NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI, #1563 most downloaded on PyPI
Programmatically author, schedule and monitor data pipelines
Latest stable
3.4.0
Last release 12 days ago
06 Nov 2026
Ships fairly regularly
a new release about every 2 weeks
Nearly every release is documented
notes for 60 of the last 60 stable releases
4 versions withdrawn
withdrawn after publishing
10 years old
306 releases · first in 2017
Drag across the ruler to compare, or tap a release to open it.
Nothing published for this version
Add backwards compatibility shim and deprecation warning for EmailOperator
📦 PyPI: https://pypi.org/project/apache-airflow/3.0.2/ 📚 Docs: https://airflow.apache.org/docs/apache-airflow/3.0.2/ 🛠️ Release Notes: https://airflow.apache.org/docs/apache-airflow/3.0.2/release_notes.html 🪶 Sources: https://airflow.apache.org/docs/apache-airflow/3.0.2/installation/installing-from-sources.html
sys.path in task runner (#51318)sys.path in dag processor (#50385)SIGSEGV signals during DAG file imports (#51171)dag.test() (#51182)ForwardRef error by reordering discriminated union definitions (#50688)BaseNotifier (#50340)upstream_mapped_index when xcom access is needed (#50641)/run API endpoint for older Task SDK clientsFlexibleForm component (#50845)+1 more when tags exceed the display limit by one (#50669)default_args handling in operator .partial() to prevent TypeError when unused keys are present (#50525)airflow tasks clear command (#49631)--local flag in dag list and dag list-import-errors CLI commands (#49380)DagProcessor stats log to show the correct parse duration (#50316)get_log API (#50547)logical_date check when validating inlets and outlets (#51464)ti update state and set task to fail if exception encountered (#51295)example_dags in standard provider to example_dags in sources (#51275)airflow-core package (#51192)task.test to Task SDK (#50827)dag.test to Task SDK (#50300,#50419)ti.run to Task SDK execution path (#50141)airflow dags test from local files (#50420)execution_time module (#50940)dagrun value for list display (#50834)secret_key config to api section (#50839)webserver configs to fab provider (#50774,#50269,#50208,#50896)dag_run nullable in Details page (#50719)ab_user table (#50343)owner_links field to DAGDetailsResponse for enhanced owner metadata in the API (#50557)AssetAlias for alias in Asset Metadata example (#50768)schedule_interval in tutorial dags (#50947)PythonOperator in tutorial dag (#50962)Full Changelog: https://github.com/apache/airflow/compare/3.0.1...3.0.2
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Nothing published for this version
Nothing published for this version
Fix a few SqlAlchemy deprecation warnings
📦 PyPI: https://pypi.org/project/apache-airflow/3.0.1/ 📚 Docs: https://airflow.apache.org/docs/apache-airflow/3.0.1/index.html 🛠️ Release Notes: https://airflow.apache.org/docs/apache-airflow/3.0.1/release_notes.html 🪶 Sources: https://airflow.apache.org/docs/apache-airflow/3.0.1/installation/installing-from-sources.html
dag_code records with no serialized dag (#49478)dag_code and serialized_dag tables on 3.0 upgrade (#49563)scheduler_interval field on downgrade (#49583)base_url in api server (#49545)max_active_tis_per_dag is not respected by dynamically mapped tasks (#49708)SimpleAuthManager (#49697)(#49866)bundle_version to DagRun response (#49726)task_ids in xcom_pull the same as multiple when provided as part of a list (#49692)DAGModel stale and associate bundle on import errors to aid migration from 2.10.5 (#49769)pip with avoiding resolution too deep issues in Python 3.12 (#49853)BashSensor (#49935)map_index_template on task completion (#49809)ContinuousTimetable false triggering when last run ends in future (#45175)Stats (#50088)TaskGroup (#49996)mapIndex to clear the relevant task instances. (#50256)STRAIGHT_JOIN prefix for MySQL query optimization in get_sorted_triggers (#46303)sqlalchemy[asyncio] extra is in core deps (#49452)HANDLER_SUPPORTS_TRIGGERER (#49370)gitpython as a core dependency (#49537)@babel/runtime from 7.26.0 to 7.27.0 (#49479)get_current_context (#49630)RunBackfillForm (#49609)backfill_id (#49691)(#49716)SimpleAllAdminMiddleware to allow api usage without auth header in request (#49599)react-router and react-router-dom from 7.4.0 to 7.5.2 (#49742)root_dag_id in dagbag and restore logic (#49668)airflow.cfg files across all containers in default docker-compose.yaml (#49681)/airflow-core (#49512)apache-airflow meta package (#49846)@task.kuberenetes_cmd (#46913)vite from 5.4.17 to 5.4.19 for Airflow UI (#49162)(#50074)map_index filter option to GetTICount and GetTaskStates (#49818)stats ui endpoint (#49985)state attribute to RuntimeTaskInstance for easier ti.state access in Task Context (#50031)dag_run_conf to RunBackfillForm (#49763)dateInterval validation and error handling (#50072)Task Instances [{map_index}] tab to mapped task details (#50085)security/api.rst (#49675)max_consecutive_failed_dag_runs default value to zero in TaskSDK dag (#49795) (#49803)example_params_ui_tutorial) (#49905)Full Changelog: https://github.com/apache/airflow/compare/3.0.0...3.0.1
Nothing published for this version
📣 We are proud to announce the General Availability of Apache Airflow® 3.0, the most significant release in the project’s history.
📣 We are proud to announce the General Availability of Apache Airflow® 3.0, the most significant release in the project’s history.
Airflow 3.0 builds on the foundation of Airflow 2 and introduces a new service-oriented architecture, a modern React-based UI, enhanced security, and a host of long-requested features such as DAG versioning, improved backfills, event-driven scheduling, and support for remote execution.
You can read more about what 3.0 brings in https://airflow.apache.org/blog/airflow-three-point-oh-is-here/.
📦 PyPI: https://pypi.org/project/apache-airflow/3.0.0/ 📚 Docs: https://airflow.apache.org/docs/apache-airflow/3.0.0 🛠️ Release Notes: https://airflow.apache.org/docs/apache-airflow/3.0.0/release_notes.html 🪶 Sources: https://airflow.apache.org/docs/apache-airflow/3.0.0/installation/installing-from-sources.html
This is the result of 300+ developers within the Airflow community working together tirelessly for many months! A huge thank you to all of them for their contributions.
Resources
We are proud to announce the General Availability of Apache Airflow 3.0 — the most significant release in the project's history. This version introduces a service-oriented architecture, a stable DAG authoring interface, expanded support for event-driven and ML workflows, and a fully modernized UI built on React. Airflow 3.0 reflects years of community investment and lays the foundation for the next era of scalable, modular orchestration.
Service-Oriented Architecture: A new Task Execution API and airflow api-server enable task execution in remote environments with improved isolation and flexibility (AIP-72).
Edge Executor: A new executor that supports distributed, event-driven, and edge-compute workflows (AIP-69), now generally available.
Stable Authoring Interface: DAG authors should now use the new airflow.sdk namespace to import core DAG constructs like @dag, @task, and DAG.
Scheduler-Managed Backfills: Backfills are now scheduled and tracked like regular DAG runs, with native UI and API support (AIP-78).
DAG Versioning: Airflow now tracks structural changes to DAGs over time, enabling inspection of historical DAG definitions via the UI and API (AIP-66).
Asset-Based Scheduling: The dataset model has been renamed and redesigned as assets, with a new @asset decorator and cleaner event-driven DAG definition (AIP-74, AIP-75).
Support for ML and AI Workflows: DAGs can now run with logical_date=None, enabling use cases such as model inference, hyperparameter tuning, and non-interval workflows (AIP-83).
Removal of Legacy Features: SLAs, SubDAGs, DAG and Xcom pickling, and several internal context variables have been removed. Use the upgrade tools to detect deprecated usage.
Split CLI and API Changes: The CLI has been split into airflow and airflowctl (AIP-81), and REST API now defaults to logical_date=None when triggering a new DAG run.
Modern React UI: A complete UI overhaul built on React and FastAPI includes version-aware views, backfill management, and improved DAG and task introspection (AIP-38, AIP-84).
Migration Tooling: Use ruff and airflow config update to validate DAGs and configurations. Upgrade requires Airflow 2.7 or later and Python 3.9–3.12.
Airflow 3.0 introduces the most significant set of changes since the 2.0 release, including architectural shifts, new execution models, and improvements to DAG authoring and scheduling.
Airflow now supports a service-oriented architecture, enabling tasks to be executed remotely via a new Task Execution API. This API decouples task execution from the scheduler and introduces a stable contract for running tasks outside of Airflow's traditional runtime environment.
To support this, Airflow introduces the Task SDK — a lightweight runtime environment for running Airflow tasks in external systems such as containers, edge environments, or other runtimes. This lays the groundwork for language-agnostic task execution and brings improved isolation, portability, and extensibility to Airflow-based workflows.
Airflow 3.0 also introduces a new airflow.sdk namespace that exposes the core authoring interfaces for defining DAGs and tasks. DAG authors should now import objects like DAG, @dag, and @task from airflow.sdk rather than internal modules. This new namespace provides a stable, forward-compatible interface for DAG authoring across future versions of Airflow.
Airflow 3.0 introduces the Edge Executor as a generally available feature, enabling execution of tasks in distributed or remote compute environments. Designed for event-driven and edge-compute use cases, the Edge Executor integrates with the Task Execution API to support task orchestration beyond the traditional Airflow runtime. This advancement facilitates hybrid and cross-environment orchestration patterns, allowing task workers to operate closer to data or application layers.
Backfills are now fully managed by the scheduler, rather than being launched as separate command-line jobs. This change unifies backfill logic with regular DAG execution and ensures that backfill runs follow the same scheduling, versioning, and observability models as other DAG runs.
Airflow 3.0 also introduces native UI and REST API support for initiating and monitoring backfills, making them more accessible and easier to integrate into automated workflows. These improvements lay the foundation for smarter, safer historical reprocessing — now available directly through the Airflow UI and API.
Airflow 3.0 introduces native DAG versioning. DAG structure changes (e.g., renamed tasks, dependency shifts) are now tracked directly in the metadata database. This allows users to inspect historical DAG structures through the UI and API, and lays the foundation for safer backfills, improved observability, and runtime-determined DAG logic.
Note: DAG bundles are not initialized in the triggerer. In practice, this means that triggers cannot come from a DAG bundle. This is because the triggerer does not deal with changes in trigger code over time, as everything happens in the main process. Triggers can come from anywhere else on sys.path instead.
Airflow 3.0 ships with a completely redesigned user interface built on React and FastAPI. This modern architecture improves responsiveness, enables more consistent navigation across views, and unlocks new UI capabilities — including support for DAG versioning, asset-centric DAG definitions, and more intuitive filtering and search.
The new UI replaces the legacy Flask-based frontend and introduces a foundation for future extensibility and community contributions.
The concept of Datasets has been renamed to Assets, unifying terminology with common practices in the modern data ecosystem. The internal model has also been reworked to better support future features like asset partitions and validations.
The @asset decorator and related changes to the DAG parser enable clearer, asset-centric DAG definitions, allowing Airflow to more naturally support event-driven and data-aware scheduling patterns.
This renaming impacts modules, classes, functions, configuration keys, and internal models. Key changes include:
Dataset → Asset
DatasetEvent → AssetEvent
DatasetAlias → AssetAlias
airflow.datasets.* → airflow.sdk.*
airflow.timetables.simple.DatasetTriggeredTimetable → airflow.timetables.simple.AssetTriggeredTimetable
airflow.timetables.datasets.DatasetOrTimeSchedule → airflow.timetables.assets.AssetOrTimeSchedule
airflow.listeners.spec.dataset.on_dataset_created → airflow.listeners.spec.asset.on_asset_created
airflow.listeners.spec.dataset.on_dataset_changed → airflow.listeners.spec.asset.on_asset_changed
core.dataset_manager_class → core.asset_manager_class
core.dataset_manager_kwargs → core.asset_manager_kwargs
Airflow 3.0 removes the legacy schedule_interval and timetable parameters. DAGs must now use the unified schedule field for all time- and event-based scheduling logic. This simplifies DAG definition and improves consistency across scheduling paradigms.
Airflow 3.0 changes the default behavior for new DAGs by setting catchup_by_default = False in the configuration file. This means DAGs that do not explicitly set catchup=... will no longer backfill missed intervals by default. This change reduces confusion for new users and better reflects the growing use of on-demand and event-driven workflows.
The default DAG schedule has been changed to None from @once.
Task code can no longer directly access the metadata database. Interactions with DAG state, task history, or DAG runs must be performed via the Airflow REST API or exposed context. This change improves architectural separation and enables remote execution.
Airflow no longer supports triggering DAG runs with a logical date in the future. This change aligns with the logical execution model and removes ambiguity in backfills and event-driven DAGs. Use logical_date=None to trigger runs with the current timestamp.
For DAG runs triggered by an Asset event or through the REST API without specifying a logical_date, Airflow now sets logical_date=None by default. These DAG runs do not have a data interval, and attempting to access data_interval_start, data_interval_end, or logical_date from the task context will raise a KeyError.
DAG authors should use dag_run.logical_date and perform appropriate checks or fallbacks if supporting multiple trigger types. This change improves consistency with event-driven semantics but may require updates to existing DAGs that assume these values are always present.
Airflow 3.0 refines task callback behavior to improve clarity and consistency. In particular, on_success_callback is no longer executed when a task is marked as SKIPPED, aligning it more closely with expected semantics.
Several default configuration values have been updated in Airflow 3.0 to better reflect modern usage patterns and simplify onboarding:
catchup_by_default is now set to False by default. DAGs will not automatically backfill unless explicitly configured to do so.
create_cron_data_intervals is now set to False by default. As a result, cron expressions will be interpreted using the CronTriggerTimetable instead of the legacy CronDataIntervalTimetable. This only affects DAGs that pass a bare cron string to schedule=; DAGs that pass an explicit timetable instance are unaffected. If you rely on the data interval semantics (data_interval_start / data_interval_end, or templated values like ds / ts derived from logical_date), set create_cron_data_intervals=True explicitly before the upgrade. Flipping the value later, after Airflow 3 DAG runs already exist, will skip one scheduled run on each affected DAG to avoid colliding with the previous run's logical_date.
SimpleAuthManager is now the default auth_manager. To continue using Flask AppBuilder-based authentication, install the apache-airflow-providers-fab provider and explicitly set auth_manager = airflow.providers.fab.auth_manager.FabAuthManager.
These changes represent the most significant evolution of the Airflow platform since the release of 2.0 — setting the stage for more scalable, event-driven, and language-agnostic orchestration in the years ahead.
Airflow 3.0 introduces several important improvements and behavior changes in how DAGs and tasks are scheduled, prioritized, and executed.
Airflow 3.0 now requires the standalone DAG processor to parse DAGs. This dedicated process improves scheduler performance, isolation, and observability. It also simplifies architecture by clearly separating DAG parsing from scheduling logic. This change may affect custom deployments that previously used embedded DAG parsing.
The priority_weight value on a task is now capped by the number of available pool slots. This ensures that resource availability remains the primary constraint in task execution order, preventing high-priority tasks from starving others when resource contention exists.
Teardown tasks will now be executed even when a DAG run is terminated early. This ensures that cleanup logic is respected, improving reliability for workflows that use teardown tasks to manage ephemeral infrastructure, temporary files, or downstream notifications.
Scheduler components now use run_with_db_retries to handle transient database issues more gracefully. This enhances Airflow's fault tolerance in high-volume environments and reduces the likelihood of scheduler restarts due to temporary database connection problems.
Airflow 3.0 fixes a bug that caused incorrect task statistics to be reported for dynamic task mapping. Stats now accurately reflect the number of mapped task instances and their statuses, improving observability and debugging for dynamic workflows.
SequentialExecutor was primarily used for local testing but is now redundant, as LocalExecutor supports SQLite with WAL mode and provides better performance with parallel execution. Users should switch to LocalExecutor or CeleryExecutor as alternatives.
Airflow 3.0 includes several changes that improve consistency, clarity, and long-term stability for DAG authors.
Airflow 3.0 introduces a new, stable public API for DAG authoring under the airflow.sdk namespace, available via the apache-airflow-task-sdk package.
The goal of this change is to decouple DAG authoring from Airflow internals (Scheduler, API Server, etc.), providing a forward-compatible, stable interface for writing and maintaining DAGs across Airflow versions.
DAG authors should now import core constructs from airflow.sdk rather than internal modules.
Key Imports from airflow.sdk:
Classes:
Asset
BaseNotifier
BaseOperator
BaseOperatorLink
BaseSensorOperator
Connection
Context
DAG
EdgeModifier
Label
ObjectStoragePath
Param
TaskGroup
Variable
Decorators and Functions:
@asset
@dag
@setup
@task
@task_group
@teardown
chain
chain_linear
cross_downstream
get_current_context
get_parsing_context
For an exhaustive list of available classes, decorators, and functions, check airflow.sdk.__all__.
All DAGs should update imports to use airflow.sdk instead of referencing internal Airflow modules directly. Legacy import paths (e.g., airflow.models.dag.DAG, airflow.decorator.task) are deprecated and will be removed in a future Airflow version. Some additional utilities and helper functions that DAGs sometimes use from airflow.utils.* and others will be progressively migrated to the Task SDK in future minor releases.
These future changes aim to complete the decoupling of DAG authoring constructs from internal Airflow services. DAG authors should expect continued improvements to airflow.sdk with no backwards-incompatible changes to existing constructs.
For example, update:
# Old (Airflow 2.x)
from airflow.models import DAG
from airflow.decorators import task
# New (Airflow 3.x)
from airflow.sdk import DAG, task
The DAG argument fail_stop has been renamed to fail_fast for improved clarity. This parameter controls whether a DAG run should immediately stop execution when a task fails. DAG authors should update any code referencing fail_stop to use the new name.
Several legacy context variables have been removed or may no longer be available in certain types of DAG runs, including:
conf
execution_date
dag_run.external_trigger
In asset-triggered and manually triggered DAG runs with logical_date=None, data interval fields such as data_interval_start and data_interval_end may not be present in the task context. DAG authors should use explicit references such as dag_run.logical_date and conditionally check for the presence of interval-related fields where applicable.
Internal task context functions such as get_parsing_context have been moved to a more appropriate location (e.g., airflow.models.taskcontext). DAG authors using these utilities directly should update import paths accordingly.
The TriggerRule.ALWAYS rule can no longer be used with teardown tasks or tasks that are expected to honor upstream dependency semantics. DAG authors should ensure that teardown logic is defined with the appropriate trigger rules for consistent task resolution behavior.
A new utility function, create_asset_aliases(), allows DAG authors to define reusable aliases for frequently referenced Assets. This improves modularity and reuse across DAG files and is particularly helpful for teams adopting asset-centric DAGs.
The Operator Extra links, which can be defined either via plugins or custom operators now do not execute any user code in the Airflow UI, but instead push the "full" links to XCom backend and the link is fetched from the XCom backend when viewing task details, for example from grid view.
Example for users with custom links class:
@attr.s(auto_attribs=True)
class CustomBaseIndexOpLink(BaseOperatorLink):
"""Custom Operator Link for Google BigQuery Console."""
index: int = attr.ib()
@property
def name(self) -> str:
return f"BigQuery Console #{self.index + 1}"
@property
def xcom_key(self) -> str:
return f"bigquery_{self.index + 1}"
def get_link(self, operator, *, ti_key):
search_queries = XCom.get_one(
task_id=ti_key.task_id, dag_id=ti_key.dag_id, run_id=ti_key.run_id, key="search_query"
)
return f"https://console.cloud.google.com/bigquery?j={search_query}"
The link has an xcom_key defined, which is how it will be stored in the XCOM backend, with key as xcom_key and value as the entire link, this case: https://console.cloud.google.com/bigquery?j=search
Operator (including Sensors), Executors & Hooks can 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. They can be imported directly from their respective provider packages.
Before:
from airflow.hooks.my_plugin import MyHook
You should instead import it as:
from my_plugin import MyHook
Airflow 3.0 expands the types of DAGs that can be expressed by removing the constraint that each DAG run must correspond to a unique data interval. This change, introduced in AIP-83, enables support for workflows that don't operate on a fixed schedule — such as model training, hyperparameter tuning, and inference tasks.
These ML- and AI-oriented DAGs often run ad hoc, are triggered by external systems, or need to execute multiple times with different parameters over the same dataset. By allowing multiple DAG runs with logical_date=None, Airflow now supports these scenarios natively without requiring workarounds.
Airflow 3.0 introduces several configuration and interface updates that improve consistency, clarify ownership of core utilities, and remove legacy behaviors that were no longer aligned with modern usage patterns.
Airflow no longer silently updates configuration options that retain deprecated default values. Users are now required to explicitly set any config values that differ from the current defaults. This change improves transparency and prevents unintentional behavior changes during upgrades.
Several configuration defaults have changed in Airflow 3.0 to better reflect modern usage patterns:
The default value of catchup_by_default is now False. DAGs will not backfill missed intervals unless explicitly configured to do so.
The default value of create_cron_data_intervals is now False. Cron expressions are now interpreted using the CronTriggerTimetable instead of the legacy CronDataIntervalTimetable. This change simplifies interval logic and aligns with the future direction of Airflow's scheduling system. Set this flag explicitly before upgrading from Airflow 2 if you rely on data interval semantics; flipping it later (after Airflow 3 DAG runs exist) will skip one scheduled run per affected DAG.
Several core components have been moved to more intuitive or stable locations:
The SecretsMasker class has been relocated to airflow.sdk.execution_time.secrets_masker.
The ObjectStoragePath utility previously located under airflow.io is now available via airflow.sdk.
These changes simplify imports and reflect broader efforts to stabilize utility interfaces across the Airflow codebase.
Asset event mappings in the task context are improved to better support asset use cases, including new features introduced in AIP-74.
Events of an asset or asset alias are now accessed directly by a concrete object to avoid ambiguity. Using a str to access events is no longer supported. Use an Asset or AssetAlias object, or Asset.ref to refer to an entity explicitly instead, such as:
outlet_events[Asset.ref(name="myasset")] # Get events for asset named "myasset". outlet_events[AssetAlias(name="myalias")] # Get events for asset alias named "myalias".
Alternatively, two helpers for_asset and for_asset_alias are added as shortcuts:
outlet_events.for_asset(name="myasset") # Get events for asset named "myasset". outlet_events.for_asset_alias(name="myalias") # Get events for asset alias named "myalias".
The internal representation of asset event triggers now also includes an explicit uri field, simplifying traceability and aligning with the broader asset-aware execution model introduced in Airflow 3.0. DAG authors interacting directly with inlet_events may need to update logic that assumes the previous structure.
In Airflow 2, the xcom_pull() method allowed pulling XComs by key without specifying task_ids, despite the fact that the underlying DB model defines task_id as part of the XCom primary key. This created ambiguity: if two tasks pushed XComs with the same key, xcom_pull() would pull whichever one happened to be first, leading to unpredictable behavior.
Airflow 3 resolves this inconsistency by requiring task_ids when pulling by key. This change aligns with the task-scoped nature of XComs as defined by the schema, ensuring predictable and consistent behavior.
DAG Authors should update their dags to use task_ids if their dags used xcom_pull without task_ids such as:
kwargs["ti"].xcom_pull(key="key")
Should be updated to:
kwargs["ti"].xcom_pull(task_ids="task1", key="key")
As part of the deprecation cleanup, several legacy configuration options have been removed. These include:
[scheduler] allow_trigger_in_future
[scheduler] use_job_schedule
[scheduler] use_local_tz
[scheduler] processor_poll_interval
[logging] dag_processor_manager_log_location
[logging] dag_processor_manager_log_stdout
[logging] log_processor_filename_template
All the webserver configurations have also been removed since API server now replaces webserver, so the configurations like below have no effect:
[webserver] allow_raw_html_descriptions
[webserver] cookie_samesite
[webserver] error_logfile
[webserver] access_logformat
[webserver] web_server_master_timeout
etc
Several configuration options previously located under the [webserver] section have been moved to the new ``[api]`` section. The following configuration keys have been moved:
[webserver] web_server_host → [api] host
[webserver] web_server_port → [api] port
[webserver] workers → [api] workers
[webserver] web_server_worker_timeout → [api] worker_timeout
[webserver] web_server_ssl_cert → [api] ssl_cert
[webserver] web_server_ssl_key → [api] ssl_key
[webserver] access_logfile → [api] access_logfile
The following DAG parsing configuration options were moved to the new ``[dag_processor]`` section:
[core] dag_file_processor_timeout → [dag_processor] dag_file_processor_timeout
[scheduler] parsing_processes → [dag_processor] parsing_processes
[scheduler] file_parsing_sort_mode → [dag_processor] file_parsing_sort_mode
[scheduler] max_callbacks_per_loop → [dag_processor] max_callbacks_per_loop
[scheduler] min_file_process_interval → [dag_processor] min_file_process_interval
[scheduler] stale_dag_threshold → [dag_processor] stale_dag_threshold
[scheduler] print_stats_interval → [dag_processor] print_stats_interval
Users should review their airflow.cfg files or use the airflow config lint command to identify outdated or removed options.
Airflow 3.0 includes improved support for upgrade validation. Use the following tools to proactively catch incompatible configs or deprecated usage patterns:
airflow config lint: Identifies removed or invalid config keys
ruff check --select AIR30 --preview: Flags removed interfaces and common migration issues
Airflow 3.0 introduces changes to both the CLI and REST API interfaces to better align with service-oriented deployments and event-driven workflows.
The Airflow CLI has been split into two distinct interfaces:
The core airflow CLI now handles only local functionality (e.g., airflow tasks test, airflow dags list).
Remote functionality, including triggering DAGs or managing connections in service-mode environments, is now handled by a separate CLI called airflowctl, distributed via the apache-airflow-client package.
This change improves security and modularity for deployments that use Airflow in a distributed or API-first context.
The legacy REST API v1, previously built with Connexion and Marshmallow, has been replaced by a modern FastAPI-based REST API v2.
This new implementation improves performance, aligns more closely with web standards, and provides a consistent developer experience across the API and UI.
Key changes include stricter validation (422 errors instead of 400), the removal of the execution_date parameter in favor of logical_date, and more consistent query parameter handling.
The v2 API is now the stable, fully supported interface for programmatic access to Airflow, and also powers the new UI - achieving full feature parity between the UI and API.
For details, see the Airflow REST API v2 documentation.
The behavior of the POST /dags/{dag_id}/dagRuns endpoint has changed. If a logical_date is not explicitly provided when triggering a DAG via the REST API, it now defaults to None.
This aligns with event-driven DAGs and manual runs in Airflow 3.0, but may break backward compatibility with scripts or tools that previously relied on Airflow auto-generating a timestamped logical_date.
Several deprecated CLI arguments and commands that were marked for removal in earlier versions have now been cleaned up in Airflow 3.0. Run airflow --help to review the current set of available commands and arguments.
Deprecated --ignore-depends-on-past cli option is replaced by --depends-on-past ignore.
--tree flag for airflow tasks list command is removed. The format of the output with that flag can be expensive to generate and extremely large, depending on the DAG. airflow dag show is a better way to visualize the relationship of tasks in a DAG.
Changing dag_id from flag (-d, --dag-id) to a positional argument in the dags list-runs CLI command.
The airflow db init and airflow db upgrade commands have been removed. Use airflow db migrate instead to initialize or migrate the metadata database. If you would like to create default connections use airflow connections create-default-connections.
airflow api-server has replaced airflow webserver cli command.
Airflow 3.0 completes the migration of several core operators, sensors, hooks, and triggers into the new apache-airflow-providers-standard package. This package now includes commonly used components such as:
PythonOperator, BashOperator
ExternalTaskSensor, FileSensor
ShortCircuitOperator, LatestOnlyOperator
SubprocessHook, FilesystemHook
DateTimeTrigger, TimeDeltaTrigger, FileTrigger
These operators, sensors, hooks, and triggers were previously bundled inside airflow-core but are now treated as provider-managed components to improve modularity, testability, and lifecycle independence.
This change enables more consistent versioning across providers and prepares Airflow for a future where all integrations — including "standard" ones — follow the same interface model.
To maintain compatibility with existing DAGs, the apache-airflow-providers-standard package is installable on both Airflow 2.x and 3.x. Users upgrading from Airflow 2.x are encouraged to begin updating import paths and testing provider installation in advance of the upgrade.
Legacy imports such as airflow.operators.python.PythonOperator are deprecated and will be removed soon. They should be replaced with:
from airflow.providers.standard.operators.python import PythonOperator
The SimpleHttpOperator has been migrated to apache-airflow-providers-http and renamed to HttpOperator
Airflow 3.0 introduces a modernized user experience that complements the new React-based UI architecture (see Significant Changes). Several areas of the interface have been enhanced to improve visibility, consistency, and navigability.
The Airflow Home page now provides a high-level operational overview of your environment. It includes health checks for core components (Scheduler, Triggerer, DAG Processor), summary stats for DAG and task instance states, and a real-time feed of asset-triggered events. This view helps users quickly identify pipeline health, recent activity, and potential failures.
The DAG List page has been refreshed with a cleaner layout and improved responsiveness. Users can browse DAGs by name, tags, or owners. While full-text search has not yet been integrated, filters and navigation have been refined for clarity in large deployments.
The Graph and Grid views now display task information in the context of the DAG version that was used at runtime. This improves traceability for DAGs that evolve over time and provides more accurate debugging of historical runs.
The Graph view now supports visualizing the full chain of asset and task dependencies, including assets consumed or produced across DAG boundaries. This allows users to inspect upstream and downstream lineage in a unified view, making it easier to trace data flows, debug triggering behavior, and understand conditional dependencies between assets and tasks.
The "Code" tab now displays the exact DAG source as parsed by the scheduler for the selected DAG version. This allows users to inspect the precise code that was executed, even for historical runs, and helps debug issues related to versioned DAG changes.
Task log access has been streamlined across views. Logs are now easier to access from both the Grid and Task Instance pages, with cleaner formatting and reduced visual noise.
New UI components support asset-centric DAGs and backfill workflows:
Asset definitions are now visible from the DAG details page, allowing users to inspect upstream and downstream asset relationships.
Backfills can be triggered and monitored directly from the UI, including support for scheduler-managed backfills introduced in Airflow 3.0.
These improvements make Airflow more accessible to operators, data engineers, and stakeholders working across both time-based and event-driven workflows.
A number of deprecated features, modules, and interfaces have been removed in Airflow 3.0, completing long-standing migrations and cleanups.
Users are encouraged to review the following removals to ensure compatibility:
SubDag support has been removed entirely, including the SubDagOperator, related CLI and API interfaces. TaskGroups are now the recommended alternative for nested DAG structures.
SLAs have been removed: The legacy SLA feature, including SLA callbacks and metrics, has been removed. A more flexible replacement mechanism, DeadlineAlerts, is planned for a future version of Airflow. Users who relied on SLA-based notifications should consider implementing custom alerting using task-level success/failure hooks or external monitoring integrations.
Pickling support has been removed: All legacy features related to DAG pickling have been fully removed. This includes the PickleDag CLI/API, as well as implicit behaviors around store_serialized_dags = False. DAGs must now be serialized using the JSON-based serialization system. Ensure any custom Python objects used in DAGs are JSON-serializable.
Context parameter cleanup: Several previously available context variables have been removed from the task execution context, including conf, execution_date, and dag_run.external_trigger. These values are either no longer applicable or have been renamed (e.g., use dag_run.logical_date instead of execution_date). DAG authors should ensure that templated fields and Python callables do not reference these deprecated keys.
Deprecated core imports have been fully removed. Any use of airflow.operators.*, airflow.hooks.*, or similar legacy import paths should be updated to import from their respective providers.
Configuration cleanup: Several legacy config options have been removed, including:
scheduler.allow_trigger_in_future: DAG runs can no longer be triggered with a future logical date. Use logical_date=None instead.
scheduler.use_job_schedule and scheduler.use_local_tz have also been removed. These options were deprecated and no longer had any effect.
Deprecated utility methods such as those in airflow.utils.helpers, airflow.utils.process_utils, and airflow.utils.timezone have been removed. Equivalent functionality can now be found in the standard Python library or Airflow provider modules.
Removal of deprecated CLI flags and behavior: Several CLI entrypoints and arguments that were marked for removal in earlier versions have been cleaned up.
To assist with the upgrade, tools like ruff (e.g., rule AIR302) and airflow config lint can help identify obsolete imports and configuration keys. These utilities are recommended for locating and resolving common incompatibilities during migration. Please see Upgrade Guide for more information.
The following table summarizes user-facing features removed in 3.0 and their recommended replacements. Not all of these are called out individually above.
Feature |
Replacement / Notes |
|---|---|
SubDagOperator / SubDAGs |
Use TaskGroups |
SLA callbacks / metrics |
Deadline Alerts (planned post-3.0) |
DAG Pickling |
Use JSON serialization; pickling is no longer supported |
Xcom Pickling |
Use custom Xcom backend; pickling is no longer supported |
execution_date context var |
Use dag_run.logical_date |
conf and dag_run.external_trigger |
Removed from context; use DAG params or dag_run APIs |
Core EmailOperator |
Use EmailOperator from the smtp provider |
none_failed_or_skipped rule |
Use none_failed_min_one_success |
dummy trigger rule |
Use always |
fail_stop argument |
Use fail_fast |
store_serialized_dags=False |
DAGs are always serialized; config has no effect |
Deprecated core imports |
Import from appropriate provider package |
SequentialExecutor & DebugExecutor |
Use LocalExecutor for testing |
.airflowignore regex |
Uses glob syntax by default |
Airflow 3 was designed with migration in mind. Many Airflow 2 DAGs will work without changes, especially if deprecation warnings were addressed in earlier releases. To support the upgrade, Airflow 3 includes validation tools such as ruff and airflow config update, as well as a simplified startup model.
For a step-by-step upgrade process, see the Upgrade Guide.
To upgrade to Airflow 3.0, you must be running Airflow 2.7 or later.
Airflow 3.0 supports the following Python versions:
Python 3.9
Python 3.10
Python 3.11
Python 3.12
Earlier versions of Airflow or Python are not supported due to architectural changes and updated dependency requirements.
Airflow now includes a Ruff-based linter with custom rules to detect DAG patterns and interfaces that are no longer compatible with Airflow 3.0. These checks are packaged under the AIR30x rule series. Example usage:
ruff check dags/ --select AIR301 --preview
ruff check dags/ --select AIR301 --fix --preview
These checks can automatically fix many common issues such as renamed arguments, removed imports, or legacy context variable usage.
Airflow 3.0 introduces a new utility to validate and upgrade your Airflow configuration file:
airflow config update
airflow config update --fix
This utility detects removed or deprecated configuration options and, if desired, updates them in-place.
Additional validation is available via:
airflow config lint
This command surfaces obsolete configuration keys and helps align your environment with Airflow 3.0 requirements.
As with previous major releases, the Airflow 3.0 upgrade includes schema changes to the metadata database. Before upgrading, it is strongly recommended that you back up your database and optionally run:
airflow db clean
to remove old task instance, log, or XCom data. To apply the new schema:
airflow db migrate
Airflow components are now started explicitly. For example:
airflow api-server # Replaces airflow webserver
airflow dag-processor # Required in all environments
These changes reflect Airflow's new service-oriented architecture.
Upgrade Guide
Airflow 3.0 represents more than a year of collaboration across hundreds of contributors and dozens of organizations. We thank everyone who helped shape this release through design discussions, code contributions, testing, documentation, and community feedback. For full details, migration guidance, and upgrade best practices, refer to the official Upgrade Guide and join the conversation on the Airflow dev and user mailing lists.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
…The core option is deprecated in Airflow 2.11.2. Both options are removed in Airflow 3 as historical templates are supported and does not cause low-se…
Fix Task Instances list view rendering raw HTML instead of clickable links for Dag Id, Task Id, and Run Id columns. (#62533)
In 2.11.1 by mistake core.use_historical_filename_templates was read by Airflow instead of logging.use_historical_filename_templates. The core option is deprecated in Airflow 2.11.2. Both options are removed in Airflow 3 as historical templates are supported and does not cause low-severity security issue in Airflow 3. (#62647)
gracefully handle 404 from worker log server for historical retry attempts (#63002)
task_instance_mutation_hook receives a TI with run_id set (#62999)
fix missing logs in UI for tasks in UP_FOR_RETRY and UP_FOR_RESCHEDULE states (#54547) (#62877)
Fixing 500 error on webserver after upgrading to FAB provider 1.5.4 (#62412)
Lazily import fs and package_index hook in providers manager #52117 (#62356)
Upgrade airflow UI to latest reasonable dependencies. (#63158)
Bump the core-ui-package-updates group across 1 directory with 87 updates (#61091)
bump filelock (#62952)
Limit Celery Provider to not install 3.17.0 as it breaks airflow 2.11 (#63046)
Bump the pip-dependency-updates group across 3 directories with 5 updates (#62808)
Upgrade to latest released build dependencies (#62613)
Nothing published for this version
…the security of the system and prevent potential vulnerabilities in case the potential of executing arbitrary code in webserver is important for Airfl…
Support for Python 3.9 has been removed, as it has reached end-of-life. Airflow 2.11.1 requires Python 3.10, 3.11, or 3.12. Note that this is unusual to remove Python version support in patch-level release of Airflow, but since Python 3.9 is already end-of-life, many libraries do not support it any more, and Airflow 2.11.1 is focused on improving security by upgrading dependencies, so we decided to remove Python 3.9 support in this patch release, to improve security of the release. Python 3.10 and 3.11 had almost no backward-incompatible changes, so you should be able to upgrade to Python 3.10 or 3.11 easily. If you were using Python 3.9 before, it is recommended to first upgrade Python version in existing installation and then upgrade to Airflow 2.11.1.
In Airflow 3.0, the timer_unit_consistency setting in the metrics section will be enabled by default and setting itself will be removed. This will standardize all timer and timing metrics to milliseconds across all metric loggers.
Users Integrating with Datadog, OpenTelemetry, or other metric backends should enable this setting. For users, using statsd, this change will not affect you.
If you need backward compatibility, you can leave this setting disabled temporarily, but enabling timer_unit_consistency is encouraged to future-proof your metrics setup. (#39908)
When you change the log template in Airflow 2.11.1, the historical log templates are not retrieved. This means that if you have existing logs that were generated using a different log template, they will not be accessible using the new log template.
This change is due to potential security issues that could arise from retrieving historical log templates, which allow Dag Authors to execute arbitrary code in webserver when retrieving logs. By disabling the retrieval of historical log templates, Airflow 2.11.1 aims to enhance the security of the system and prevent potential vulnerabilities in case the potential of executing arbitrary code in webserver is important for Airflow deployment.
Users who need to access historical logs generated with a different log template will need to manually update their log files to match the naming of their historical log files with the latest log template configured in Airflow configuration, or they can set the "core.use_historical_filename_templates" configuration option to True to enable the retrieval of historical log templates, if they are fine with the Dag Authors being able to execute arbitrary code in webserver when retrieving logs. (#61880)
Airflow 2.11.1 includes updates to a number of dependencies including connexion, Flask-Session, Werkzeug, that were not possible to upgrade before, because the dependencies did not have compatible versions with Airflow 2.11.0, but we worked together with the community to update them. Many thanks to connexion team and a number of community members to help with the updates so that we could upgrade to newer versions and get rid of some dependency versions that had known security vulnerabilities (#51681)
Add proxy values to be masked by secrets manager (#61906)
Masking details while creating connections using json & uri (#61882)
Fix redaction of illegal args (#61883)
Fix stuck queued tasks by calling executor fail method and invoking failure callbacks (#53038)
Fix recursion depth error in _redact_exception_with_context (#61797)
Avoid warning when passing none as dataset alias (#61791)
Add pool name validation to avoid XSS from the DAG file (#61732)
Prevent scheduler to crash due to RecursionError when making a SQL query (#55778)
Fix root logger level cache invalidation in LoggerMutationHelper (#61644)
update null event values to empty string in downgrade for migration revision_id d75389605139 (#57131)
Fix WeightRule spec (#53947)
Correctly treat request on reschedule sensors as resetting after each reschedule (#51410) (#52638)
Allow more empty loops before stopping log streaming (#52614) (#52636)
Ensuring XCom return value can be mapped for dynamically-mapped @task_group's (#51668)
Fix archival for cascading deletes by archiving dependent tables first (#51952) (#52211)
Stop streaming task logs if end of log mark is missing (#51904)
Fix bad width w/no options in multi-select DAG parameter (#51516)
Fix remove filter button visibility in Pools list page (#51161)
Fix delete button visibility in search filters (#51100)
Fix migration from 2.2.0 to 2.11.0 for Sqlite (#50745)
Check if stand alone dag processor is active in get_health endpoint (#48612)
Nothing published for this version
Nothing published for this version
Emit warning for deprecated BaseOperatorLink.get_link signature
DeltaTriggerTimetable for trigger-based scheduling (#47074)This change introduces DeltaTriggerTimetable, a new built-in timetable that complements the existing suite of Airflow timetables by supporting delta-based trigger schedules without relying on data intervals.
Airflow currently has two major types of timetables:
CronDataIntervalTimetable, DeltaDataIntervalTimetable)CronTriggerTimetable)However, there was no equivalent trigger-based option for delta intervals like timedelta(days=1).
As a result, even simple schedules like schedule=timedelta(days=1) were interpreted through a data interval
lens—adding unnecessary complexity for users who don't care about upstream/downstream data dependencies.
This feature is backported to Airflow 2.11.0 to help users begin transitioning before upgrading to Airflow 3.0.
schedule=timedelta(...) still defaults to DeltaDataIntervalTimetable.[scheduler] create_delta_data_intervals (default: True) allows opting in to DeltaTriggerTimetable.False, meaning DeltaTriggerTimetable becomes the default for timedelta schedules.By flipping this config in 2.11, users can preview and adopt the new scheduling behavior in advance — minimizing surprises during upgrade.
Previously, Airflow reported timing metrics in milliseconds for StatsD but in seconds for other backends
such as OpenTelemetry and Datadog. This inconsistency made it difficult to interpret or compare
timing metrics across systems.
Airflow 2.11 introduces a new config option:
[metrics] timer_unit_consistency (default: False in 2.11, True and dropped in Airflow 3.0).When enabled, all timing metrics are consistently reported in milliseconds, regardless of the backend.
This setting has become mandatory and always True in Airflow 3.0 (the config will be removed), so
enabling it in 2.11 allows users to migrate early and avoid surprises during upgrade.
This release introduces several changes to help users prepare for upgrading to Airflow 3:
execution_date now also include a logical_date field. Airflow 3 drops execution_date entirely in favor of logical_date (#44283)airflow config lint and airflow config update commands in 2.11 to help audit and migrate configs for Airflow 3.0. (#45736, #50353, #46757)Support for Python 3.8 has been removed, as it has reached end-of-life. Airflow 2.11 requires Python 3.9, 3.10, 3.11, or 3.12.
DeltaTriggerTimetable (#47074)airflow config update and airflow config lint changes to ease migration to Airflow 3 (#45736, #50353)ti.log_url timestamp format from "%Y-%m-%dT%H:%M:%S%z" to "%Y-%m-%dT%H:%M:%S.%f%z" (#50306)airflow.cfg contains a random fernet_key and secret_key (#47755)rendered_map_index via internal api (#49057)TaskInstancePydantic into TaskInstance (#48571)log_url property on TaskInstancePydantic (Internal API) (#50560)TypeError when deserializing task with execution_timeout set to None (#46822)check_query_exists returns a bool (#46707)/xcom/list got exception when applying filter on the value column (#46053)logical_date to models using execution_date (#44283)BaseOperatorLink.get_link signature (#46448)airflow.cfg variable (#48084)XCom docs to show examples of pushing multiple XComs (#46284, #47068)Nothing published for this version
Deprecate conf from Task Context
Previously when a DAG run was manually set to "failed" or to "success" state the terminal state was set to all tasks. But this was a gap for cases when setup- and teardown tasks were defined: If teardown was used to clean-up infrastructure or other resources, they were also skipped and thus resources could stay allocated.
As of now when setup tasks had been executed before and the DAG is manually set to "failed" or "success" then teardown tasks are executed. Teardown tasks are skipped if the setup was also skipped.
As a side effect this means if the DAG contains teardown tasks, then the manual marking of DAG as "failed" or "success" will need to keep the DAG in running state to ensure that teardown tasks will be scheduled. They would not be scheduled if the DAG is directly set to "failed" or "success".
trigger_rule=TriggerRule.ALWAYS in a task-generated mapping within bare tasks (#44751)ONE_DONE) in a mapped task group (#44937)FileTaskHandler only read from default executor (#46000)skip_if and run_if decorators before TaskFlow virtualenv tasks are run (#41832) (#45680)rendered_map_index (#45109) (#45122)max_form_parts, max_form_memory_size (#46243) (#45749)execute safeguard mechanism (#44646) (#46280)conf from Task Context (#44993)Nothing published for this version
Raise deprecation warning when accessing inlet or outlet events through str
priority_weight is capped in 32-bit signed integer ranges (#43611)Some database engines are limited to 32-bit integer values. As some users reported errors in
weight rolled-over to negative values, we decided to cap the value to the 32-bit integer. Even
if internally in python smaller or larger values to 64 bit are supported, priority_weight is
capped and only storing values from -2147483648 to 2147483647.
trigger_rule="always" in a dynamic mapped task (#43810)trigger_rule=TriggerRule.ALWAYS in a task-generated mapping within bare tasks (#44751)Nothing published for this version
Improves the handling of value masking when setting Airflow variables for enhanced security.
No significant changes.
stringified objects to UI via xcom if pickling is active (#42388) (#42486)selectinload instead of joinedload (#40487) (#42351)TrySelector for Mapped Tasks in Logs and Details Grid Panel (#43566)scheduler_loop_duration (#42886) (#43544)dompurify from 2.2.9 to 2.5.6 in /airflow/www (#42263) (#42270)4.5.2 (#43309) (#43318)Nothing published for this version
Nothing published for this version
Deprecate --tree flag for tasks list cli command
No significant changes.
renderedTemplates as keys to skip camelCasing (#42206) (#42208)camelcase xcom entries (#42182) (#42187)0.2.4 as it breaks our integration (#42101)LibCST (#42089)--tree flag for tasks list cli command (#41965)security_model.rst to clear unauthenticated endpoints exceptions (#42085)Nothing published for this version
Remove deprecation warning for cgitb in Plugins Manager
No significant changes.
tojson filter to example_inlet_event_extra example dag (#41890)keycloak (#41791)Nothing published for this version
…occur when the DAG is active. While this is a breaking change, the previous behavior is considered a bug.
Previously, when a DAG is paused or removed, incoming dataset events would still trigger it, and the DAG would run when it is unpaused or added back in a DAG file. This has been changed; a DAG's dataset schedule can now only be satisfied by events that occur when the DAG is active. While this is a breaking change, the previous behavior is considered a bug.
The behavior of time-based scheduling is unchanged, including the timetable part
of DatasetOrTimeSchedule.
try_number is no longer incremented during task execution (#39336)Previously, the try number (try_number) was incremented at the beginning of task execution on the worker. This was problematic for many reasons.
For one it meant that the try number was incremented when it was not supposed to, namely when resuming from reschedule or deferral. And it also resulted in
the try number being "wrong" when the task had not yet started. The workarounds for these two issues caused a lot of confusion.
Now, instead, the try number for a task run is determined at the time the task is scheduled, and does not change in flight, and it is never decremented. So after the task runs, the observed try number remains the same as it was when the task was running; only when there is a "new try" will the try number be incremented again.
One consequence of this change is, if users were "manually" running tasks (e.g. by calling ti.run() directly, or command line airflow tasks run),
try number will no longer be incremented. Airflow assumes that tasks are always run after being scheduled by the scheduler, so we do not regard this as a breaking change.
/logout endpoint in FAB Auth Manager is now CSRF protected (#40145)The /logout endpoint's method in FAB Auth Manager has been changed from GET to POST in all existing
AuthViews (AuthDBView, AuthLDAPView, AuthOAuthView, AuthOIDView, AuthRemoteUserView), and
now includes CSRF protection to enhance security and prevent unauthorized logouts.
This new feature adds capability for Apache Airflow to emit 1) airflow system traces of scheduler, triggerer, executor, processor 2) DAG run traces for deployed DAG runs in OpenTelemetry format. Previously, only metrics were supported which emitted metrics in OpenTelemetry. This new feature will add richer data for users to use OpenTelemetry standard to emit and send their trace data to OTLP compatible endpoints.
(@skip_if, @run_if) to make it simple to apply whether or not to skip a Task. (#41116)This feature adds a decorator to make it simple to skip a Task.
Previously known as hybrid executors, this new feature allows Airflow to use multiple executors concurrently. DAGs, or even individual tasks, can be configured
to use a specific executor that suits its needs best. A single DAG can contain tasks all using different executors. Please see the Airflow documentation for
more details. Note: This feature is still experimental. See documentation on Executor <https://airflow.apache.org/docs/apache-airflow/stable/core-concepts/executor/index.html#using-multiple-executors-concurrently>_ for a more detailed description.
Airflow integrates Scarf to collect basic usage data during operation. Deployments can opt-out of data collection by setting the [usage_data_collection]enabled option to False, or the SCARF_ANALYTICS=false environment variable.
See FAQ on this <https://airflow.apache.org/docs/apache-airflow/stable/faq.html#does-airflow-collect-any-telemetry-data>_ for more information.
accessors to read dataset events defined as inlet (#39367)dag test (#40010)endDate in task instance tooltip. (#39547)accessors to read dataset events defined as inlet (#39367, #39893)run_if & skip_if decorators (#41116)renderedjson component (#40964)get_extra_dejson method with nested parameter which allows you to specify if you want the nested json as string to be also deserialized (#39811)__getattr__ to task decorator stub (#39425)RemovedIn20Warning in airflow task command (#39244)db migrate error messages (#39268)suppress_and_warn warning (#39263)declarative_base from sqlalchemy.orm instead of sqlalchemy.ext.declarative (#39134)on_task_instance_failed access to the error that caused the failure (#38155)output_processor parameter to BashProcessor (#40843)never_fail in BaseSensor (#40915)start_date (#40878)external_task_group_id to WorkflowTrigger (#39617)BaseSensorOperator introduce skip_policy parameter (#40924)__init__ (#41086)OTel Traces (#40874)pydocstyle rules to pyproject.toml (#40569)pydocstyle rule D213 in ruff. (#40448, #40464)Dag.test() to run with an executor if desired (#40205)AirflowInternalRuntimeError for raise non catchable errors (#38778)pytest to 8.0+ (#39450)back_populates between DagScheduleDatasetReference.dag and DagModel.schedule_dataset_references (#39392)B028 (no-explicit-stacklevel) in core (#39123)ImportError to ParseImportError for avoid shadowing with builtin exception (#39116)SubDagOperator examples warnings (#39057)model_dump instead of dict for serialize Pydantic V2 model (#38933)ws from 7.5.5 to 7.5.10 in /airflow/www (#40288)filesystems and dataset-uris to "how to create your own provider" page (#40801)otel_on to True in example airflow.cfg (#40712)task_id from send_email to send_email_notification in taskflow.rst (#41060)Nothing published for this version
Nothing published for this version
Nothing published for this version
Add [webserver]update_fab_perms to deprecated configs
scheduled_duration and queued_duration changed (#37936)scheduled_duration and queued_duration metrics are now emitted in milliseconds instead of seconds.
By convention all statsd metrics should be emitted in milliseconds, this is later expected in e.g. prometheus statsd-exporter.
Experimental support for OpenTelemetry was added in 2.7.0 since then fixes and improvements were added and now we announce the feature as stable.
[webserver]update_fab_perms to deprecated configs (#40317)httpx to requests in file_task_handler (#39799)SchedulerJobRunner._process_executor_events (#40563)Nothing published for this version
Do not provide deprecated execution_date in @apply_lineage
No significant changes.
AirflowSecurityManagerV2 leave transactions in the idle in transaction state (#39935)SafeDogStatsdLogger to use get_validator to enable pattern matching (#39370)has_access (#39421)execution_date in @apply_lineage (#39327)sql_alchemy_engine_args config example (#38971)yandex provider to avoid mypy errors (#39990)provider_info_cache decorator (#39750)defer (#39742)idx_last_scheduling_decision on dag_run table (#39275)CronDataIntervalTimetable (#39780)Nothing published for this version
Undeprecate BaseXCom.get_one method for now
10.17.0 or later (#38071)If you use Stackdriver logging, you must use Google provider version 10.17.0 or later. Airflow 2.9.1 now passes gcp_log_name to the StackdriverTaskHandler instead of name, and this will fail on earlier provider versions.
This fixes a bug where the log name configured in [logging] remove_base_log_folder was overridden when Airflow configured logging, resulting in task logs going to the wrong destination.
href for nav bar (#39282)firefox (#39261)log_url (#39183)UX (#39119)ux in react dag page (#39122)AUTH_ROLE_PUBLIC is set in check_authentication (#39012)map_index_template so it renders for failed tasks as long as it was defined before the point of failure (#38902)Undeprecate BaseXCom.get_one method for now (#38991)inherit_cache attribute for CreateTableAs custom SA Clause (#38985)SAWarning 'Coercing Subquery object into a select() for use in IN()' (#38926)cartesian product in AirflowSecurityManagerV2 (#38913)methodtools.lru_cache instead of functools.lru_cache in class methods (#37757)airflow dags backfill only if -I / --ignore-first-depends-on-past provided (#38676)TriggerDagRunOperator deprecate execution_date in favor of logical_date (#39285)@deprecated decorator (#39205)is_authorized_custom_view from auth manager to handle custom actions (#39167)minischeduler skip (#38976)undici from 5.28.3 to 5.28.4 in /airflow/www (#38751)PythonOperator op_kwargs (#39242)user and role commands (#39224)k8s 1.29 to supported version in docs (#39168)DagBag class docstring to include all params (#38814)Nothing published for this version
Nothing published for this version
Bump aiohttp min version to avoid CVE-2024-23829 and CVE-2024-23334
Lifecycle events:
on_startingbefore_stoppingDagRun State Change Events:
on_dag_run_runningon_dag_run_successon_dag_run_failedTaskInstance State Change Events:
on_task_instance_runningon_task_instance_successon_task_instance_failedAfter discussion <https://lists.apache.org/thread/r06j306hldg03g2my1pd4nyjxg78b3h4>__
and a voting process <https://lists.apache.org/thread/pgcgmhf6560k8jbsmz8nlyoxosvltph2>__,
the Airflow's PMC and Committers have reached a resolution to no longer maintain MsSQL as a supported Database Backend.
As of Airflow 2.9.0 support of MsSQL has been removed for Airflow Database Backend.
A migration script which can help migrating the database before upgrading to Airflow 2.9.0 is available in
airflow-mssql-migration repo on Github <https://github.com/apache/airflow-mssql-migration>_.
Note that the migration script is provided without support and warranty.
This does not affect the existing provider packages (operators and hooks), DAGs can still access and process data from MsSQL.
Datasets must use a URI that conform to rules laid down in AIP-60, and the value
will be automatically normalized when the DAG file is parsed. See
documentation on Datasets <https://airflow.apache.org/docs/apache-airflow/stable/authoring-and-scheduling/datasets.html>_ for
a more detailed description on the rules.
You may need to change your Dataset identifiers if they look like a URI, but are used in a less mainstream way, such as relying on the URI's auth section, or have a case-sensitive protocol name.
get_permitted_menu_items in BaseAuthManager has been renamed filter_permitted_menu_items (#37627)The Audit Log event name for REST API events will be prepended with api. or ui., depending on if it came from the Airflow UI or externally.
There are a few caveats though:
Pendulum2 does not support Python 3.12. For Python 3.12 you need to use
Pendulum 3 <https://pendulum.eustace.io/blog/announcing-pendulum-3-0-0.html>_
Minimum SQLAlchemy version supported when Pandas is installed for Python 3.12 is 1.4.36 released in
April 2022. Airflow 2.9.0 increases the minimum supported version of SQLAlchemy to 1.4.36 for all
Python versions.
Not all Providers support Python 3.12. At the initial release of Airflow 2.9.0 the following providers are released without support for Python 3.12:
apache.beam - pending on Apache Beam support for 3.12 <https://github.com/apache/beam/issues/29149>_papermill - pending on Releasing Python 3.12 compatible papermill client version
including this merged issue <https://github.com/nteract/papermill/pull/771>_There's now a limit to the length of data that can be stored in the Rendered Template Fields.
The limit is set to 4096 characters. If the data exceeds this limit, it will be truncated. You can change this limit
by setting the [core]max_template_field_length configuration option in your airflow config.
Xcom table column value type has changed from blob to longblob. This will allow you to store relatively big data in Xcom but process can take a significant amount of time if you have a lot of large data stored in Xcom.
To downgrade from revision: b4078ac230a1, ensure that you don't have Xcom values larger than 65,535 bytes. Otherwise, you'll need to clean those rows or run airflow db clean xcom to clean the Xcom table.
Matomo as an option for analytics_tool. (#38221)hashable (#37465)queuedEvent endpoint to get/delete DatasetDagRunQueue (#37176)DatasetOrTimeSchedule (#36710)on_skipped_callback to BaseOperator (#36374)@task.bash TaskFlow decorator (#30176, #37875)ExternalPythonOperator use version from sys.version_info (#38377)run_id column to log table (#37731)tryNumber to grid task instance tooltip (#37911)ExternalPythonOperator (#37409)Pathlike (#36947)nowait and skip_locked into with_row_locks (#36889)dag/dagRun in the REST API (#36641)Connexion from auth manager interface (#36209)total_entries count on the event logs endpoint (#38625)tz in next run ID info (#38482)chakra styles to keep dropdowns in filter bar (#38456)__exit__ is called in decorator context managers (#38383)BaseAuthManager.is_authorized_custom_view abstract (#37915)/get_logs_with_metadata endpoint (#37756)encoding to the SQL engine in SQLAlchemy v2 (#37545)consuming_dags attr eagerly before dataset listener (#36247)importlib_metadata with compat to Python 3.10/3.12 stdlib (#38366)__new__ magic method of BaseOperatorMeta to avoid bad mixing classic and decorated operators (#37937)sys.version_info for determine Python Major.Minor (#38372)blinker add where it requires (#38140)> 39.0.0 (#38112)assert outside of the tests (#37718)flask._request_ctx_stack (#37522)login attribute in airflow.__init__.py (#37565)datetime.datetime.utcnow by airflow.utils.timezone.utcnow in core (#35448)is_authorized_cluster_activity from auth manager (#36175)exception to templates ref list (#36656)Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Fix incorrect serialization of FixedTimezone
No significant changes.
FixedTimezone (#38139)ObjectStoragePath (#37769)pytest_rewrites (#38095, #38139)pandas to <2.2 (#37748)croniter to fix an issue with 29 Feb cron expressions (#38198)Nothing published for this version
Add "MENU" permission in auth manager
airflow_pre_installed_providers.txt artifact (#37679)BranchDayOfWeekOperator (#37813)ERD generating doc improvement (#37808)Nothing published for this version
Replace other Python 3.11 and 3.12 deprecations
allowed_deserialization_classes flag now follows a glob pattern (#36147).For example if one wants to add the class airflow.tests.custom_class to the
allowed_deserialization_classes list, it can be done by writing the full class
name (airflow.tests.custom_class) or a pattern such as the ones used in glob
search (e.g., airflow.*, airflow.tests.*).
If you currently use a custom regexp path make sure to rewrite it as a glob pattern.
Alternatively, if you still wish to match it as a regexp pattern, add it under the new
list allowed_deserialization_classes_regexp instead.
This was done under the policy that we do not want users like Viewer, Ops, and other users apart from Admin to have access to audit_logs. The intention behind this change is to restrict users with less permissions from viewing user details like First Name, Email etc. from the audit_logs when they are not permitted to.
The impact of this change is that the existing users with non admin rights won't be able to view or access the audit_logs, both from the Browse tab or from the DAG run.
AirflowTimeoutError is no longer except by default through Exception (#35653).The AirflowTimeoutError is now inheriting BaseException instead of
AirflowException->Exception.
See https://docs.python.org/3/library/exceptions.html#exception-hierarchy
This prevents code catching Exception from accidentally
catching AirflowTimeoutError and continuing to run.
AirflowTimeoutError is an explicit intent to cancel the task, and should not
be caught in attempts to handle the error and return some default value.
Catching AirflowTimeoutError is still possible by explicitly excepting
AirflowTimeoutError or BaseException.
This is discouraged, as it may allow the code to continue running even after
such cancellation requests.
Code that previously depended on performing strict cleanup in every situation
after catching Exception is advised to use finally blocks or
context managers. To perform only the cleanup and then automatically
re-raise the exception.
See similar considerations about catching KeyboardInterrupt in
https://docs.python.org/3/library/exceptions.html#KeyboardInterrupt
IMPORT_ERROR from DAG related permissions to view related permissions (#37292)AirflowTaskTimeout to inherit BaseException (#35653)namedtuple (#37168)Treeview function (#37162)access_entity is specified (#37290)dateTimeAttrFormat constant (#37285)@Sentry.enrich_errors (#37002)dryrun auto-fetch (#36941)/variables endpoint (#36820)pendulum.from_timestamp usage (#37160)CLI instead of specific one (#37651)undici from 5.26.3 to 5.28.3 in /airflow/www (#37493)3.12 exclusions in providers/pyproject.toml (#37404)markdown from core dependencies (#37396)pageSize method. (#37319)Python 3.11 and 3.12 deprecations (#37478)airflow_pre_installed_providers.txt into sdist distribution (#37388)universal-pathlib to < 0.2.0 (#37311)queue_when (#36997)config.yml for environment variable sql_alchemy_connect_args (#36526)Alembic to 1.13.1 (#36928)flask-session to <0.6 (#36895)CLI flags available (#37231)otel config descriptions (#37229)Objectstore tutorial with prereqs section (#36983)package/module names (#36927)__init__ of operators automatically (#33786)Nothing published for this version
Nothing published for this version
Nothing published for this version
Get rid of pyarrow-hotfix for CVE-2023-47248
pendulum package set to 3 (#36281).Support for pendulum 2.1.2 will be saved for a while, presumably until the next feature version of Airflow. It is advised to upgrade user code to use pendulum 3 as soon as possible.
We standardized Airflow dependency configuration to follow latest development in Python packaging by
using pyproject.toml. Airflow is now compliant with those accepted PEPs:
PEP-440 Version Identification and Dependency Specification <https://www.python.org/dev/peps/pep-0440/>__PEP-517 A build-system independent format for source trees <https://www.python.org/dev/peps/pep-0517/>__PEP-518 Specifying Minimum Build System Requirements for Python Projects <https://www.python.org/dev/peps/pep-0518/>__PEP-561 Distributing and Packaging Type Information <https://www.python.org/dev/peps/pep-0561/>__PEP-621 Storing project metadata in pyproject.toml <https://www.python.org/dev/peps/pep-0621/>__PEP-660 Editable installs for pyproject.toml based builds (wheel based) <https://www.python.org/dev/peps/pep-0660/>__PEP-685 Comparison of extra names for optional distribution dependencies <https://www.python.org/dev/peps/pep-0685/>__Also we implement multiple license files support coming from Draft, not yet accepted (but supported by hatchling) PEP:
PEP 639 Improving License Clarity with Better Package Metadata <https://peps.python.org/pep-0639/>__This has almost no noticeable impact on users if they are using modern Python packaging and development tools, generally
speaking Airflow should behave as it did before when installing it from PyPI and it should be much easier to install
it for development purposes using pip install -e ".[devel]".
The differences from the user side are:
- (following PEP-685) instead of _ and .
(as it was before in some extras). When you install airflow with such extras (for example dbt.core or
all_dbs) you should use - instead of _ and ..In most modern tools this will work in backwards-compatible way, but in some old version of those tools you might need to
replace _ and . with -. You can also get warnings that the extra you are installing does not exist - but usually
this warning is harmless and the extra is installed anyway. It is, however, recommended to change to use - in extras in your dependency
specifications for all Airflow extras.
Released airflow package does not contain devel, devel-*, doc and doc-gen extras.
Those extras are only available when you install Airflow from sources in --editable mode. This is
because those extras are only used for development and documentation building purposes and are not needed
when you install Airflow for production use. Those dependencies had unspecified and varying behaviour for
released packages anyway and you were not supposed to use them in released packages.
The all and all-* extras were not always working correctly when installing Airflow using constraints
because they were also considered as development-only dependencies. With this change, those dependencies are
now properly handling constraints and they will install properly with constraints, pulling the right set
of providers and dependencies when constraints are used.
The graphviz dependency has been problematic as Airflow required dependency - especially for
ARM-based installations. Graphviz packages require binary graphviz libraries - which is already a
limitation, but they also require to install graphviz Python bindings to be build and installed.
This does not work for older Linux installation but - more importantly - when you try to install
Graphviz libraries for Python 3.8, 3.9 for ARM M1 MacBooks, the packages fail to install because
Python bindings compilation for M1 can only work for Python 3.10+.
This is not a breaking change technically - the CLIs to render the DAGs is still there and IF you already have graphviz installed, it will continue working as it did before. The only problem when it does not work is where you do not have graphviz installed it will raise an error and inform that you need it.
Graphviz will remain to be installed for most users:
The only change will be a new installation of new version of Airflow from the scratch, where graphviz will need to be specified as extra or installed separately in order to enable DAG rendering option.
taskinstance list (#36693)AUTH_ROLE_PUBLIC=admin (#36750)op subtypes (#35536)typing.Union in _infer_multiple_outputs for Python 3.10+ (#36728)multiple_outputs is inferred correctly even when using TypedDict (#36652)Dagrun.update_state (#36712)EventsTimetable schedule past events if catchup=False (#36134)tis_query in _process_executor_events (#36655)call_regular_interval (#36608)DagRun fails while running dag test (#36517)_manage_executor_state by refreshing TIs in batch (#36502)MAX_CONTENT_LENGTH (#36401)kubernetes decorator type annotation consistent with operator (#36405)api/dag/*/dagrun from anonymous user (#36275)DAG.is_fixed_time_schedule (#36370)httpx import in file_task_handler for performance (#36753)pyarrow-hotfix for CVE-2023-47248 (#36697)graphviz dependency optional (#36647)pandas dependency to 1.2.5 for all providers and airflow (#36698)/airflow/www (#36700)docker decorator type annotations (#36406)batch_is_authorized_dag to check if user has permission to read DAGs (#36279)numpy example with practical exercise demonstrating top-level code (#35097)dags.rst with information on DAG pausing (#36540)metrics.rst for param dagrun.schedule_delay (#36404)One column per quarter.