin path:

def my_stat_name_handler(stat):
    return stat

class AirflowTestPlugin(AirflowPlugin):
    name = "test_plugin"
    stat_name_handler = my_stat_name_handler

then your airflow.cfg file should look like this:


This change is intended to simplify the statsd configuration.

Logging configuration has been moved to new section

The following configurations have been moved from [core] to the new [logging] section.

  • base_log_folder
  • remote_logging
  • remote_log_conn_id
  • remote_base_log_folder
  • encrypt_s3_logs
  • logging_level
  • fab_logging_level
  • logging_config_class
  • colored_console_log
  • colored_log_format
  • colored_formatter_class
  • log_format
  • simple_log_format
  • task_log_prefix_template
  • log_filename_template
  • log_processor_filename_template
  • dag_processor_manager_log_location
  • task_log_reader

Metrics configuration has been moved to new section

The following configurations have been moved from [scheduler] to the new [metrics] section.

  • statsd_on
  • statsd_host
  • statsd_port
  • statsd_prefix
  • statsd_allow_list
  • stat_name_handler
  • statsd_datadog_enabled
  • statsd_datadog_tags
  • statsd_custom_client_path

Changes to Elasticsearch logging provider

When JSON output to stdout is enabled, log lines will now contain the log_id & offset fields, this should make reading task logs from elasticsearch on the webserver work out of the box. Example configuration:

remote_logging = True
host = http://es-host:9200
write_stdout = True
json_format = True

Note that the webserver expects the log line data itself to be present in the message field of the document.

Remove gcp_service_account_keys option in airflow.cfg file

This option has been removed because it is no longer supported by the Google Kubernetes Engine. The new recommended service account keys for the Google Cloud management method is Workload Identity.

Fernet is enabled by default

The fernet mechanism is enabled by default to increase the security of the default installation. In order to restore the previous behavior, the user must consciously set an empty key in the fernet_key option of section [core] in the airflow.cfg file.

At the same time, this means that the apache-airflow[crypto] extra-packages are always installed. However, this requires that your operating system has libffi-dev installed.

Changes to propagating Kubernetes worker annotations

kubernetes_annotations configuration section has been removed. A new key worker_annotations has been added to existing kubernetes section instead. That is to remove restriction on the character set for k8s annotation keys. All key/value pairs from kubernetes_annotations should now go to worker_annotations as a json. I.e. instead of e.g.

annotation_key = annotation_value
annotation_key2 = annotation_value2

it should be rewritten to

worker_annotations = { "annotation_key" : "annotation_value", "annotation_key2" : "annotation_value2" }

Remove run_duration

We should not use the run_duration option anymore. This used to be for restarting the scheduler from time to time, but right now the scheduler is getting more stable and therefore using this setting is considered bad and might cause an inconsistent state.

Rename pool statsd metrics

Used slot has been renamed to running slot to make the name self-explanatory and the code more maintainable.

This means pool.used_slots.<pool_name> metric has been renamed to pool.running_slots.<pool_name>. The Used Slots column in Pools Web UI view has also been changed to Running Slots.

Removal of Mesos Executor

The Mesos Executor is removed from the code base as it was not widely used and not maintained. Mailing List Discussion on deleting it.

Change dag loading duration metric name

Change DAG file loading duration metric from dag.loading-duration.<dag_id> to dag.loading-duration.<dag_file>. This is to better handle the case when a DAG file has multiple DAGs.

Sentry is disabled by default

Sentry is disabled by default. To enable these integrations, you need set sentry_on option in [sentry] section to "True".

Simplified GCSTaskHandler configuration

In previous versions, in order to configure the service account key file, you had to create a connection entry. In the current version, you can configure google_key_path option in [logging] section to set the key file path.

Users using Application Default Credentials (ADC) need not take any action.

The change aims to simplify the configuration of logging, to prevent corruption of the instance configuration by changing the value controlled by the user - connection entry. If you configure a backend secret, it also means the webserver doesn't need to connect to it. This simplifies setups with multiple GCP projects, because only one project will require the Secret Manager API to be enabled.

Changes to the core operators/hooks

We strive to ensure that there are no changes that may affect the end user and your files, but this release may contain changes that will require changes to your DAG files.

This section describes the changes that have been made, and what you need to do to update your DAG File, if you use core operators or any other.

BaseSensorOperator now respects the trigger_rule of downstream tasks

Previously, BaseSensorOperator with setting soft_fail=True skips itself and skips all its downstream tasks unconditionally, when it fails i.e the trigger_rule of downstream tasks is not respected.

In the new behavior, the trigger_rule of downstream tasks is respected. User can preserve/achieve the original behaviour by setting the trigger_rule of each downstream task to all_success.

BaseOperator uses metaclass

BaseOperator class uses a BaseOperatorMeta as a metaclass. This meta class is based on abc.ABCMeta. If your custom operator uses different metaclass then you will have to adjust it.

Remove SQL support in BaseHook

Remove get_records and get_pandas_df and run from BaseHook, which only apply for sql like hook, If want to use them, or your custom hook inherit them, please use airflow.hooks.dbapi.DbApiHook

Assigning task to a DAG using bitwise shift (bit-shift) operators are no longer supported

Previously, you could assign a task to a DAG as follows:

dag = DAG('my_dag')
dummy = DummyOperator(task_id='dummy')

dag >> dummy

This is no longer supported. Instead, we recommend using the DAG as context manager:

with DAG('my_dag') as dag:
    dummy = DummyOperator(task_id='dummy')

Removed deprecated import mechanism

The deprecated import mechanism has been removed so the import of modules becomes more consistent and explicit.

For example: from airflow.operators import BashOperator becomes from airflow.operators.bash_operator import BashOperator

Changes to sensor imports

Sensors are now accessible via airflow.sensors and no longer via airflow.operators.sensors.

For example: from airflow.operators.sensors import BaseSensorOperator becomes from airflow.sensors.base import BaseSensorOperator

Skipped tasks can satisfy wait_for_downstream

Previously, a task instance with wait_for_downstream=True will only run if the downstream task of the previous task instance is successful. Meanwhile, a task instance with depends_on_past=True will run if the previous task instance is either successful or skipped. These two flags are close siblings yet they have different behavior. This inconsistency in behavior made the API less intuitive to users. To maintain consistent behavior, both successful or skipped downstream task can now satisfy the wait_for_downstream=True flag.



The chain and cross_downstream methods are now moved to airflow.models.baseoperator module from airflow.utils.helpers module.

The baseoperator module seems to be a better choice to keep closely coupled methods together. Helpers module is supposed to contain standalone helper methods that can be imported by all classes.

The chain method and cross_downstream method both use BaseOperator. If any other package imports any classes or functions from helpers module, then it automatically has an implicit dependency to BaseOperator. That can often lead to cyclic dependencies.

More information in AIRFLOW-6392

In Airflow <2.0 you imported those two methods like this:

from airflow.utils.helpers import chain
from airflow.utils.helpers import cross_downstream

In Airflow 2.0 it should be changed to:

from airflow.models.baseoperator import chain
from airflow.models.baseoperator import cross_downstream


BranchPythonOperator will now return a value equal to the task_id of the chosen branch, where previously it returned None. Since it inherits from BaseOperator it will do an xcom_push of this value if do_xcom_push=True. This is useful for downstream decision-making.


SQLSensor now consistent with python bool() function and the allow_null parameter has been removed.

It will resolve after receiving any value that is casted to True with python bool(value). That changes the previous response receiving NULL or '0'. Earlier '0' has been treated as success criteria. NULL has been treated depending on value of allow_nullparameter. But all the previous behaviour is still achievable setting param success to lambda x: x is None or str(x) not in ('0', '').


The TriggerDagRunOperator now takes a conf argument to which a dict can be provided as conf for the DagRun. As a result, the python_callable argument was removed. PR:


provide_context argument on the PythonOperator was removed. The signature of the callable passed to the PythonOperator is now inferred and argument values are always automatically provided. There is no need to explicitly provide or not provide the context anymore. For example:

def myfunc(execution_date):

python_operator = PythonOperator(task_id='mytask', python_callable=myfunc, dag=dag)

Notice you don't have to set provide_context=True, variables from the task context are now automatically detected and provided.

All context variables can still be provided with a double-asterisk argument:

def myfunc(**context):
    print(context)  # all variables will be provided to context

python_operator = PythonOperator(task_id='mytask', python_callable=myfunc)

The task context variable names are reserved names in the callable function, hence a clash with op_args and op_kwargs results in an exception:

def myfunc(dag):
    # raises a ValueError because "dag" is a reserved name
    # valid signature example: myfunc(mydag)

python_operator = PythonOperator(

The change is backwards compatible, setting provide_context will add the provide_context variable to the kwargs (but won't do anything).

PR: #5990


FileSensor is now takes a glob pattern, not just a filename. If the filename you are looking for has *, ?, or [ in it then you should replace these with [*], [?], and [[].


SubDagOperator is changed to use Airflow scheduler instead of backfill to schedule tasks in the subdag. User no longer need to specify the executor in SubDagOperator.







The do_xcom_push flag (a switch to push the result of an operator to xcom or not) was appearing in different incarnations in different operators. It's function has been unified under a common name (do_xcom_push) on BaseOperator. This way it is also easy to globally disable pushing results to xcom.

The following operators were affected:

  • DatastoreExportOperator (Backwards compatible)
  • DatastoreImportOperator (Backwards compatible)
  • KubernetesPodOperator (Not backwards compatible)
  • SSHOperator (Not backwards compatible)
  • WinRMOperator (Not backwards compatible)
  • BashOperator (Not backwards compatible)
  • DockerOperator (Not backwards compatible)
  • SimpleHttpOperator (Not backwards compatible)

See AIRFLOW-3249 for details


In previous versions, the LatestOnlyOperator forcefully skipped all (direct and undirect) downstream tasks on its own. From this version on the operator will only skip direct downstream tasks and the scheduler will handle skipping any further downstream dependencies.

No change is needed if only the default trigger rule all_success is being used.

If the DAG relies on tasks with other trigger rules (i.e. all_done) being skipped by the LatestOnlyOperator, adjustments to the DAG need to be made to accommodate the change in behaviour, i.e. with additional edges from the LatestOnlyOperator.

The goal of this change is to achieve a more consistent and configurale cascading behaviour based on the BaseBranchOperator (see AIRFLOW-2923 and AIRFLOW-1784).

Changes to the core Python API

We strive to ensure that there are no changes that may affect the end user, and your Python files, but this release may contain changes that will require changes to your plugins, DAG File or other integration.

Only changes unique to this provider are described here. You should still pay attention to the changes that have been made to the core (including core operators) as they can affect the integration behavior of this provider.

This section describes the changes that have been made, and what you need to do to update your Python files.

Removed sub-package imports from airflow/

The imports LoggingMixin, conf, and AirflowException have been removed from airflow/ All implicit references of these objects will no longer be valid. To migrate, all usages of each old path must be replaced with its corresponding new path.

Old Path (Implicit Import) New Path (Explicit Import)
airflow.LoggingMixin airflow.utils.log.logging_mixin.LoggingMixin
airflow.conf airflow.configuration.conf
airflow.AirflowException airflow.exceptions.AirflowException

Variables removed from the task instance context

The following variables were removed from the task instance context:

  • end_date
  • latest_date
  • tables


Formerly the core code was maintained by the original creators - Airbnb. The code that was in the contrib package was supported by the community. The project was passed to the Apache community and currently the entire code is maintained by the community, so now the division has no justification, and it is only due to historical reasons.

To clean up, Weekday enum has been moved from airflow.contrib.utils into airflow.utils module.


The connection module has new deprecated methods:

  • Connection.parse_from_uri
  • Connection.log_info
  • Connection.debug_info

and one deprecated function:

  • parse_netloc_to_hostname

Previously, users could create a connection object in two ways

conn_1 = Connection(conn_id="conn_a", uri="mysql://AAA/")
# or
conn_2 = Connection(conn_id="conn_a")

Now the second way is not supported.

Connection.log_info and Connection.debug_info method have been deprecated. Read each Connection field individually or use the default representation (__repr__).

The old method is still works but can be abandoned at any time. The changes are intended to delete method that are rarely used.


DAG.create_dagrun accepts run_type and does not require run_id This change is caused by adding run_type column to DagRun.

Previous signature:

def create_dagrun(self,


def create_dagrun(self,

If user provides run_id then the run_type will be derived from it by checking prefix, allowed types : manual, scheduled, backfill (defined by airflow.utils.types.DagRunType).

If user provides run_type and execution_date then run_id is constructed as {run_type}__{execution_data.isoformat()}.

Airflow should construct dagruns using run_type and execution_date, creation using run_id is preserved for user actions.


Use DagRunType.SCHEDULED.value instead of DagRun.ID_PREFIX

All the run_id prefixes for different kind of DagRuns have been grouped into a single enum in airflow.utils.types.DagRunType.

Previously, there were defined in various places, example as ID_PREFIX class variables for DagRun, BackfillJob and in _trigger_dag function.


>> from airflow.models.dagrun import DagRun

Replaced by:

>> from airflow.utils.types import DagRunType
>> DagRunType.SCHEDULED.value


We remove airflow.utils.file.TemporaryDirectory Since Airflow dropped support for Python < 3.5 there's no need to have this custom implementation of TemporaryDirectory because the same functionality is provided by tempfile.TemporaryDirectory.

Now users instead of import from airflow.utils.files import TemporaryDirectory should do from tempfile import TemporaryDirectory. Both context managers provide the same interface, thus no additional changes should be required.


We removed airflow.AirflowMacroPlugin class. The class was there in airflow package but it has not been used (apparently since 2015). It has been removed.


CONTEXT_MANAGER_DAG was removed from settings. It's role has been taken by DagContext in 'airflow.models.dag'. One of the reasons was that settings should be rather static than store dynamic context from the DAG, but the main one is that moving the context out of settings allowed to untangle cyclic imports between DAG, BaseOperator, SerializedDAG, SerializedBaseOperator which was part of AIRFLOW-6010.



Function redirect_stderr and redirect_stdout from airflow.utils.log.logging_mixin module has been deleted because it can be easily replaced by the standard library. The functions of the standard library are more flexible and can be used in larger cases.

The code below

import logging

from airflow.utils.log.logging_mixin import redirect_stderr, redirect_stdout

logger = logging.getLogger("custom-logger")
with redirect_stdout(logger, logging.INFO), redirect_stderr(logger, logging.WARN):
    print("I love Airflow")

can be replaced by the following code:

from contextlib import redirect_stdout, redirect_stderr
import logging

from airflow.utils.log.logging_mixin import StreamLogWriter

logger = logging.getLogger("custom-logger")

with redirect_stdout(StreamLogWriter(logger, logging.INFO)), \
        redirect_stderr(StreamLogWriter(logger, logging.WARN)):
    print("I Love Airflow")


Now, additional arguments passed to BaseOperator cause an exception. Previous versions of Airflow took additional arguments and displayed a message on the console. When the message was not noticed by users, it caused very difficult to detect errors.

In order to restore the previous behavior, you must set an True in the allow_illegal_arguments option of section [operators] in the airflow.cfg file. In the future it is possible to completely delete this option.


Passing store_serialized_dags argument to DagBag.init and accessing DagBag.store_serialized_dags property are deprecated and will be removed in future versions.

Previous signature:

    include_examples=conf.getboolean('core', 'LOAD_EXAMPLES'),
    safe_mode=conf.getboolean('core', 'DAG_DISCOVERY_SAFE_MODE'),


    include_examples=conf.getboolean('core', 'LOAD_EXAMPLES'),
    safe_mode=conf.getboolean('core', 'DAG_DISCOVERY_SAFE_MODE'),

If you were using positional arguments, it requires no change but if you were using keyword arguments, please change store_serialized_dags to read_dags_from_db.

Similarly, if you were using DagBag().store_serialized_dags property, change it to DagBag().read_dags_from_db.

Changes in google provider package

We strive to ensure that there are no changes that may affect the end user and your Python files, but this release may contain changes that will require changes to your configuration, DAG Files or other integration e.g. custom operators.

Only changes unique to this provider are described here. You should still pay attention to the changes that have been made to the core (including core operators) as they can affect the integration behavior of this provider.

This section describes the changes that have been made, and what you need to do to update your if you use operators or hooks which integrate with Google services (including Google Cloud - GCP).

Direct impersonation added to operators communicating with Google services

Directly impersonating a service account has been made possible for operators communicating with Google services via new argument called impersonation_chain (google_impersonation_chain in case of operators that also communicate with services of other cloud providers). As a result, GCSToS3Operator no longer derivatives from GCSListObjectsOperator.

Normalize gcp_conn_id for Google Cloud

Previously not all hooks and operators related to Google Cloud use gcp_conn_id as parameter for GCP connection. There is currently one parameter which apply to most services. Parameters like datastore_conn_id, bigquery_conn_id, google_cloud_storage_conn_id and similar have been deprecated. Operators that require two connections are not changed.

Following components were affected by normalization:

  • airflow.operators.sql_to_gcs.BaseSQLToGoogleCloudStorageOperator
  • airflow.operators.adls_to_gcs.AdlsToGoogleCloudStorageOperator
  • airflow.operators.gcs_to_s3.GoogleCloudStorageToS3Operator
  • airflow.operators.gcs_to_gcs.GoogleCloudStorageToGoogleCloudStorageOperator
  • airflow.operators.bigquery_to_gcs.BigQueryToCloudStorageOperator
  • airflow.operators.local_to_gcs.FileToGoogleCloudStorageOperator
  • airflow.operators.cassandra_to_gcs.CassandraToGoogleCloudStorageOperator
  • airflow.operators.bigquery_to_bigquery.BigQueryToBigQueryOperator

Changes to import paths and names of GCP operators and hooks

According to AIP-21 operators related to Google Cloud has been moved from contrib to core. The following table shows changes in import paths.

Old path New path
airflow.contrib.operators.adls_to_gcs.AdlsToGoogleCloudStorageOperator airflow.operators.adls_to_gcs.AdlsToGoogleCloudStorageOperator
airflow.contrib.operators.bigquery_to_bigquery.BigQueryToBigQueryOperator airflow.operators.bigquery_to_bigquery.BigQueryToBigQueryOperator
airflow.contrib.operators.bigquery_to_gcs.BigQueryToCloudStorageOperator airflow.operators.bigquery_to_gcs.BigQueryToCloudStorageOperator
airflow.contrib.operators.bigquery_to_mysql_operator.BigQueryToMySqlOperator airflow.operators.bigquery_to_mysql.BigQueryToMySqlOperator
airflow.contrib.operators.gcs_to_bq.GoogleCloudStorageToBigQueryOperator airflow.operators.gcs_to_bq.GoogleCloudStorageToBigQueryOperator
airflow.contrib.operators.gcs_to_gcs.GoogleCloudStorageToGoogleCloudStorageOperator airflow.operators.gcs_to_gcs.GoogleCloudStorageToGoogleCloudStorageOperator
airflow.contrib.operators.gcs_to_s3.GoogleCloudStorageToS3Operator airflow.operators.gcs_to_s3.GCSToS3Operator
airflow.contrib.operators.mssql_to_gcs.MsSqlToGoogleCloudStorageOperator airflow.operators.mssql_to_gcs.MsSqlToGoogleCloudStorageOperator
airflow.contrib.operators.mysql_to_gcs.MySqlToGoogleCloudStorageOperator airflow.operators.mysql_to_gcs.MySqlToGoogleCloudStorageOperator
airflow.contrib.operators.postgres_to_gcs_operator.PostgresToGoogleCloudStorageOperator airflow.operators.postgres_to_gcs.PostgresToGoogleCloudStorageOperator
airflow.contrib.operators.sql_to_gcs.BaseSQLToGoogleCloudStorageOperator airflow.operators.sql_to_gcs.BaseSQLToGoogleCloudStorageOperator

Unify default conn_id for Google Cloud

Previously not all hooks and operators related to Google Cloud use google_cloud_default as a default conn_id. There is currently one default variant. Values like google_cloud_storage_default, bigquery_default, google_cloud_datastore_default have been deprecated. The configuration of existing relevant connections in the database have been preserved. To use those deprecated GCP conn_id, you need to explicitly pass their conn_id into operators/hooks. Otherwise, google_cloud_default will be used as GCP's conn_id by default.

To use project_id argument consistently across GCP hooks and operators, we did the following changes:

  • Changed order of arguments in DataflowHook.start_python_dataflow. Uses with positional arguments may break.
  • Changed order of arguments in DataflowHook.is_job_dataflow_running. Uses with positional arguments may break.
  • Changed order of arguments in DataflowHook.cancel_job. Uses with positional arguments may break.
  • Added optional project_id argument to DataflowCreateJavaJobOperator constructor.
  • Added optional project_id argument to DataflowTemplatedJobStartOperator constructor.
  • Added optional project_id argument to DataflowCreatePythonJobOperator constructor.

To provide more precise control in handling of changes to objects in underlying GCS Bucket the constructor of this sensor now has changed.

  • Old Behavior: This constructor used to optionally take previous_num_objects: int.
  • New replacement constructor kwarg: previous_objects: Optional[Set[str]].

Most users would not specify this argument because the bucket begins empty and the user wants to treat any files as new.

Example of Updating usage of this sensor: Users who used to call:

GCSUploadSessionCompleteSensor(bucket='my_bucket', prefix='my_prefix', previous_num_objects=1)

Will now call:

GCSUploadSessionCompleteSensor(bucket='my_bucket', prefix='my_prefix', previous_num_objects={'.keep'})

Where '.keep' is a single file at your prefix that the sensor should not consider new.

To simplify BigQuery operators (no need of Cursor) and standardize usage of hooks within all GCP integration methods from BiqQueryBaseCursor were moved to BigQueryHook. Using them by from Cursor object is still possible due to preserved backward compatibility but they will raise DeprecationWarning. The following methods were moved:

Old path New path

Since BigQuery is the part of the GCP it was possible to simplify the code by handling the exceptions by usage of the decorator however it changes exceptions raised by the following methods:

  • raises AirflowException instead of Exception.
  • raises AirflowException instead of ValueError.
  • raises AirflowException instead of ValueError.

Idempotency was added to BigQueryCreateEmptyTableOperator and BigQueryCreateEmptyDatasetOperator. But to achieve that try / except clause was removed from create_empty_dataset and create_empty_table methods of BigQueryHook.

The change in GCP operators implies that GCP Hooks for those operators require now keyword parameters rather than positional ones in all methods where project_id is used. The methods throw an explanatory exception in case they are called using positional parameters.

Other GCP hooks are unaffected.

In the PubSubPublishOperator and PubSubHook.publsh method the data field in a message should be bytestring (utf-8 encoded) rather than base64 encoded string.

Due to the normalization of the parameters within GCP operators and hooks a parameters like project or topic_project are deprecated and will be substituted by parameter project_id. In PubSubHook.create_subscription hook method in the parameter subscription_project is replaced by subscription_project_id. Template fields are updated accordingly and old ones may not work.

It is required now to pass key-word only arguments to PubSub hook.

These changes are not backward compatible.

The gcp_conn_id parameter in GKEPodOperator is required. In previous versions, it was possible to pass the None value to the gcp_conn_id in the GKEStartPodOperator operator, which resulted in credentials being determined according to the Application Default Credentials strategy.

Now this parameter requires a value. To restore the previous behavior, configure the connection without specifying the service account.

Detailed information about connection management is available: Google Cloud Connection.

  • The following parameters have been replaced in all the methods in GCSHook:

    • bucket is changed to bucket_name
    • object is changed to object_name
  • The maxResults parameter in GoogleCloudStorageHook.list has been renamed to max_results for consistency.

The 'properties' and 'jars' properties for the Dataproc related operators (DataprocXXXOperator) have been renamed from dataproc_xxxx_properties and dataproc_xxx_jars to dataproc_properties and dataproc_jarsrespectively. Arguments for dataproc_properties dataproc_jars

To obtain pylint compatibility the filter argument in CloudDataTransferServiceCreateJobOperator has been renamed to request_filter.

To obtain pylint compatibility the filter argument in CloudDataTransferServiceHook.list_transfer_job and CloudDataTransferServiceHook.list_transfer_operations has been renamed to request_filter.

In general all hook methods are decorated with @GoogleBaseHook.fallback_to_default_project_id thus parameters to hook can only be passed via keyword arguments.

  • create_empty_table method accepts now table_resource parameter. If provided all other parameters are ignored.
  • create_empty_dataset will now use values from dataset_reference instead of raising error if parameters were passed in dataset_reference and as arguments to method. Additionally validation of dataset_reference is done using Dataset.from_api_repr. Exception and log messages has been changed.
  • update_dataset requires now new fields argument (breaking change)
  • delete_dataset has new signature (dataset_id, project_id, ...) previous one was (project_id, dataset_id, ...) (breaking change)
  • get_tabledata returns list of rows instead of API response in dict format. This method is deprecated in favor of list_rows. (breaking change)

Change python3 as Dataflow Hooks/Operators default interpreter

Now the py_interpreter argument for DataFlow Hooks/Operators has been changed from python2 to python3.

To simplify the code, the decorator provide_gcp_credential_file has been moved from the inner-class.

Instead of @GoogleBaseHook._Decorators.provide_gcp_credential_file, you should write @GoogleBaseHook.provide_gcp_credential_file

It is highly recommended to have 1TB+ disk size for Dataproc to have sufficient throughput:

Hence, the default value for master_disk_size in DataprocCreateClusterOperator has been changed from 500GB to 1TB.

<airflow class="providers google c"></airflow>loud.operators.bigquery.BigQueryGetDatasetTablesOperator

We changed signature of BigQueryGetDatasetTablesOperator.


BigQueryGetDatasetTablesOperator(dataset_id: str, dataset_resource: dict, ...)


BigQueryGetDatasetTablesOperator(dataset_resource: dict, dataset_id: Optional[str] = None, ...)

Changes in amazon provider package

We strive to ensure that there are no changes that may affect the end user, and your Python files, but this release may contain changes that will require changes to your configuration, DAG Files or other integration e.g. custom operators.

Only changes unique to this provider are described here. You should still pay attention to the changes that have been made to the core (including core operators) as they can affect the integration behavior of this provider.

This section describes the changes that have been made, and what you need to do to update your if you use operators or hooks which integrate with Amazon services (including Amazon Web Service - AWS).

Migration of AWS components

All AWS components (hooks, operators, sensors, example DAGs) will be grouped together as decided in AIP-21. Migrated components remain backwards compatible but raise a DeprecationWarning when imported from the old module. Migrated are:

Old path New path
private attrs and methods on AWSBatchOperator

The default value for the aws_conn_id was accidentally set to 's3_default' instead of 'aws_default' in some of the emr operators in previous versions. This was leading to EmrStepSensor not being able to find their corresponding emr cluster. With the new changes in the EmrAddStepsOperator, EmrTerminateJobFlowOperator and EmrCreateJobFlowOperator this issue is solved.

The AwsBatchOperator was refactored to extract an AwsBatchClient (and inherit from it). The changes are mostly backwards compatible and clarify the public API for these classes; some private methods on AwsBatchOperator for polling a job status were relocated and renamed to surface new public methods on AwsBatchClient (and via inheritance on AwsBatchOperator). A couple of job attributes are renamed on an instance of AwsBatchOperator; these were mostly used like private attributes but they were surfaced in the public API, so any use of them needs to be updated as follows:

  • AwsBatchOperator().jobId -> AwsBatchOperator().job_id
  • AwsBatchOperator().jobName -> AwsBatchOperator().job_name

The AwsBatchOperator gets a new option to define a custom model for waiting on job status changes. The AwsBatchOperator can use a new waiters parameter, an instance of AwsBatchWaiters, to specify that custom job waiters will be used to monitor a batch job. See the latest API documentation for details.

Replace parameter max_retires with max_retries to fix typo.

Note: The order of arguments has changed for check_for_prefix. The bucket_name is now optional. It falls back to the connection schema attribute. The delete_objects now returns None instead of a response, since the method now makes multiple api requests when the keys list length is > 1000.

Changes in other provider packages

We strive to ensure that there are no changes that may affect the end user and your Python files, but this release may contain changes that will require changes to your configuration, DAG Files or other integration e.g. custom operators.

Only changes unique to providers are described here. You should still pay attention to the changes that have been made to the core (including core operators) as they can affect the integration behavior of this provider.

This section describes the changes that have been made, and what you need to do to update your if you use any code located in airflow.providers package.

Changed return type of list_prefixes and list_keys methods in S3Hook

Previously, the list_prefixes and list_keys methods returned None when there were no results. The behavior has been changed to return an empty list instead of None in this case.

Removed Hipchat integration

Hipchat has reached end of life and is no longer available.

For more information please see


Replace parameter sandbox with domain. According to change in simple-salesforce package.

Rename sign_in function to get_conn.


Rename parameter name from format to segment_format in PinotAdminHook function create_segment fro pylint compatible


Rename parameter name from filter to partition_filter in HiveMetastoreHook function get_partitions for pylint compatible


Remove unnecessary parameter nlst in FTPHook function list_directory for pylint compatible


Remove unnecessary parameter open in PostgresHook function copy_expert for pylint compatible


Change parameter name from visibleTo to visible_to in OpsgenieAlertOperator for pylint compatible




  • The order of arguments has changed for has_mail_attachment, retrieve_mail_attachments and download_mail_attachments.
  • A new mail_filter argument has been added to each of those.


The HTTPHook is now secured by default: verify=True (before: verify=False) This can be overwritten by using the extra_options param as {'verify': False}.


  • upgraded cloudant version from >=0.5.9,<2.0 to >=2.0
  • removed the use of the schema attribute in the connection
  • removed db function since the database object can also be retrieved by calling cloudant_session['database_name']

For example:

from airflow.providers.cloudant.hooks.cloudant import CloudantHook

with CloudantHook().get_conn() as cloudant_session:
    database = cloudant_session['database_name']

See the docs for more information on how to use the new cloudant version.


When initializing a Snowflake hook or operator, the value used for snowflake_conn_id was always snowflake_conn_id, regardless of whether or not you specified a value for it. The default snowflake_conn_id value is now switched to snowflake_default for consistency and will be properly overridden when specified.

Other changes

This release also includes changes that fall outside any of the sections above.

Standardised "extra" requirements

We standardised the Extras names and synchronized providers package names with the main airflow extras.

We deprecated a number of extras in 2.0.

Deprecated extras New extras
atlas apache.atlas
aws amazon
cassandra apache.cassandra
druid apache.druid
gcp google
gcp_api google
hdfs apache.hdfs
hive apache.hive
kubernetes cncf.kubernetes
mssql microsoft.mssql
pinot apache.pinot
webhdfs apache.webhdfs
winrm apache.winrm

For example:

If you want to install integration for Microsoft Azure, then instead of pip install apache-airflow[atlas] you should use pip install apache-airflow[apache.atlas].


On November 2020, new version of PIP (20.3) has been released with a new, 2020 resolver. This resolver does not yet work with Apache Airflow and might lead to errors in installation - depends on your choice of extras. In order to install Airflow you need to either downgrade pip to version 20.2.4 pip install --upgrade pip==20.2.4 or, in case you use Pip 20.3, you need to add option --use-deprecated legacy-resolver to your pip install command.

If you want to install integration for Microsoft Azure, then instead of

pip install 'apache-airflow[azure_blob_storage,azure_data_lake,azure_cosmos,azure_container_instances]'

you should execute pip install 'apache-airflow[azure]'

If you want to install integration for Amazon Web Services, then instead of pip install 'apache-airflow[s3,emr]', you should execute pip install 'apache-airflow[aws]'

The deprecated extras will be removed in 3.0.

Simplify the response payload of endpoints /dag_stats and /task_stats

The response of endpoints /dag_stats and /task_stats help UI fetch brief statistics about DAGs and Tasks. The format was like

    "example_http_operator": [
            "state": "success",
            "count": 0,
            "dag_id": "example_http_operator",
            "color": "green"
            "state": "running",
            "count": 0,
            "dag_id": "example_http_operator",
            "color": "lime"

The dag_id was repeated in the payload, which makes the response payload unnecessarily bigger.

Now the dag_id will not appear repeated in the payload, and the response format is like

    "example_http_operator": [
            "state": "success",
            "count": 0,
            "color": "green"
            "state": "running",
            "count": 0,
            "color": "lime"

Airflow 1.10.14

[scheduler] max_threads config has been renamed to [scheduler] parsing_processes

From Airflow 1.10.14, max_threads config under [scheduler] section has been renamed to parsing_processes.

This is to align the name with the actual code where the Scheduler launches the number of processes defined by [scheduler] parsing_processes to parse the DAG files.

Airflow CLI changes in line with 2.0

The Airflow CLI has been organized so that related commands are grouped together as subcommands, which means that if you use these commands in your scripts, they will now raise a DeprecationWarning and you have to make changes to them before you upgrade to Airflow 2.0.

This section describes the changes that have been made, and what you need to do to update your script.

The ability to manipulate users from the command line has been changed. airflow create_user, airflow delete_user and airflow list_users has been grouped to a single command airflow users with optional flags create, list and delete.

The airflow list_dags command is now airflow dags list, airflow pause is airflow dags pause, etc.

In Airflow 1.10 and 2.0 there is an airflow config command but there is a difference in behavior. In Airflow 1.10, it prints all config options while in Airflow 2.0, it's a command group. airflow config is now airflow config list. You can check other options by running the command airflow config --help

Compatibility with the old CLI has been maintained, but they will no longer appear in the help

You can learn about the commands by running airflow --help. For example to get help about the celery group command, you have to run the help command: airflow celery --help.

Old command New command Group
airflow worker airflow celery worker celery
airflow flower airflow celery flower celery
airflow trigger_dag airflow dags trigger dags
airflow delete_dag airflow dags delete dags
airflow show_dag airflow dags show dags
airflow list_dag airflow dags list dags
airflow dag_status airflow dags status dags
airflow backfill airflow dags backfill dags
airflow list_dag_runs airflow dags list-runs dags
airflow pause airflow dags pause dags
airflow unpause airflow dags unpause dags
airflow next_execution airflow dags next-execution dags
airflow test airflow tasks test tasks
airflow clear airflow tasks clear tasks
airflow list_tasks airflow tasks list tasks
airflow task_failed_deps airflow tasks failed-deps tasks
airflow task_state airflow tasks state tasks
airflow run airflow tasks run tasks
airflow render airflow tasks render tasks
airflow initdb airflow db init db
airflow resetdb airflow db reset db
airflow upgradedb airflow db upgrade db
airflow checkdb airflow db check db
airflow shell airflow db shell db
airflow pool airflow pools pools
airflow create_user airflow users create users
airflow delete_user airflow users delete users
airflow list_users airflow users list users
airflow rotate_fernet_key airflow rotate-fernet-key
airflow sync_perm airflow sync-perm

Airflow 1.10.13

TimeSensor is now timezone aware

Previously TimeSensor always compared the target_time with the current time in UTC.

Now it will compare target_time with the current time in the timezone of the DAG, defaulting to the default_timezone in the global config.

Removed Kerberos support for HDFS hook

The HDFS hook's Kerberos support has been removed due to removed python-krbV dependency from PyPI and generally lack of support for SSL in Python3 (Snakebite-py3 we use as dependency has no support for SSL connection to HDFS).

SSL support still works for WebHDFS hook.

Unify user session lifetime configuration

In previous version of Airflow user session lifetime could be configured by session_lifetime_days and force_log_out_after options. In practise only session_lifetime_days had impact on session lifetime, but it was limited to values in day. We have removed mentioned options and introduced new session_lifetime_minutes option which simplify session lifetime configuration.


force_log_out_after = 0
session_lifetime_days = 30


session_lifetime_minutes = 43200

Adding Operators, Hooks and Sensors via Airflow Plugins is deprecated

The ability to import Operators, Hooks and Sensors via the plugin mechanism has been deprecated and will raise warnings in Airflow 1.10.13 and will be removed completely in Airflow 2.0.

Check to see how you can create and import Custom Hooks, Operators and Sensors.

Airflow 1.10.12

Clearing tasks skipped by SkipMixin will skip them

Previously, when tasks skipped by SkipMixin (such as BranchPythonOperator, BaseBranchOperator and ShortCircuitOperator) are cleared, they execute. Since 1.10.12, when such skipped tasks are cleared, they will be skipped again by the newly introduced NotPreviouslySkippedDep.

The pod_mutation_hook function will now accept a kubernetes V1Pod object

As of airflow 1.10.12, using the airflow.contrib.kubernetes.Pod class in the pod_mutation_hook is now deprecated. Instead we recommend that users treat the pod parameter as a kubernetes.client.models.V1Pod object. This means that users now have access to the full Kubernetes API when modifying airflow pods

pod_template_file option now available in the KubernetesPodOperator

Users can now offer a path to a yaml for the KubernetesPodOperator using the pod_template_file parameter.

Airflow 1.10.11

Use NULL as default value for dag.description

Now use NULL as default value for dag.description in dag table

Restrict editing DagRun State in the old UI (Flask-admin based UI)

Before 1.10.11 it was possible to edit DagRun State in the /admin/dagrun/ page to any text.

In Airflow 1.10.11+, the user can only choose the states from the list.

Experimental API will deny all request by default.

The previous default setting was to allow all API requests without authentication, but this poses security risks to users who miss this fact. This changes the default for new installs to deny all requests by default.

Note: This will not change the behavior for existing installs, please update check your airflow.cfg

If you wish to have the experimental API work, and aware of the risks of enabling this without authentication (or if you have your own authentication layer in front of Airflow) you can get the previous behaviour on a new install by setting this in your airflow.cfg:

auth_backend = airflow.api.auth.backend.default

XCom Values can no longer be added or changed from the Webserver

Since XCom values can contain pickled data, we would no longer allow adding or changing XCom values from the UI.

Default for run_as_user configured has been changed to 50000 from 0

The UID to run the first process of the Worker PODs when using has been changed to 50000 from the previous default of 0. The previous default was an empty string but the code used 0 if it was empty string.


run_as_user =


run_as_user = 50000

This is done to avoid running the container as root user.

Airflow 1.10.10

Setting Empty string to a Airflow Variable will return an empty string

Previously when you set an Airflow Variable with an empty string (''), the value you used to get back was None. This will now return an empty string (''')


>> Variable.set('test_key', '')
>> Variable.get('test_key')

The above code returned None previously, now it will return ''.

Make behavior of none_failed trigger rule consistent with documentation

The behavior of the none_failed trigger rule is documented as "all parents have not failed (failed or upstream_failed) i.e. all parents have succeeded or been skipped." As previously implemented, the actual behavior would skip if all parents of a task had also skipped.

Add new trigger rule none_failed_or_skipped

The fix to none_failed trigger rule breaks workflows that depend on the previous behavior. If you need the old behavior, you should change the tasks with none_failed trigger rule to none_failed_or_skipped.

Success Callback will be called when a task in marked as success from UI

When a task is marked as success by a user from Airflow UI - on_success_callback will be called

Airflow 1.10.9

No breaking changes.

Airflow 1.10.8

Failure callback will be called when task is marked failed

When task is marked failed by user or task fails due to system failures - on failure call back will be called as part of clean up

See AIRFLOW-5621 for details

Airflow 1.10.7

Changes in experimental API execution_date microseconds replacement

The default behavior was to strip the microseconds (and milliseconds, etc) off of all dag runs triggered by by the experimental REST API. The default behavior will change when an explicit execution_date is passed in the request body. It will also now be possible to have the execution_date generated, but keep the microseconds by sending replace_microseconds=false in the request body. The default behavior can be overridden by sending replace_microseconds=true along with an explicit execution_date

Infinite pool size and pool size query optimisation

Pool size can now be set to -1 to indicate infinite size (it also includes optimisation of pool query which lead to poor task n^2 performance of task pool queries in MySQL).

Viewer won't have edit permissions on DAG view.

Google Cloud Storage Hook

The GoogleCloudStorageDownloadOperator can either write to a supplied filename or return the content of a file via xcom through store_to_xcom_key - both options are mutually exclusive.

Airflow 1.10.6

BaseOperator::render_template function signature changed

Previous versions of the BaseOperator::render_template function required an attr argument as the first positional argument, along with content and context. This function signature was changed in 1.10.6 and the attr argument is no longer required (or accepted).

In order to use this function in subclasses of the BaseOperator, the attr argument must be removed:

result = self.render_template('myattr', self.myattr, context)  # Pre-1.10.6 call
result = self.render_template(self.myattr, context)  # Post-1.10.6 call

Changes to aws_default Connection's default region

The region of Airflow's default connection to AWS (aws_default) was previously set to us-east-1 during installation.

The region now needs to be set manually, either in the connection screens in Airflow, via the ~/.aws config files, or via the AWS_DEFAULT_REGION environment variable.

Some DAG Processing metrics have been renamed

The following metrics are deprecated and won't be emitted in Airflow 2.0:

  • scheduler.dagbag.errors and dagbag_import_errors -- use dag_processing.import_errors instead
  • dag_file_processor_timeouts -- use dag_processing.processor_timeouts instead
  • collect_dags -- use dag_processing.total_parse_time instead
  • dag.loading-duration.<basename> -- use dag_processing.last_duration.<basename> instead
  • dag_processing.last_runtime.<basename> -- use dag_processing.last_duration.<basename> instead

Airflow 1.10.5

No breaking changes.

Airflow 1.10.4

Export MySQL timestamps as UTC

MySqlToGoogleCloudStorageOperator now exports TIMESTAMP columns as UTC by default, rather than using the default timezone of the MySQL server. This is the correct behavior for use with BigQuery, since BigQuery assumes that TIMESTAMP columns without time zones are in UTC. To preserve the previous behavior, set ensure_utc to False.

Changes to DatastoreHook

  • removed argument version from get_conn function and added it to the hook's __init__ function instead and renamed it to api_version
  • renamed the partialKeys argument of function allocate_ids to partial_keys

Changes to GoogleCloudStorageHook

  • the discovery-based api (googleapiclient.discovery) used in GoogleCloudStorageHook is now replaced by the recommended client based api (google-cloud-storage). To know the difference between both the libraries, read PR: #5054

  • as a part of this replacement, the multipart & num_retries parameters for GoogleCloudStorageHook.upload method have been deprecated.

    The client library uses multipart upload automatically if the object/blob size is more than 8 MB - source code. The client also handles retries automatically

  • the generation parameter is deprecated in GoogleCloudStorageHook.delete and GoogleCloudStorageHook.insert_object_acl.

Updating to google-cloud-storage >= 1.16 changes the signature of the upstream client.get_bucket() method from get_bucket(bucket_name: str) to get_bucket(bucket_or_name: Union[str, Bucket]). This method is not directly exposed by the airflow hook, but any code accessing the connection directly (GoogleCloudStorageHook().get_conn().get_bucket(...) or similar) will need to be updated.

Changes in writing Logs to Elasticsearch

The elasticsearch_ prefix has been removed from all config items under the [elasticsearch] section. For example elasticsearch_host is now just host.

Removal of non_pooled_task_slot_count and non_pooled_backfill_task_slot_count

non_pooled_task_slot_count and non_pooled_backfill_task_slot_count are removed in favor of a real pool, e.g. default_pool.

By default tasks are running in default_pool. default_pool is initialized with 128 slots and user can change the number of slots through UI/CLI. default_pool cannot be removed.

pool config option in Celery section to support different Celery pool implementation

The new pool config option allows users to choose different pool implementation. Default value is "prefork", while choices include "prefork" (default), "eventlet", "gevent" or "solo". This may help users achieve better concurrency performance in different scenarios.

For more details about Celery pool implementation, please refer to:

Change to method signature in BaseOperator and DAG classes

The signature of the get_task_instances method in the BaseOperator and DAG classes has changed. The change does not change the behavior of the method in either case.

For BaseOperator

Old signature:

def get_task_instances(self, session, start_date=None, end_date=None):

New signature:

def get_task_instances(self, start_date=None, end_date=None, session=None):


Old signature:

def get_task_instances(
    self, session, start_date=None, end_date=None, state=None):

New signature:

def get_task_instances(
    self, start_date=None, end_date=None, state=None, session=None):

In either case, it is necessary to rewrite calls to the get_task_instances method that currently provide the session positional argument. New calls to this method look like:

# if you can rely on @provide_session
# if you need to provide the session

Airflow 1.10.3

New dag_discovery_safe_mode config option

If dag_discovery_safe_mode is enabled, only check files for DAGs if they contain the strings "airflow" and "DAG". For backwards compatibility, this option is enabled by default.

RedisPy dependency updated to v3 series

If you are using the Redis Sensor or Hook you may have to update your code. See redis-py porting instructions to check if your code might be affected (MSET, MSETNX, ZADD, and ZINCRBY all were, but read the full doc).


It is no longer required to set one of the environment variables to avoid a GPL dependency. Airflow will now always use text-unidecode if unidecode was not installed before.

new sync_parallelism config option in celery section

The new sync_parallelism config option will control how many processes CeleryExecutor will use to fetch celery task state in parallel. Default value is max(1, number of cores - 1)

Rename of BashTaskRunner to StandardTaskRunner

BashTaskRunner has been renamed to StandardTaskRunner. It is the default task runner so you might need to update your config.

task_runner = StandardTaskRunner

Modification to config file discovery

If the AIRFLOW_CONFIG environment variable was not set and the ~/airflow/airflow.cfg file existed, airflow previously used ~/airflow/airflow.cfg instead of $AIRFLOW_HOME/airflow.cfg. Now airflow will discover its config file using the $AIRFLOW_CONFIG and $AIRFLOW_HOME environment variables rather than checking for the presence of a file.

Changes in Google Cloud related operators

Most GCP-related operators have now optional PROJECT_ID parameter. In case you do not specify it, the project id configured in GCP Connection is used. There will be an AirflowException thrown in case PROJECT_ID parameter is not specified and the connection used has no project id defined. This change should be backwards compatible as earlier version of the operators had PROJECT_ID mandatory.

Operators involved:

  • GCP Compute Operators
    • GceInstanceStartOperator
    • GceInstanceStopOperator
    • GceSetMachineTypeOperator
  • GCP Function Operators
    • GcfFunctionDeployOperator
  • GCP Cloud SQL Operators
    • CloudSqlInstanceCreateOperator
    • CloudSqlInstancePatchOperator
    • CloudSqlInstanceDeleteOperator
    • CloudSqlInstanceDatabaseCreateOperator
    • CloudSqlInstanceDatabasePatchOperator
    • CloudSqlInstanceDatabaseDeleteOperator

Other GCP operators are unaffected.

Changes in Google Cloud related hooks

The change in GCP operators implies that GCP Hooks for those operators require now keyword parameters rather than positional ones in all methods where project_id is used. The methods throw an explanatory exception in case they are called using positional parameters.

Hooks involved:

  • GceHook
  • GcfHook
  • CloudSqlHook

Other GCP hooks are unaffected.

Changed behaviour of using default value when accessing variables

It's now possible to use None as a default value with the default_var parameter when getting a variable, e.g.

foo = Variable.get("foo", default_var=None)
if foo is None:

(Note: there is already Variable.setdefault() which me be helpful in some cases.)

This changes the behaviour if you previously explicitly provided None as a default value. If your code expects a KeyError to be thrown, then don't pass the default_var argument.

Removal of airflow_home config setting

There were previously two ways of specifying the Airflow "home" directory (~/airflow by default): the AIRFLOW_HOME environment variable, and the airflow_home config setting in the [core] section.

If they had two different values different parts of the code base would end up with different values. The config setting has been deprecated, and you should remove the value from the config file and set AIRFLOW_HOME environment variable if you need to use a non default value for this.

(Since this setting is used to calculate what config file to load, it is not possible to keep just the config option)

Change of two methods signatures in GCPTransferServiceHook

The signature of the create_transfer_job method in GCPTransferServiceHook class has changed. The change does not change the behavior of the method.

Old signature:

def create_transfer_job(self, description, schedule, transfer_spec, project_id=None):

New signature:

def create_transfer_job(self, body):

It is necessary to rewrite calls to method. The new call looks like this:

body = {
  'status': 'ENABLED',
  'projectId': project_id,
  'description': description,
  'transferSpec': transfer_spec,
  'schedule': schedule,

The change results from the unification of all hooks and adjust to the official recommendations for the Google Cloud.

The signature of wait_for_transfer_job method in GCPTransferServiceHook has changed.

Old signature:

def wait_for_transfer_job(self, job):

New signature:

def wait_for_transfer_job(self, job, expected_statuses=(GcpTransferOperationStatus.SUCCESS, )):

The behavior of wait_for_transfer_job has changed:

Old behavior:

wait_for_transfer_job would wait for the SUCCESS status in specified jobs operations.

New behavior:

You can now specify an array of expected statuses. wait_for_transfer_job now waits for any of them.

The default value of expected_statuses is SUCCESS so that change is backwards compatible.

Moved two classes to different modules

The class GoogleCloudStorageToGoogleCloudStorageTransferOperator has been moved from airflow.contrib.operators.gcs_to_gcs_transfer_operator to airflow.contrib.operators.gcp_transfer_operator

the class S3ToGoogleCloudStorageTransferOperator has been moved from airflow.contrib.operators.s3_to_gcs_transfer_operator to airflow.contrib.operators.gcp_transfer_operator

The change was made to keep all the operators related to GCS Transfer Services in one file.

The previous imports will continue to work until Airflow 2.0

Fixed typo in --driver-class-path in SparkSubmitHook

The driver_classapth argument to SparkSubmit Hook and Operator was generating --driver-classpath on the spark command line, but this isn't a valid option to spark.

The argument has been renamed to driver_class_path and the option it generates has been fixed.

Airflow 1.10.2

New dag_processor_manager_log_location config option

The DAG parsing manager log now by default will be log into a file, where its location is controlled by the new dag_processor_manager_log_location config option in core section.

DAG level Access Control for new RBAC UI

Extend and enhance new Airflow RBAC UI to support DAG level ACL. Each dag now has two permissions(one for write, one for read) associated('can_dag_edit', 'can_dag_read'). The admin will create new role, associate the dag permission with the target dag and assign that role to users. That user can only access / view the certain dags on the UI that he has permissions on. If a new role wants to access all the dags, the admin could associate dag permissions on an artificial view(all_dags) with that role.

We also provide a new cli command(sync_perm) to allow admin to auto sync permissions.

Modification to ts_nodash macro

ts_nodash previously contained TimeZone information along with execution date. For Example: 20150101T000000+0000. This is not user-friendly for file or folder names which was a popular use case for ts_nodash. Hence this behavior has been changed and using ts_nodash will no longer contain TimeZone information, restoring the pre-1.10 behavior of this macro. And a new macro ts_nodash_with_tz has been added which can be used to get a string with execution date and timezone info without dashes.


  • ts_nodash: 20150101T000000
  • ts_nodash_with_tz: 20150101T000000+0000

Semantics of next_ds/prev_ds changed for manually triggered runs

next_ds/prev_ds now map to execution_date instead of the next/previous schedule-aligned execution date for DAGs triggered in the UI.

User model changes

This patch changes the User.superuser field from a hardcoded boolean to a Boolean() database column. User.superuser will default to False, which means that this privilege will have to be granted manually to any users that may require it.

For example, open a Python shell and

from airflow import models, settings

session = settings.Session()
users = session.query(models.User).all()  # [admin, regular_user]

users[1].superuser  # False

admin = users[0]
admin.superuser = True

Custom auth backends interface change

We have updated the version of flask-login we depend upon, and as a result any custom auth backends might need a small change: is_active, is_authenticated, and is_anonymous should now be properties. What this means is if previously you had this in your user class

def is_active(self):

then you need to change it like this

def is_active(self):

Support autodetected schemas to GoogleCloudStorageToBigQueryOperator

GoogleCloudStorageToBigQueryOperator is now support schema auto-detection is available when you load data into BigQuery. Unfortunately, changes can be required.

If BigQuery tables are created outside of airflow and the schema is not defined in the task, multiple options are available:

define a schema_fields:


or define a schema_object:


or enabled autodetect of schema:


Airflow 1.10.1

min_file_parsing_loop_time config option temporarily disabled

The scheduler.min_file_parsing_loop_time config option has been temporarily removed due to some bugs.

StatsD Metrics

The scheduler_heartbeat metric has been changed from a gauge to a counter. Each loop of the scheduler will increment the counter by 1. This provides a higher degree of visibility and allows for better integration with Prometheus using the StatsD Exporter. The scheduler's activity status can be determined by graphing and alerting using a rate of change of the counter. If the scheduler goes down, the rate will drop to 0.

EMRHook now passes all of connection's extra to CreateJobFlow API

EMRHook.create_job_flow has been changed to pass all keys to the create_job_flow API, rather than just specific known keys for greater flexibility.

However prior to this release the "emr_default" sample connection that was created had invalid configuration, so creating EMR clusters might fail until your connection is updated. (Ec2KeyName, Ec2SubnetId, TerminationProtection and KeepJobFlowAliveWhenNoSteps were all top-level keys when they should be inside the "Instances" dict)

LDAP Auth Backend now requires TLS

Connecting to an LDAP server over plain text is not supported anymore. The certificate presented by the LDAP server must be signed by a trusted certificate, or you must provide the cacert option under [ldap] in the config file.

If you want to use LDAP auth backend without TLS then you will have to create a custom-auth backend based on

Airflow 1.10

Installation and upgrading requires setting SLUGIFY_USES_TEXT_UNIDECODE=yes in your environment or AIRFLOW_GPL_UNIDECODE=yes. In case of the latter a GPL runtime dependency will be installed due to a dependency (python-nvd3 -> python-slugify -> unidecode).

Replace DataProcHook.await calls to DataProcHook.wait

The method name was changed to be compatible with the Python 3.7 async/await keywords

Setting UTF-8 as default mime_charset in email utils

Add a configuration variable(default_dag_run_display_number) to control numbers of dag run for display

Add a configuration variable(default_dag_run_display_number) under webserver section to control the number of dag runs to show in UI.

Default executor for SubDagOperator is changed to SequentialExecutor

New Webserver UI with Role-Based Access Control

The current webserver UI uses the Flask-Admin extension. The new webserver UI uses the Flask-AppBuilder (FAB) extension. FAB has built-in authentication support and Role-Based Access Control (RBAC), which provides configurable roles and permissions for individual users.

To turn on this feature, in your airflow.cfg file (under [webserver]), set the configuration variable rbac = True, and then run airflow command, which will generate the file in your $AIRFLOW_HOME.

Setting up Authentication

FAB has built-in authentication support for DB, OAuth, OpenID, LDAP, and REMOTE_USER. The default auth type is AUTH_DB.

For any other authentication type (OAuth, OpenID, LDAP, REMOTE_USER), see the Authentication section of FAB docs for how to configure variables in file.

Once you modify your config file, run airflow db init to generate new tables for RBAC support (these tables will have the prefix ab_).

Creating an Admin Account

Once configuration settings have been updated and new tables have been generated, create an admin account with airflow create_user command.

Using your new UI

Run airflow webserver to start the new UI. This will bring up a log in page, enter the recently created admin username and password.

There are five roles created for Airflow by default: Admin, User, Op, Viewer, and Public. To configure roles/permissions, go to the Security tab and click List Roles in the new UI.

Breaking changes

  • AWS Batch Operator renamed property queue to job_queue to prevent conflict with the internal queue from CeleryExecutor - AIRFLOW-2542
  • Users created and stored in the old users table will not be migrated automatically. FAB's built-in authentication support must be reconfigured.
  • Airflow dag home page is now /home (instead of /admin).
  • All ModelViews in Flask-AppBuilder follow a different pattern from Flask-Admin. The /admin part of the URL path will no longer exist. For example: /admin/connection becomes /connection/list, /admin/connection/new becomes /connection/add, /admin/connection/edit becomes /connection/edit, etc.
  • Due to security concerns, the new webserver will no longer support the features in the Data Profiling menu of old UI, including Ad Hoc Query, Charts, and Known Events.
  • HiveServer2Hook.get_results() always returns a list of tuples, even when a single column is queried, as per Python API 2.
  • UTC is now the default timezone: Either reconfigure your workflows scheduling in UTC or set default_timezone as explained in

airflow.contrib.sensors.hdfs_sensors renamed to airflow.contrib.sensors.hdfs_sensor

We now rename airflow.contrib.sensors.hdfs_sensors to airflow.contrib.sensors.hdfs_sensor for consistency purpose.

MySQL setting required

We now rely on more strict ANSI SQL settings for MySQL in order to have sane defaults. Make sure to have specified explicit_defaults_for_timestamp=1 in your my.cnf under [mysqld]

Celery config

To make the config of Airflow compatible with Celery, some properties have been renamed:

celeryd_concurrency -> worker_concurrency
celery_result_backend -> result_backend
celery_ssl_active -> ssl_active
celery_ssl_cert -> ssl_cert
celery_ssl_key -> ssl_key

Resulting in the same config parameters as Celery 4, with more transparency.

GCP Dataflow Operators

Dataflow job labeling is now supported in Dataflow{Java,Python}Operator with a default "airflow-version" label, please upgrade your google-cloud-dataflow or apache-beam version to 2.2.0 or greater.

BigQuery Hooks and Operator

The bql parameter passed to BigQueryOperator and BigQueryBaseCursor.run_query has been deprecated and renamed to sql for consistency purposes. Using bql will still work (and raise a DeprecationWarning), but is no longer supported and will be removed entirely in Airflow 2.0

Redshift to S3 Operator

With Airflow 1.9 or lower, Unload operation always included header row. In order to include header row, we need to turn off parallel unload. It is preferred to perform unload operation using all nodes so that it is faster for larger tables. So, parameter called include_header is added and default is set to False. Header row will be added only if this parameter is set True and also in that case parallel will be automatically turned off (PARALLEL OFF)

Google cloud connection string

With Airflow 1.9 or lower, there were two connection strings for the Google Cloud operators, both google_cloud_storage_default and google_cloud_default. This can be confusing and therefore the google_cloud_storage_default connection id has been replaced with google_cloud_default to make the connection id consistent across Airflow.

Logging Configuration

With Airflow 1.9 or lower, FILENAME_TEMPLATE, PROCESSOR_FILENAME_TEMPLATE, LOG_ID_TEMPLATE, END_OF_LOG_MARK were configured in These have been moved into the configuration file, and hence if you were using a custom configuration file the following defaults need to be added.

fab_logging_level = WARN
log_filename_template = {{{{ ti.dag_id }}}}/{{{{ ti.task_id }}}}/{{{{ ts }}}}/{{{{ try_number }}}}.log
log_processor_filename_template = {{{{ filename }}}}.log

elasticsearch_log_id_template = {{dag_id}}-{{task_id}}-{{execution_date}}-{{try_number}}
elasticsearch_end_of_log_mark = end_of_log

The previous setting of log_task_reader is not needed in many cases now when using the default logging config with remote storages. (Previously it needed to be set to s3.task or similar. This is not needed with the default config anymore)

Change of per-task log path

With the change to Airflow core to be timezone aware the default log path for task instances will now include timezone information. This will by default mean all previous task logs won't be found. You can get the old behaviour back by setting the following config options:

log_filename_template = {{ ti.dag_id }}/{{ ti.task_id }}/{{ execution_date.strftime("%%Y-%%m-%%dT%%H:%%M:%%S") }}/{{ try_number }}.log

Airflow 1.9

SSH Hook updates, along with new SSH Operator & SFTP Operator

SSH Hook now uses the Paramiko library to create an ssh client connection, instead of the sub-process based ssh command execution previously (<1.9.0), so this is backward incompatible.

  • update SSHHook constructor
  • use SSHOperator class in place of SSHExecuteOperator which is removed now. Refer to for usage info.
  • SFTPOperator is added to perform secure file transfer from serverA to serverB. Refer to for usage info.
  • No updates are required if you are using ftpHook, it will continue to work as is.

S3Hook switched to use Boto3

The airflow.hooks.S3_hook.S3Hook has been switched to use boto3 instead of the older boto (a.k.a. boto2). This results in a few backwards incompatible changes to the following classes: S3Hook:

  • the constructors no longer accepts s3_conn_id. It is now called aws_conn_id.
  • the default connection is now "aws_default" instead of "s3_default"
  • the return type of objects returned by get_bucket is now boto3.s3.Bucket
  • the return type of get_key, and get_wildcard_key is now an boto3.S3.Object.

If you are using any of these in your DAGs and specify a connection ID you will need to update the parameter name for the connection to "aws_conn_id": S3ToHiveTransfer, S3PrefixSensor, S3KeySensor, RedshiftToS3Transfer.

Logging update

The logging structure of Airflow has been rewritten to make configuration easier and the logging system more transparent.

A quick recap about logging

A logger is the entry point into the logging system. Each logger is a named bucket to which messages can be written for processing. A logger is configured to have a log level. This log level describes the severity of the messages that the logger will handle. Python defines the following log levels: DEBUG, INFO, WARNING, ERROR or CRITICAL.

Each message that is written to the logger is a Log Record. Each log record contains a log level indicating the severity of that specific message. A log record can also contain useful metadata that describes the event that is being logged. This can include details such as a stack trace or an error code.

When a message is given to the logger, the log level of the message is compared to the log level of the logger. If the log level of the message meets or exceeds the log level of the logger itself, the message will undergo further processing. If it doesn’t, the message will be ignored.

Once a logger has determined that a message needs to be processed, it is passed to a Handler. This configuration is now more flexible and can be easily be maintained in a single file.

Changes in Airflow Logging

Airflow's logging mechanism has been refactored to use Python’s built-in logging module to perform logging of the application. By extending classes with the existing LoggingMixin, all the logging will go through a central logger. Also the BaseHook and BaseOperator already extend this class, so it is easily available to do logging.

The main benefit is easier configuration of the logging by setting a single centralized python file. Disclaimer; there is still some inline configuration, but this will be removed eventually. The new logging class is defined by setting the dotted classpath in your ~/airflow/airflow.cfg file:

# Logging class
# Specify the class that will specify the logging configuration
# This class has to be on the python classpath
logging_config_class = my.path.default_local_settings.LOGGING_CONFIG

The logging configuration file needs to be on the PYTHONPATH, for example $AIRFLOW_HOME/config. This directory is loaded by default. Any directory may be added to the PYTHONPATH, this might be handy when the config is in another directory or a volume is mounted in case of Docker.

The config can be taken from airflow/config_templates/ as a starting point. Copy the contents to ${AIRFLOW_HOME}/config/, and alter the config as is preferred.

# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License.  You may obtain a copy of the License at
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# KIND, either express or implied.  See the License for the
# specific language governing permissions and limitations
# under the License.

import os

from airflow import configuration as conf

# TODO: Logging format and level should be configured
# in this file instead of from airflow.cfg. Currently
# there are other log format and level configurations in
# and Please see AIRFLOW-1455.

LOG_LEVEL = conf.get('core', 'LOGGING_LEVEL').upper()
LOG_FORMAT = conf.get('core', 'log_format')

BASE_LOG_FOLDER = conf.get('core', 'BASE_LOG_FOLDER')
PROCESSOR_LOG_FOLDER = conf.get('scheduler', 'child_process_log_directory')

FILENAME_TEMPLATE = '{{ ti.dag_id }}/{{ ti.task_id }}/{{ ts }}/{{ try_number }}.log'
PROCESSOR_FILENAME_TEMPLATE = '{{ filename }}.log'

    'version': 1,
    'disable_existing_loggers': False,
    'formatters': {
        'airflow.task': {
            'format': LOG_FORMAT,
        'airflow.processor': {
            'format': LOG_FORMAT,
    'handlers': {
        'console': {
            'class': 'logging.StreamHandler',
            'formatter': 'airflow.task',
            'stream': 'ext://sys.stdout'
        'file.task': {
            'class': 'airflow.utils.log.file_task_handler.FileTaskHandler',
            'formatter': 'airflow.task',
            'base_log_folder': os.path.expanduser(BASE_LOG_FOLDER),
            'filename_template': FILENAME_TEMPLATE,
        'file.processor': {
            'class': 'airflow.utils.log.file_processor_handler.FileProcessorHandler',
            'formatter': 'airflow.processor',
            'base_log_folder': os.path.expanduser(PROCESSOR_LOG_FOLDER),
            'filename_template': PROCESSOR_FILENAME_TEMPLATE,
        # When using s3 or gcs, provide a customized LOGGING_CONFIG
        # in airflow_local_settings within your PYTHONPATH, see
        # for details
        # 's3.task': {
        #     'class': 'airflow.utils.log.s3_task_handler.S3TaskHandler',
        #     'formatter': 'airflow.task',
        #     'base_log_folder': os.path.expanduser(BASE_LOG_FOLDER),
        #     's3_log_folder': S3_LOG_FOLDER,
        #     'filename_template': FILENAME_TEMPLATE,
        # },
        # 'gcs.task': {
        #     'class': 'airflow.utils.log.gcs_task_handler.GCSTaskHandler',
        #     'formatter': 'airflow.task',
        #     'base_log_folder': os.path.expanduser(BASE_LOG_FOLDER),
        #     'gcs_log_folder': GCS_LOG_FOLDER,
        #     'filename_template': FILENAME_TEMPLATE,
        # },
    'loggers': {
        '': {
            'handlers': ['console'],
            'level': LOG_LEVEL
        'airflow': {
            'handlers': ['console'],
            'level': LOG_LEVEL,
            'propagate': False,
        'airflow.processor': {
            'handlers': ['file.processor'],
            'level': LOG_LEVEL,
            'propagate': True,
        'airflow.task': {
            'handlers': ['file.task'],
            'level': LOG_LEVEL,
            'propagate': False,
        'airflow.task_runner': {
            'handlers': ['file.task'],
            'level': LOG_LEVEL,
            'propagate': True,

To customize the logging (for example, use logging rotate), define one or more of the logging handles that Python has to offer. For more details about the Python logging, please refer to the official logging documentation.

Furthermore, this change also simplifies logging within the DAG itself:

root@ae1bc863e815:/airflow# python
Python 3.6.2 (default, Sep 13 2017, 14:26:54)
[GCC 4.9.2] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from airflow.settings import *
>>> from datetime import datetime
>>> from airflow.models.dag import DAG
>>> from airflow.operators.dummy import DummyOperator
>>> dag = DAG('simple_dag', start_date=datetime(2017, 9, 1))
>>> task = DummyOperator(task_id='task_1', dag=dag)
>>> task.log.error('I want to say something..')
[2017-09-25 20:17:04,927] {<stdin>:1} ERROR - I want to say something..

Template path of the file_task_handler

The file_task_handler logger has been made more flexible. The default format can be changed, {dag_id}/{task_id}/{execution_date}/{try_number}.log by supplying Jinja templating in the FILENAME_TEMPLATE configuration variable. See the file_task_handler for more information.

I'm using S3Log or GCSLogs, what do I do!?

If you are logging to Google cloud storage, please see the Google cloud platform documentation for logging instructions.

If you are using S3, the instructions should be largely the same as the Google cloud platform instructions above. You will need a custom logging config. The REMOTE_BASE_LOG_FOLDER configuration key in your airflow config has been removed, therefore you will need to take the following steps:

  • Copy the logging configuration from airflow/config_templates/
  • Place it in a directory inside the Python import path PYTHONPATH. If you are using Python 2.7, ensuring that any files exist so that it is importable.
  • Update the config by setting the path of REMOTE_BASE_LOG_FOLDER explicitly in the config. The REMOTE_BASE_LOG_FOLDER key is not used anymore.
  • Set the logging_config_class to the filename and dict. For example, if you place on the base of your PYTHONPATH, you will need to set logging_config_class = custom_logging_config.LOGGING_CONFIG in your config as Airflow 1.8.

New Features

Dask Executor

A new DaskExecutor allows Airflow tasks to be run in Dask Distributed clusters.

Deprecated Features

These features are marked for deprecation. They may still work (and raise a DeprecationWarning), but are no longer supported and will be removed entirely in Airflow 2.0

  • If you're using the google_cloud_conn_id or dataproc_cluster argument names explicitly in contrib.operators.Dataproc{*}Operator(s), be sure to rename them to gcp_conn_id or cluster_name, respectively. We've renamed these arguments for consistency. (AIRFLOW-1323)

  • post_execute() hooks now take two arguments, context and result (AIRFLOW-886)

    Previously, post_execute() only took one argument, context.

  • contrib.hooks.gcp_dataflow_hook.DataFlowHook starts to use --runner=DataflowRunner instead of DataflowPipelineRunner, which is removed from the package google-cloud-dataflow-0.6.0.

  • The pickle type for XCom messages has been replaced by json to prevent RCE attacks. Note that JSON serialization is stricter than pickling, so if you want to e.g. pass raw bytes through XCom you must encode them using an encoding like base64. By default pickling is still enabled until Airflow 2.0. To disable it set enable_xcom_pickling = False in your Airflow config.

Airflow 1.8.1

The Airflow package name was changed from airflow to apache-airflow during this release. You must uninstall a previously installed version of Airflow before installing 1.8.1.

Airflow 1.8


The database schema needs to be upgraded. Make sure to shutdown Airflow and make a backup of your database. To upgrade the schema issue airflow upgradedb.

Upgrade systemd unit files

Systemd unit files have been updated. If you use systemd please make sure to update these.

Please note that the webserver does not detach properly, this will be fixed in a future version.

Tasks not starting although dependencies are met due to stricter pool checking

Airflow 1.7.1 has issues with being able to over subscribe to a pool, ie. more slots could be used than were available. This is fixed in Airflow 1.8.0, but due to past issue jobs may fail to start although their dependencies are met after an upgrade. To workaround either temporarily increase the amount of slots above the amount of queued tasks or use a new pool.

Less forgiving scheduler on dynamic start_date

Using a dynamic start_date (e.g. start_date = is not considered a best practice. The 1.8.0 scheduler is less forgiving in this area. If you encounter DAGs not being scheduled you can try using a fixed start_date and renaming your DAG. The last step is required to make sure you start with a clean slate, otherwise the old schedule can interfere.

New and updated scheduler options

Please read through the new scheduler options, defaults have changed since 1.7.1.


In order to increase the robustness of the scheduler, DAGS are now processed in their own process. Therefore each DAG has its own log file for the scheduler. These log files are placed in child_process_log_directory which defaults to <AIRFLOW_HOME>/scheduler/latest. You will need to make sure these log files are removed.

DAG logs or processor logs ignore and command line settings for log file locations.


Previously the command line option num_runs was used to let the scheduler terminate after a certain amount of loops. This is now time bound and defaults to -1, which means run continuously. See also num_runs.


Previously num_runs was used to let the scheduler terminate after a certain amount of loops. Now num_runs specifies the number of times to try to schedule each DAG file within run_duration time. Defaults to -1, which means try indefinitely. This is only available on the command line.


After how much time should an updated DAG be picked up from the filesystem.


CURRENTLY DISABLED DUE TO A BUG How many seconds to wait between file-parsing loops to prevent the logs from being spammed.


The frequency with which the scheduler should relist the contents of the DAG directory. If while developing +dags, they are not being picked up, have a look at this number and decrease it when necessary.


By default the scheduler will fill any missing interval DAG Runs between the last execution date and the current date. This setting changes that behavior to only execute the latest interval. This can also be specified per DAG as catchup = False / True. Command line backfills will still work.

Faulty DAGs do not show an error in the Web UI

Due to changes in the way Airflow processes DAGs the Web UI does not show an error when processing a faulty DAG. To find processing errors go the child_process_log_directory which defaults to <AIRFLOW_HOME>/scheduler/latest.

New DAGs are paused by default

Previously, new DAGs would be scheduled immediately. To retain the old behavior, add this to airflow.cfg:

dags_are_paused_at_creation = False

Airflow Context variable are passed to Hive config if conf is specified

If you specify a hive conf to the run_cli command of the HiveHook, Airflow add some convenience variables to the config. In case you run a secure Hadoop setup it might be required to allow these variables by adjusting you hive configuration to add airflow\.ctx\..* to the regex of user-editable configuration properties. See the Hive docs on Configuration Properties for more info.

Google Cloud Operator and Hook alignment

All Google Cloud Operators and Hooks are aligned and use the same client library. Now you have a single connection type for all kinds of Google Cloud Operators.

If you experience problems connecting with your operator make sure you set the connection type "Google Cloud".

Also the old P12 key file type is not supported anymore and only the new JSON key files are supported as a service account.

Deprecated Features

These features are marked for deprecation. They may still work (and raise a DeprecationWarning), but are no longer supported and will be removed entirely in Airflow 2.0

  • Hooks and operators must be imported from their respective submodules

    airflow.operators.PigOperator is no longer supported; from airflow.operators.pig_operator import PigOperator is. (AIRFLOW-31, AIRFLOW-200)

  • Operators no longer accept arbitrary arguments

    Previously, Operator.__init__() accepted any arguments (either positional *args or keyword **kwargs) without complaint. Now, invalid arguments will be rejected. (

  • The config value secure_mode will default to True which will disable some insecure endpoints/features

Known Issues

There is a report that the default of "-1" for num_runs creates an issue where errors are reported while parsing tasks. It was not confirmed, but a workaround was found by changing the default back to None.

To do this edit, find the following:

        'num_runs': Arg(
            ("-n", "--num_runs"),
            default=-1, type=int,
            help="Set the number of runs to execute before exiting"),

and change default=-1 to default=None. If you have this issue please report it on the mailing list.


Changes to Configuration

Email configuration change

To continue using the default smtp email backend, change the email_backend line in your config file from:

email_backend = airflow.utils.send_email_smtp


email_backend =

S3 configuration change

To continue using S3 logging, update your config file so:

s3_log_folder = s3://my-airflow-log-bucket/logs


remote_base_log_folder = s3://my-airflow-log-bucket/logs
remote_log_conn_id = <your desired s3 connection>