Airflow Updates & Release Notes

Apache Airflow 3.3.0 introduces a first-class state store for tasks and assets (AIP-103), a Language Task SDK for writing tasks in Java and Go (AIP-108), a major expansion of asset partitioning, and pluggable retry policies.

We’re proud to announce the release of Apache Airflow 3.3.0! Where 3.2 brought precision to data with asset partitioning, 3.3 gives your tasks memory and multi-language support: a first-class state store for tasks and assets, a Language Task SDK for writing task logic in Java and Go, a major expansion of asset partitioning, and pluggable retry policies.

🎯 Release Highlights

📦 PyPI: https://pypi.org/project/apache-airflow/3.3.0/
📚 Docs: https://airflow.apache.org/docs/apache-airflow/3.3.0/
🛠️ Release Notes: https://airflow.apache.org/docs/apache-airflow/3.3.0/release_notes.html
🐳 Docker Image: docker pull apache/airflow:3.3.0
🚏 Constraints: https://github.com/apache/airflow/tree/constraints-3.3.0

🗃️ Task & Asset State Store (AIP-103): Tasks That Remember

Until now, if a task needed to remember something across retries or runs — a cursor, a checkpoint, a high-water mark — you reached for XComs, an external store, or a clever Variable hack. Airflow 3.3 makes durable task state a first-class concept.

Tasks can persist arbitrary key-value state that survives across retries and runs via a new task_state_store accessor, and assets can carry their own state via asset_state_store — both available directly from the Task SDK. State lives in the metadata database by default, or in a custom worker-side backend ([workers] state_store_backend), supports per-key retention with periodic garbage collection and an optional clear_on_success, and is fully manageable through the Core API and Execution API.

Task State Store

Task state is scoped to a specific task instance and persists across retries. Use it to track coordination state like remote job IDs, cursors, or progress checkpoints:

@task
def extract_data(**context):
    task_state = context["task_state_store"]
    # Resume from where we left off on retry
    cursor = task_state.get("last_cursor", default=0)
    records = fetch_records(since=cursor)
    new_cursor = records[-1]["id"]
    task_state.set("last_cursor", new_cursor)
    return records

Persisted state is visible per task instance in the UI, and you can add or edit entries (with an optional custom expiry) directly.

Asset State Store

Asset state is scoped to an asset, so any task that produces or consumes it can read and write the same state. Ideal for shared watermarks:

from datetime import datetime, timezone
from airflow.sdk import Asset, task

warehouse = Asset("s3://warehouse/orders")

@task(outlets=[warehouse])
def load_to_warehouse(**context):
    asset_state = context["asset_state_store"]
    last_loaded_at = asset_state.get("last_loaded_at", default=None)
    rows = [r for r in fetch_all_rows() if last_loaded_at is None or r["created_at"] > last_loaded_at]
    if not rows:
        return
    load(rows)
    asset_state.set("last_loaded_at", datetime.now(tz=timezone.utc).isoformat())

Asset state is browsable from the asset view, including which task last wrote each entry.

Durable / Crash-Safe Execution

The clearest demonstration of what AIP-103 enables is ResumableJobMixin, which SparkSubmitOperator now uses. If the Airflow worker dies mid-run, the retry reconnects to the existing Spark job instead of submitting a new one — this is on by default (durable=True):

SparkSubmitOperator(
    task_id="submit_spark_job",
    application="my_app.py",
    durable=True,  # survives worker failures, reconnects to the existing job on retry
)

Configuration

By default, state persists in the Airflow metadata database:

[state_store]
backend = airflow.state.metastore.MetastoreBackend
# Set to 0 to disable time-based cleanup
default_retention_days = 30
# Set True for automatic post-success cleanup
clear_on_success = False
# 64 KB cap on REST API writes; 0 = no limit
max_value_storage_bytes = 65535

For large payloads or credentialed storage, route writes through a worker-side backend. The Execution API still records a reference string in the database:

[workers]
state_store_backend = mypackage.state.CustomStateBackend

Key Capabilities

• Durable task state that survives retries and reruns via the task_state_store accessor
• Asset state carried alongside assets via asset_state_store
• Pluggable backend: metadata DB by default, or a custom worker-side backend ([workers] state_store_backend)
• Retention & GC: per-key retention with periodic garbage collection for the task state store; asset state persists indefinitely until explicitly deleted. Optional clear_on_success for task state.
• Fully managed via API: Core API + Execution API support, and new UI views for the asset and task stores (#67292)

🌐 Language Task SDK (AIP-108): Write Tasks in Java and Go

Airflow has always been Python-first for both orchestration and task logic. 3.3 keeps the orchestration in Python but lets individual task implementations be written in Java or Go and run on the standard Airflow workers you already operate. A new Coordinator layer routes a task to its language runtime — JavaCoordinator for the JVM, ExecutableCoordinator for self-contained native binaries such as Go — runs your code there, and proxies Variables, Connections, and XComs back through the Execution API. Your team’s existing Java or Go code becomes a first-class Airflow task without a Python rewrite.

⚠️ Experimental in 3.3. The SDK APIs and wire protocol may change between releases. The Java and Go SDKs ship as separate artifacts (a Maven/Gradle dependency and a Go module + coordinator), not in the apache-airflow pip package.

Author two things: a Python stub Dag + the language task

In this release the Dag shape is still declared in Python with @task.stub, while the implementation lives in your Go binary or Java jar. The dag_id and task ids must match across both so the coordinator can locate the right artifact. The queue= on each stub routes it to a coordinator.

One run of the combined Dag chains Python, Go, and Java tasks in a single graph — all three task types succeed and their native logs stream straight into the Airflow UI.

Authoring & build tooling

• Go — write tasks in a bundle package, then pack a deployable binary (executable + embedded source + manifest) with the airflow-go-pack tool: go tool airflow-go-pack ./example/bundle
• Java — apply the org.apache.airflow.sdk Gradle plugin and run the bundle task to produce a deploy-ready fat jar in build/bundle/: ./gradlew bundle

Configuration

Declare one coordinator per language runtime under [sdk] and bind each to the queues your stub tasks use:

[sdk]
# One coordinator per language runtime
coordinators = {
  "go": {"classpath": "airflow.sdk.coordinators.executable.ExecutableCoordinator", "kwargs": {"executables_root": ["~/airflow/executable-bundles"]}},
  "jdk-17": {"classpath": "airflow.sdk.coordinators.java.JavaCoordinator", "kwargs": {"jars_root": ["~/airflow/jars"]}}
}
# Map @task.stub(queue=...) to a coordinator
queue_to_coordinator = {"golang": "go", "java": "jdk-17"}

Key Capabilities

• Native Airflow model access: read and write XComs, Variables, and Connections from Go and Java through the same Execution API the Python SDK uses
• Airflow logging + remote logging: task logs stream over the coordinator into Airflow’s normal logging stack, so they show up in the UI and honor your remote logging (S3/GCS/…) setup. The SDKs bridge native loggers (Java ships System.Logger / SLF4J adapters; Go bridges log/slog)
• Truly cross-language Dag: one Dag can chain Python, Go, and Java tasks, and XCom round-trips between any of them — Python → Go/Java and back
• Task-level retries: set retries on the Python stub task (e.g. @task.stub(queue=…, retries=2)); the Go and Java SDKs honor it — a failed task with retries left moves to up_for_retry and runs again, instead of failing terminally
• Works across executors: runs on LocalExecutor, CeleryExecutor, and KubernetesExecutor (KubernetesExecutor needs some extra configuration)

🧩 Asset Partitioning: More Mappers, Windows, and Wait Policies

Building on the asset partitioning introduced in 3.2.0, Airflow 3.3 substantially expands how a single upstream asset event fans out to partitioned downstream Dag runs.

New partition mappers — FanOutMapper (one-to-many) and FixedKeyMapper + SegmentWindow (categorical rollup) — join RollupMapper, and compose with time windows (day / week / month / quarter / year) and a wait_policy (WaitForAll or MinimumCount(n)) to control exactly when partitioned runs fire.

Key Capabilities

• New mappers: FanOutMapper for one-to-many fan-out, FixedKeyMapper + SegmentWindow for categorical rollup (#67716, #66030)
• Time windows that fan out forward or backward in time (day/week/month/quarter/year)
• Wait policies: fire on WaitForAll or MinimumCount(n) of the upstream window (#66848)
• Per-mapper fan-out cap plus the global [scheduler] partition_mapper_max_downstream_keys bound (#67184)
• Forward or backward fan-out via direction=Window.Direction.FORWARD / Window.Direction.BACKWARD (default FORWARD) (#67475)

🔁 Pluggable Retry Policies (AIP-105)

Task retry behaviour is now pluggable. In addition to a fixed retries count, you can attach a custom retry policy that decides whether and when a task is retried — enabling strategies such as retrying only on specific exceptions, or backing off based on your own logic. (#65474)

🎯 Dag Results: Call a Dag and Get a Value Back

You can now designate a task as a Dag’s result with the @result decorator, then call GET /api/v2/dags/{dag_id}/dagRuns/{dag_run_id}/wait to block until the run finishes and get the value back. The endpoint streams newline-delimited JSON (NDJSON) — the run state as it progresses, and on completion the result task’s return value keyed by task id — which makes it easy to embed an Airflow Dag behind an API endpoint.

⚠️ Experimental in 3.3.0 and may change in future releases. (#64563)

📦 Dag Bundle Version Control on Clear, Rerun, and Backfill

Airflow 2.x always reran with the latest code; 3.x introduced bundle versioning that defaults to the original version. 3.3 adds per-run control over whether a cleared, rerun, or backfilled Dag run uses the latest bundle version or the original, resolved by precedence: an explicit request parameter run_on_latest_version (exposed as a CLI flag for backfill only) → the Dag-level rerun_with_latest_version → [core] rerun_with_latest_version → finally False for clear/rerun and True for backfills (preserving historical behaviour). (#63884)

🖥️ UI Enhancements

• Asset & task store views: Inspect persisted task and asset state directly in the UI (#67292), including a column linking to the task instance that wrote each entry (#68395) and a custom expiration picker for the task store (#68394)
• awaiting_input HITL state: Human-in-the-Loop tasks awaiting input now run off the triggerer with a dedicated state (#68028)
• Bulk actions: Bulk-mark Dag runs as success/failed (#68278, #67948) and bulk-clear task instance selections (#68029)
• Details tab for mapped task instances (#68340) and additional task-instance attributes in the details section (#68378)
• Team name in the asset graph view for multi-team deployments (#68457)
• Full-screen toggle in the code viewer (#68044)

🔭 Observability & Operations

• OpenTelemetry Histograms: Timer and timing metrics are now recorded as Histograms instead of Gauges, preserving count, sum, and bucket distribution across recordings. ⚠️ Breaking for existing dashboards/alerts — queries built on the old Gauge series need updating. (#64207)
• Tagged Dag-processing metric: dag_processing.last_run.seconds_ago is now emitted with file_path, bundle_name, and file_name tags instead of baking the file name into the metric path. ⚠️ Breaking — dashboards/alerts parsing the old …seconds_ago.{dag_file} path must switch to the tagged form. (#62487)
• New Deadlines page: A Deadlines page under the Browse menu, available to any role with can_read + menu_access on Dag Runs (#67586)
• Remote logging resolution refactor: Remote task-log handler resolution now lives in the shared airflow_shared.logging.factory with a single, well-defined precedence, lazy resolution on first use, and a from_config() contract for provider RemoteLogIO classes (#67056)

⬆️ Upgrading to 3.3.0

Before upgrading, review the Significant Changes in the release notes. In particular, the OpenTelemetry metrics changes above are breaking for existing dashboards and alerts, so plan to update any queries built on the previous Gauge series or the old dag_processing.last_run.seconds_ago.{dag_file} metric path.

🙏 Community Appreciation

This release represents the collaborative effort of hundreds of contributors from around the world. Special thanks to our release manager and all the developers, documentarians, testers, and community members who made Airflow 3.3.0 possible.

Thanks to contributors like you, the Airflow project continues to thrive. Whether you’re filing issues, submitting PRs, improving documentation, or helping others in the community, every contribution matters.

🔗 Get Involved

• Try the Release: Upgrade your development environment and explore the new features
• Join the Conversation: Connect with us on Slack and the dev mailing list
• Contribute: Check out our contribution guide
• Provide Feedback: Share your experiences and suggestions on GitHub

Apache Airflow 3.3.0 gives your tasks memory and multi-language support. We can’t wait to see what you build with it!

📦 PyPI: https://pypi.org/project/apache-airflow/3.3.0/
📚 Docs: https://airflow.apache.org/docs/apache-airflow/3.3.0/
🛠 Release Notes: https://airflow.apache.org/docs/apache-airflow/3.3.0/release_notes.html
🐳 Docker Image: "docker pull apache/airflow:3.3.0"
🚏 Constraints: https://github.com/apache/airflow/tree/constraints-3.3.0

Significant Changes

Asset Partitioning (#64571, #65447, #66030, #66848, #67184, #67475, #67716, #68978)

Building on the asset partitioning introduced in 3.2.0, Airflow 3.3.0 substantially expands how a single upstream asset event fans out to partitioned downstream Dag runs. New partition mappers —
RollupMapper (many-to-one), FanOutMapper (one-to-many), and FixedKeyMapper +
SegmentWindow (categorical rollup) — compose with time windows (day/week/month/quarter/year) and a wait_policy (WaitForAll or MinimumCount(n)) to control when partitioned runs fire. Windows can fan out forward or backward in time, and total fan-out per upstream event is bounded by the new [scheduler] partition_mapper_max_downstream_keys config (configurable per mapper). Airflow 3.3.0 also adds the PartitionedAtRuntime timetable, which lets a Dag declare that its partition key(s) are assigned when the run starts rather than mapped from an upstream event.

For detailed usage instructions, see :doc:/authoring-and-scheduling/assets.

Task and Asset State Store (#65759, #66073, #66160, #66463, #66586, #66859, #67041, #67292, #67319)

Airflow 3.3.0 introduces a first-class state store for tasks and assets (AIP-103). Tasks can persist arbitrary key-value state that survives across retries and runs via a new task_state_store accessor, and assets can carry their own state via asset_state_store — both available from the Task SDK. State is kept in the metadata database by default, or in a custom worker-side backend ([workers] state_store_backend), supports per-key retention with periodic garbage collection and an optional clear_on_success, and is fully manageable through the Core API and Execution API.

For detailed usage instructions, see :doc:/core-concepts/task-and-asset-state-store.

Pluggable Retry Policies (#65474)

Task retry behaviour is now pluggable (AIP-105). In addition to a fixed retries count, you can attach a custom retry policy that decides whether and when a task is retried, enabling strategies such as retrying only on specific exceptions or backing off based on custom logic.

For detailed usage instructions, see :ref:concepts:retry-policies.

Language Task SDK (Java and Go) (#65958, #67161, #67635, #67699)

Airflow 3.3.0 adds a Coordinator layer (AIP-108) that lets individual task implementations be written in non-Python languages while the Dag and its scheduling stay in Python. A task is declared in the Dag with @task.stub(queue=...); the worker routes it to a configured coordinator (JavaCoordinator for JVM languages, ExecutableCoordinator for self-contained native binaries such as Go) that runs the task in a language runtime and proxies Variables, Connections, and XComs back through the Execution API.

.. warning::

The Coordinator layer and the Java/Go SDKs are experimental in 3.3.0 and may change in future versions based on user feedback.

For detailed usage instructions, see :doc:/authoring-and-scheduling/language-sdks/index.

Dag bundle version on clear, rerun, and backfill (#63884)

The new rerun_with_latest_version setting controls whether a cleared, rerun, or backfilled Dag run uses the latest bundle version or the original version from the initial run. The default is resolved by precedence: an explicit request parameter/CLI flag, then the Dag-level rerun_with_latest_version, then [core] rerun_with_latest_version, and finally False for clear/rerun and True for backfills (preserving historical behaviour). Airflow 2.x always reran with the latest code; 3.x introduced bundle versioning defaulting to the original version, and this setting gives users control.

See :doc:/administration-and-deployment/dag-bundles for full details.

Provider example Dags as dedicated bundles (#66161)

Example Dags shipped by provider distributions are now discovered via ProvidersManager and registered as their own Dag bundles — one per provider, named apache-airflow-providers-<distribution>-example-dags (or <distribution>-example-dags for third-party providers). The [core] load_examples option still gates whether they are registered.

REST API clients that filtered bundle_name by "dags-folder" for provider-shipped example Dags must update to the new per-provider bundle names; Dag identifiers are unchanged.

Remote logging resolution decoupled from airflow.logging_config (#67056)

Remote task log handler resolution is now owned by the shared airflow_shared.logging.factory module and applies a single, well-defined precedence:
a user-defined [logging] logging_config_class exporting REMOTE_TASK_LOG /
DEFAULT_REMOTE_CONN_ID (existing custom configs keep working);
ProvidersManager scheme dispatch — the scheme of [logging] remote_base_log_folder selects a provider RemoteLogIO class, instantiated via a no-argument from_config() classmethod;
a transitional legacy fallback reading airflow_local_settings.py (to be removed in Airflow 4.0).

airflow.logging_config.load_logging_config is deprecated (it now emits DeprecationWarning and delegates to new private helpers), and configure_logging no longer eagerly resolves the remote handler — resolution is lazy on first use. Providers that registered a remote-logging: block but do not implement from_config are skipped with a warning and fall through to the legacy path.

Migration: replace direct calls to airflow.logging_config.load_logging_config() with the new helpers, and have provider remote-log handler classes implement a no-argument from_config classmethod that reads airflow.providers.common.compat.sdk.conf.

OpenTelemetry timer metrics now use Histogram (#64207)

OpenTelemetry timer and timing metrics are now recorded as Histograms instead of Gauges, preserving count, sum, and bucket distribution across recordings.

Dag-processing "seconds ago" metric is now tagged (#62487)

dag_processing.last_run.seconds_ago.{dag_file} is now a legacy metric. The new dag_processing.last_run.seconds_ago is emitted with file_path, bundle_name and file_name tags (file_path + bundle_name uniquely identify the Dag file). The legacy metric is still emitted by default and can be disabled via [metrics] legacy_names_on.

New Deadlines page under Browse (#67586)

A new Deadlines page is available under the Browse menu, accessible to any role that already has can_read and menu_access on Dag Runs.

New Features

• Add partition clear support to the REST API matching the CLI, with a clearPartitions endpoint and partition_key/partition_date window selectors on clearDagRuns (#68702)
• Add [core] mp_start_method and [core] mp_forkserver_preload configuration options (which can be overridden per [scheduler]/[triggerer]/[dag_processor]) to control the multiprocessing start method (#68875)
• Add a durable toggle to ResumableJobMixin to opt out of resumable execution (#68623)
• Add a @result decorator to mark a TaskFlow task as the Dag's result task (#64563)
• Add [triggerer] shared_stream_cohort_grace_period to reduce missed events on triggerer restart (#68888)
• Propagate partition_date from producer Dag runs to consumers of partitioned assets (#67285)
• Make the task and asset state store accessible from triggers via AssetStateStoreAccessors (#67839)
• Add OpenTelemetry head sampling support (#68591)
• Add async XCom accessors for async tasks (#68299)
• Add an async aget_hook method to BaseHook for async tasks (#68506)
• Allow custom partition Window subclasses via a plugin registry (#68717)
• Support an extra field for the Coordinator (#68694)
• Apply rerun_with_latest_version to TriggerDagRunOperator reruns (#67273)
• Scope the XCom Execution API to teams in multi-team mode (#68850)
• Enforce pool team ownership in the scheduling loop (#68649)
• UI: Add team name to the asset graph view (#68457)
• Populate partition_date for partitioned Dag runs whose composite asset key has a single time-based dimension (#68442)
• UI: Add a column to the asset store table linking to the task instance that wrote it (#68395)
• UI: Add a custom expiration datetime picker for the task store modal (#68394)
• UI: Add additional task instance attributes to the task instance details section (#68378)
• UI: Add a Details tab to the mapped task instance view (#68340)
• UI: Add bulk marking of Dag runs as success or failed from multi-select (#68278)
• Add --team-name support to the pool CLI commands (#68110)
• UI: Add a full-screen toggle to the code viewer (#68044)
• Add API endpoint support for consumer team asset filtering (#68034)
• UI: Add bulk clear selection for task instances (#68029)
• Add awaiting_input task state for Human-in-the-Loop, running off the triggerer (#68028)
• Add bulk API to mark Dag runs as success or failed (#67948)
• Record writer info for every asset store write for better cross-linkage (#67902)
• Register XCom output_type classes from a worker-side Dag walk (#67875)
• Add FixedKeyMapper and SegmentWindow for categorical asset-partition rollup (#67716)
• Add a bulk POST /dags/{dag_id}/clearDagRuns API endpoint (#67709)
• Return Pydantic model instances through XCom for structured output (#67644)
• Add the ability to apply a note when clearing a Dag run or task instances (#67639)
• Add consumer_teams to AssetAccessControl in the Task SDK (#67625)
• Populate trigger team_name at creation time for multi-team support (#67605)
• UI: Add bulk Clear on the Dag Runs list page (#67564)
• Add multi-team query filtering to triggerer trigger assignment (#67517)
• Add forward fan-out support via the forward kwarg on Window (AIP-76) (#67475)
• Add patch task state API and expires_at support in the set API (AIP-103) (#67319)
• Add a team_name column to the trigger table for multi-team triggerer support (#67305)
• UI: Add asset and task store views (#67292)
• Add a --team-name CLI argument to the triggerer for multi-team (#67254)
• Add an allow_global option to asset access control (#67251)
• Add mTLS and private CA support to the API client and server (#67214)
• Add Markdown documentation support for TaskGroups (#67207)
• Add a per-mapper max_fan_out override for partition fan-out cap (#67184)
• Add timezone support to the SDK temporal partition mappers (#67164)
• Add ResumableJobMixin with SparkSubmitOperator for surviving worker failures (#67118)
• Add bulk delete for Dag runs (#67095)
• Add a nav_top_level option for plugin nav items (#67084)
• Add Core API endpoints for task state and asset state (AIP-103) (#67041)
• Replace allow_producer_teams with access_control on Asset (#66954)
• Add worker-side custom state backend support (AIP-103) (#66859)
• Let partitioned Dag runs fire on a partial upstream window with wait_policy (#66848)
• Consume task-emitted partition keys on asset events (AIP-76) (#66782)
• Add per-task state key retention from operators (AIP-103) (#66699)
• Add a callback_execution_timeout config for deadline callbacks (#66609)
• Add a clear_on_success config to wipe task state on success (AIP-103) (#66586)
• Add a partitions clear CLI command to reset DagRun partition fields (#66520)
• Make CORS allow_credentials configurable (#66503)
• Add periodic task state garbage collection and retention support (AIP-103) (#66463)
• Add URI sanitizers and asset factories for new schemes (#66426)
• Add a teams sync CLI command (#66418)
• Add remote log upload support for callback subprocesses (#66379)
• Add by-name/by-uri asset state routes and AssetUriRef support (AIP-103) (#66336)
• UI: Add support for rendering multi-type params (#66278)
• Filter Dags by teams when registering asset changes (#66168)
• Wire up Task SDK communication and context access for task/asset state (AIP-103) (#66160)
• UI: Add marking a task group as success or failed (#66146)
• Add Execution API endpoints for task and asset states (AIP-103) (#66073)
• Add FanOutMapper for one-to-many partition fan-out (#66030)
• Add Variable.keys() to list variable keys by prefix in the Task SDK (#66022)
• Add an airflow dags clear command for partition-range reprocessing (#66004)
• Add a memray_detailed_tracing option for deeper memory profiling (#65996)
• UI: Add support for different graph directions in the asset graph view (#65948)
• UI: Add a Clear All Mapped Tasks button (#65813)
• Add allow_producer_teams to the Asset SDK class (#65790)
• UI: Show expected duration based on historical average in Dag Run details (#65722)
• Add team name to the task context (#65617)
• Add an on_kill() hook to BaseTrigger to handle user actions on triggers (#65590)
• Allow accessing a Dag's members via [] (#65586)
• Add pluggable retry policies for Airflow tasks (AIP-105) (#65474)
• Add support for format="Duration" in params (#65469)
• UI: Add pagination to the grid view (#65388)
• Add partition_key to the task context (#65359)
• Make the blocked-thread warning threshold configurable (#65009)
• Add name fields to SDK deadline alerts (#64926)
• Add dynamic interval resolution support via Variables for deadline alerts (#64751)
• Add an is_backfillable property to Dag API responses (#64644)
• Return dag-specified results in the dag run wait API (#64577)
• Hold a Dag run until all upstream partitions arrive (AIP-76) (#64571)
• Add a way to mark a return-value XCom as the dag result (#64522)
• Allow accessing a TaskGroup's members via [] (#64430)
• UI: Redo the Gantt chart (#64335)
• UI: Add task-level filters to the Dag graph tab (#64271)
• UI: Add bulk Clear, Mark Success/Fail, and delete for multiple task instances (#64141)
• Check that multi-team is enabled when a team name is provided to the API (#63994)
• Add a DagRunType for operators (#63733)
• UI: Add search functionality to the task log viewer (#63467)
• Add patching of task group instances in the API (#62812)
• Add deadlines API endpoints (#62583)
• Add async connection testing via workers for security isolation (#62343)
• Add run_after to TriggerDagRunOperator (#62259)
• UI: Display deadlines on the Dag Run and Overview tabs (#62195)
• Re-enable the start_from_trigger feature with template-field rendering (#55068)
• Backfill partitioned Dags by partition-date range (#67537)
• Add a producer-side acknowledgement channel to shared-stream triggers (#67523)
• UI: Add partition_date to the Dag run detail page (#68977)
• Expose the upstream partition_key on triggering_asset_events and dag_run.consumed_asset_events (AIP-76) (#69120)
• Allow get/set/delete/clear of AssetStateStoreAccessor to run on the triggerer (#68966)

Bug Fixes

• Fix KubernetesExecutor scheduler crash caused by a pod_override that cannot be pickled when running in-cluster (#68831)
• UI: Fix dashboard alert clamping and collapse controls (#68893)
• Stabilize mapped-task XCom result ordering in the Dag run wait endpoint by ordering on task_id/map_index (#68550)
• Only log task state cleanup when a worker state store backend is configured (#68878)
• Fix in-process Execution API loop stopped while transport still in use (#68865)
• Fix task state store custom expiry datetime missing timezone on save (#68823)
• Do not leak threads from InProcessExecutionAPI (#68840)
• Fix partitioned backfill widening a sub-day window to the whole day (#68718)
• UI: Fix inconsistent padding between Dag Runs and Task Instances list views (#68689)
• Skip asset-change registration for tasks with no outlets (#68687)
• Percent-encode API client path params for keys with slashes (#68667)
• Fix bulk create+overwrite silently resetting unset fields on pools and connections (#68645)
• Fix triggerer crash when a trigger subclass does not call super().init() (#68636)
• Fix Task SDK swallowing errors when Variable.set() or Variable.delete() fails (#68542)
• Populate partition_date when manually triggering partitioned Dags (#68458)
• Improve warning visibility for invalid JSON when editing variables (#68268)
• Fix the triggerer log server port configuration key (#67785)
• Enforce ti:self scope on /execution/task-reschedules/{ti}/start_date (#67628)
• UI: Fix misleading Calendar "Total Runs" coloring behavior (#67595)
• Fix Stats not being initialized in the API server lifespan (#68514)
• Fix BackfillDagRun.partition_key type annotation (#68432)
• Fix backward compatibility for DagRunInfo partition fields (#68342)
• Fix airflow db clean failing on foreign-key-referenced dag_version rows (#68339)
• Fix 500 error when listing event logs with a NULL timestamp (#68338)
• Fix MySQL downgrade from 3.3.0 for the deadline_alert.interval JSON conversion (#68337)
• Fix older and custom secrets backends breaking on Airflow 3.2 (#68302)
• Fix secrets backend connection errors being silently swallowed at DEBUG level (#68301)
• UI: Fix the instance name title shown on non-Dag pages (#68288)
• Fix scheduler not populating partition_date for temporal asset partitions (#68266)
• UI: Fix wrong language being auto-detected from browser preferences (#68258)
• Honor retry_policy on non-deferrable TriggerDagRunOperator wait failures (#68254)
• Fix scheduler crash loop when the last task instance predates Dag versioning (#68253)
• Fix team consumer asset filtering (#68242)
• UI: Fix sluggish multi-selection behavior in tables (#68229)
• UI: Fix mapped task instance links for tasks without a start date (#68194)
• Fix setup/teardown auto-inclusion when clearing or marking tasks (#68193)
• UI: Remove redundant columns from the XCom panel on the task instance page (#68188)
• UI: Fix Gantt tooltip showing the wrong start date on queued/scheduled segments (#68176)
• Fix Java SDK coordinator rejecting IPv4-mapped IPv6 connections (#68169)
• Fix Java SDK tasks being rejected by the coordinator connection-ownership check (#68147)
• UI: Fix language key for the Dag bundle filter (#68131)
• Fix DagFileProcessorManager silent hang on database lock contention (#68118)
• Mask all connection extra and variable values in the API audit log (#68049)
• Fix spurious "Failed to detach context" error on Execution API disconnects (#68039)
• UI: Fix Dag code highlighting for triple-quoted and escaped-brace f-strings (#68026)
• Fix cursor encoding for column-form sort parameters in the REST API (#67973)
• Fix SimpleAuthManager not preserving the deep-link next URL on first login (#67965)
• Guard the task stats emission to prevent errors (#67955)
• UI: Fix task instance state badge staying stale after a Mark-as action (#67950)
• Register nested Pydantic models for XCom deserialization (#67932)
• Fix example_asset_store consumer crash (#67922)
• Raise InvalidJwtError in JWTValidator.avalidated_claims() when the key ID does not match (#67909)
• Fix Kubernetes executor pod_override being stringified without the cncf provider (#67895)
• UI: Prevent duplicate task instance summary stream refreshes after mutations (#67892)
• Reject negative default_retention_days in the Task SDK and core API routes (#67890)
• Fix none_failed_min_one_success trigger rule checks (#67873)
• Remove trigger kwargs from the REST API response (#67868)
• UI: Fix long parameter names overflowing the Trigger Dag modal (#67859)
• Fix misleading log message in the task runner clear-on-success block (#67836)
• Fix scheduler crash when logging orphaned task resets (#67822)
• UI: Fix dashboard pool summary showing incorrect deferred slot usage (#67818)
• Fix trigger datetime deserialization (#67795)
• UI: Fix Graph layout for TaskGroup tasks wired to external nodes (#67720)
• Fix airflow dags clear clearing the wrong day for non-UTC partitioned timetables (#67717)
• Fix per-index evaluation of ONE_FAILED in mapped task groups (#67684)
• UI: Fix dialog dismissal for the Chakra upgrade (#67674)
• UI: Hide dashboard metric percentages when a state count is capped (#67664)
• Apply per-file authorization to the dag-source endpoint (#67662)
• Fix airflow dags next-execution --table crash when no next run exists (#67642)
• UI: Fix the time picker omitting seconds (#67636)
• Filter scheduling-dependencies graph edges by readable-Dag access (#67627)
• Mask per-key secrets-backend-kwarg overrides on the Config API (#67622)
• Fix GET /auth/login missing a 400 response in the OpenAPI spec (#67571)
• Fix GET /pools incorrectly documenting a 404 response in the OpenAPI spec (#67570)
• Add a compatibility layer for import errors caused by AirflowSecretsBackendAccessDenied (#67560)
• UI: Fix rendering of None child state (#67552)
• Fix sort order for mapped task instances (#67551)
• Fix import errors total-entries count with multiple Dags per file (#67550)
• UI: Prefer active over queued state for collapsed groups (#67543)
• Fix callback state not updating from executor events due to a UUID type mismatch (#67542)
• Reject wildcard origin in CORS config instead of toggling credentials (#67502)
• Guard the finally-block logger in the HTTP access log middleware (#67501)
• Strip CR/LF from user-supplied logical date before logging (#67500)
• Redact secret-looking query parameters in the HTTP access log (#67498)
• UI: Fix Calendar view to respect the user-selected timezone (#67497)
• Escape LIKE wildcards in non-search filter parameters (#67496)
• Fix missing redaction of secret values in variable JSON (#67495)
• Fix bulk CREATE+OVERWRITE team-context authorization bypass (#67493)
• UI: Return 400 instead of 500 from structure_data on a malformed asset expression (#67489)
• Fix SimpleAuthManager redirect to the next URL after login (#67483)
• Return 400 instead of 500 from materialize_asset on invalid input (#67445)
• UI: Restore the Monaco find widget in the Dag Code view (#67391)
• UI: Fix an HTTPException import that turned a 400 into a 500 in the dags endpoint (#67363)
• Restore fail_fast handling when reschedule exceeds the MySQL TIMESTAMP limit (#67353)
• Fix the Triggered Dag button not being visible during queued/running state (#67327)
• Fix variables import with structured falsy values (#67060)
• Avoid logging Execution API bearer credentials (#67059)
• Sanitize Dag processor metric file names (#67029)
• Return a 422 when the database rejects an API payload (#66888)
• Prevent AlreadyRunningBackfill error caused by an invalid date range request (#66874)
• Restrict owner-link and extra-link href values to safe schemes (http, https, mailto, relative) (#66741)
• Add a session parameter to the BaseStateBackend interface to fix custom backends (#66708)
• Allow deadline callbacks within the same Dag module (#66702)
• UI: Fix relative React plugin bundle URLs in dev mode (#66618)
• Validate Dag trigger conf as a JSON object or null (#66617)
• Require a trust sentinel for state.user injection in get_user() (#66562)
• Use hmac.compare_digest for SimpleAuthManager password comparison (CWE-208) (#66556)
• Set SameSite=Lax on the SimpleAuthManager all-admins login cookie (#66502)
• Reserve /auth and /pluginsv2 from plugin URL prefixes (#66501)
• Use a cryptographically secure RNG for SimpleAuthManager passwords (#66500)
• Fix Triggerer runner_health_check_threshold log formatting (#66486)
• Fix Dag processor callback cleanup for versioned bundle files (#66484)
• Default AIRFLOW_UID to 50000 in the airflow-init chown lines (#66481)
• Strip CR/LF from MySQL URL query values before forwarding to my.cnf (#66325)
• Fix CronMixin not resolving cron presets before validation (#66102)
• Fix AirflowSDKConfigParser missing the mask_secrets method (#66077)
• Fix resolve_xcom_backend to rely on the config schema default (#65938)
• Fix a missing import cast error in the dag_run API route (#65748)
• Mask Dag processor connection and variable responses (#65704)
• UI: Show import error for deactivated Dags (#65687)
• Forward MySQL SSL params from sql_alchemy_conn to airflow db shell (#65575)
• Disable SQLite FK checks in the 0111 migration downgrade (#65545)
• Handle Variable values that cannot be decrypted gracefully in the stable REST API (#65452)
• Retry TriggerDagRunOperator when the triggered DagRun fails (#65390)
• Fix task run exceptions never being caught by Sentry (#65161)
• Fix the bulk task instance authorization error message rendering (#64719)
• Fix trigger template rendering failure when operator template_fields differ from trigger attributes (#64715)
• Fix task_defer with non-JSON next_kwargs in TaskInstance (#64714)
• Add the error as context["exception"] in InProcessTestSupervisor (#64568)
• Fix NPM security alerts in the simple auth manager (#64309)
• Fix Dag run trigger to surface errors instead of swallowing them (#64130)
• Fix Task SDK Connection extras built from a URI constructor (#64120)
• Add insert/update-on-conflict for rendered task instance fields (#63874)
• Fix timeout_with_traceback crashes on Windows and non-main threads (#63664)
• UI: Wrap long lines in the rendered templates view (#63492)
• Block path traversal via ".." in dag_id and run_id (#63296)
• Fix the scheduler health check command in docker-compose.yaml (#62280)
• Fix unmapped task deadlock when upstream tasks are removed (#62034)
• Forward termination signals from the supervisor to the task subprocess (#61627)
• Check destination team permission when using bulk APIs for connections, variables, and pools (#68573)
• Fix the execution API /health check failing on the empty-path route (#68578)
• Fix Dag run partition key filter breaking on composite keys containing | (#68459)
• Fix the partition clear date range for non-UTC partitioned timetables (#68460)
• Validate that partition keys are non-empty and within the column length (#68443)
• Fix the scheduler serving stale Dag code after an in-place serialized Dag version update (#68558)
• Determine the latest Dag version by version number to avoid collisions when timestamps tie (#68389)
• Fix new runs and reruns executing an outdated bundle version when the Dag serialization is unchanged (#68336)
• Fix remote logging from the task supervisor (#68370)
• Upload task logs even when the final state update fails (#67935)
• Escape URLs in the Task SDK client when looking up Dag operations (#68129)
• Fix task scheduling when multi-team is enabled (#68634)
• Fix secret values not being masked in rendered templates when keys use dot or dash separators (#68624)
• Fix jwt_audience for the public API being read from two different config sections (#67494)
• Fix duplicate deadline-miss callbacks firing from multiple HA scheduler replicas (#64737)
• Fix scheduler crash on non-ASCII Dag names when OpenTelemetry metrics are enabled (#68023)
• Fix Dag processor crash on non-ASCII names in OpenTelemetry gauge and timer metrics (#68284)
• Report duplicate plugin names as import errors instead of silently ignoring them (#66649)
• Require edit permission for async connection tests that update an existing connection (#68127)
• Restore the deprecated [core] execution_api_server_url mapping to [workers] execution_api_server_url (#63949)
• Fix dag.test() not re-syncing sibling Dags across repeated calls (#66205)
• UI: Invalidate per-attempt task instance caches after actions so logs and details are not stale (#67212)
• Fix task runner failure on a duplicate task instance success-state update (#63355)
• Fix a race condition on the order_by parameter when listing Dag runs via the REST API (#68948)
• Exclude non-successful Dag runs from the DeadlineReference.AVERAGE_RUNTIME deadline calculation so failed runs no longer skew the computed deadline (#68949)
• Allow InProcessExecutionAPI to start without api_auth.jwt_secret configured (#68982)
• Make airflow dags test wait for Human-in-the-loop input instead of looping indefinitely on parked HITL tasks (#69104)
• Fix the Java coordinator rejecting macOS dual-stack loopback connections (#68973)
• Fix an asset-event ingestion crash for Dags using FixedKeyMapper (#69326)
• UI: Fix the details panel header overlapping the tabs (#69318)
• Fix new Dag versions being created when a task's retry_policy was serialized (#69315)
• Fix retry-policy overrides not being persisted to task-instance history (#69241)
• Fix deadline callback data not being persisted (#69259)

Miscellaneous

• Propagate the resolved task log level and [logging] namespace_levels to language SDK runtimes (#68712)
• Forward run-identity attributes (dag_id, run_id, run_type) to the trace sampler so a custom head sampler can differentiate by run kind (#68592)
• Remove all_map_indices from task_state_store.clear() in the task context (#68880)
• Optimize the dag processor by caching bundle-to-team name lookups (#68730)
• Rename the misleading last_automated_run param to reference_run (#68714)
• Add a team_name tag to the remaining multi-team metrics (#68601)
• Add a team_name tag to dag processor metrics for multi-team deployments (#68599)
• Add a team_name tag to asset metrics for multi-team deployments (#68367)
• UI: Persist dashboard alert collapse state and clamp long alerts (#68329)
• Optimize bulk variable deletion to avoid N+1 queries (#68508)
• UI: Unify the Dag Code tab toolbar styling with the Logs toolbar (#68449)
• Optimize bulk Dag run authorization to avoid N+1 team-name queries (#68286)
• Improve airflow dags command to use bulk clear (#68280)
• Add the task_state_store table to the airflow db clean mechanism (#68218)
• Add metrics and traces to ResumableJobMixin for crash recovery (#68213)
• Improve ResumableJobMixin crash-recovery observability with better logging (#68206)
• Pass DagRun to task_instance_mutation_hook for run-aware task mutation (#68198)
• Add team_name to multi-team metrics (#68108)
• Reduce redundant Dag team lookups in authorization checks (#68020)
• Enhance ResumableJobMixin.get_job_status with context for better job status tracking (#68009)
• Propagate OpenTelemetry trace headers from the client to Execution API server-side spans (#67904)
• Widen the type hint for the DagRun.get_task_instances / fetch_task_instances state parameter (#67880)
• UI: Use the bulk clear Dag runs endpoint for bulk Dag run clear (#67846)
• Add a default parameter to the task and asset state get() method (#67842)
• Make core API routes for task and asset states interact only with the database (#67835)
• Optimize Dag processor file-queue deduplication from O(N^2) to O(N) (#67750)
• Add allow_consumer_teams and allow_global_consumers columns to TaskOutletAssetReference (#67730)
• Speed up the Dags list and dashboard queries on large DagRun tables (#67721)
• Make partition_key provenance-only and inherit it onto asset events (#67718)
• Speed up Dag serialization by skipping a redundant asset roundtrip (#67702)
• Cache BaseOperator.init signature in operator serialization (#67701)
• Optimize TaskGroup.topological_sort for reverse-declared Dags (#67688)
• Update serialization for producer-side asset access control (#67658)
• Allow outlets to be added and accessed in AssetStateAccessor (#67619)
• Unify task/asset state storage between the Core API and Execution API (#67547)
• Decorate custom state references with an envelope for UI clarity (#67530)
• Simplify authoring of task and asset states by allowing JSON types (#67418)
• Replace Sphinx Redoc with Swagger for the API docs (#67390)
• Emit OpenTelemetry spans around listener hook calls (#67347)
• UI: Update verbiage for lower-priority backfill runs (#67338)
• Fix N+1 query in the bulk task instance delete endpoint (#67304)
• Speed up TaskGroup.topological_sort with an int-indexed projected sweep (#67288)
• UI: Use react-query native error state for bulk action hooks (#67284)
• Wrap executor.heartbeat() in a timer to localize scheduler loop slowdowns (#66808)
• Emit dagrun.first_task_start_delay separately from scheduling delay (#66807)
• Share one poll loop across sibling event triggers (#66584)
• UI: Upgrade icons, spacing, and default component themes (#66569)
• Warn when SimpleAuthManager runs in a production-shaped deployment (#66563)
• Migrate Stackdriver logging config to the RemoteLogIO pattern (#66513)
• Add a BundleVersion dataclass and version_data persistence to DagVersion (#66491)
• Avoid lazy-loading timetable fields for latest DagRuns (#66488)
• Move allow_producer_teams to DagScheduleAssetReference (#66487)
• Pass user teams to the create_asset_event endpoint (#66367)
• Load USFederalHolidayCalendar lazily to reduce memory usage when loading examples (#66303)
• Propagate task OpenTelemetry trace context through IPC into Execution API requests (#66151)
• Surface worker Dag parse duration in the task log (#66138)
• Skip deserializing trigger_kwargs when loading serialized Dags (#66002)
• Honor AUTH_ROLE_PUBLIC in the FastAPI API server (#65685)
• Add extended sysinfo for the Edge worker (#65472)
• Clarify logs when a Dag is being processed in the Dag processor (#65196)
• Add indexes on task_instance.dag_version_id and dag_run.created_dag_version_id (#64818)
• Improve creation of RuntimeTaskInstance in TriggerRunner for start_for_trigger functionality (#64298)
• Mark the Triggerer supervisor as a server context so it can read metastore connections (#64022)
• Load hook metadata from YAML without importing the hook class (#63826)
• Add detailed task spans (#63568)
• Downgrade logging on query JSON parsing and add a JSON load condition (#62044)
• UI: Add a Deadlines section with a time-range selector to the Dashboard page (#68038)
• UI: Add a modal for editing notes with Markdown support (#68362)
• UI: Improve the Human-In-The-Loop form UX (#68397)
• Add a team_name tag to executor metrics for multi-team deployments (#68593)
• Make task and asset state store row size limits configurable (#68133)
• UI: Add notification UX for Human-In-The-Loop actions (#68346)
• Allow synchronous deadline callbacks (SyncCallback) to access Connections and Variables (#65269)
• Add a team_name tag to deadline metrics for multi-team deployments (#68589)
• Add a team_name tag to scheduler metrics for multi-team deployments (#68594)
• Defer the Cadwyn import so FastAPI/Starlette stay off the Task SDK worker path, reducing per-worker memory (#69029)

Doc Only Changes

• Complete the Taiwanese Mandarin (zh-TW) translation (#68870)
• Add missing Korean (ko) translations (#68600)
• Close German (de) translation gaps (#68356)
• Add a segment fan-out example to the asset partition example Dag (#68722)
• Fix runtime-partition example Dags using unreachable schedules (#68719)
• Add an example Dag for the task state store with mapped tasks (#68670)
• Fix the gap in the Taiwanese Mandarin (zh-TW) translation (#68668)
• Add wait-policy examples to the asset partition example Dag (#68658)
• Add a contributing guide for language SDKs (#68330)
• Add sdk.TIRunContext documentation for the Go SDK (#68319)
• Add a Go Task SDK authoring guide to the docs (#68223)
• Update supported-versions doc to mark 2.11.2 as EOL (#68212)
• Add documentation for ResumableJobMixin and resumable tasks (#68136)
• Add CLI examples for team-scoped pools (#68111)
• Add docs for multi-team triggerer support (#67608)
• Clarify trigger rule behavior for the removed upstream state (#67452)
• Fix outdated image links in dags.rst (#67357)
• Add an example and docs for runtime asset partitioning (AIP-76) (#67307)
• Add documentation for the Task and Asset Store (AIP-103) (#67299)
• Add a dynamic task mapping no-op example (#67022)
• Add documentation about adding access_control to the Asset object (#66949)
• Add a how-to for Dag-level retry via on_failure_callback (#66277)
• Fix documentation after PR 62645 (#65843)
• Add documentation for team-based asset event filtering (#65690)
• Document on_kill()/cleanup() for triggers (#65671)
• Explain xcom_pull behaviour without task_ids in the docs (#65406)
• Improve standalone authentication documentation for Airflow 3.x (#65330)
• Clarify manual Dag run data interval semantics in Airflow 3 (#64740)
• Document and test xcom_pull run_id usage for triggered Dag runs (#63030)
• Update params in the backfill documentation (#61821)
• Document the apache-airflow-mypy package in the core docs (#68561)
• Fix typos and formatting in the Fundamentals documentation (#68524)
• Complete the Hindi (hi) UI translation (#68574)
• Fill the Taiwanese Mandarin (zh-TW) UI translation gap (#68563)
• Document that Dag bundle kwargs should reference a Connection rather than inline credentials (#69105)
• Add example plugins and expand the asset-partitions documentation (#69017)
• Java SDK docs: JUL setup, pinning java_executable, and a config-reload note (#69020)
• Correct the example config for the coordinators (#68940)