Activity

1. potiuk

   added

   kind:metaHigh-level information important to the communityHigh-level information important to the community

   testing statusStatus of testing releasesStatus of testing releases

   on Oct 8, 2026
2. jingi723 commented on Oct 8, 2026

   

   jingi723

   on Oct 8, 2026

   Last edited by jingi723

   Contributor

   More actions

   Verified my changes in apache-airflow-providers-amazon==9.38.0rc1 for #73756 and #73881 using the PyPI wheel, with Airflow 3.3.2 / Python 3.12.14 on Linux ARM64. Updated with live AWS verification in ap-northeast-2.

   • Keep S3 Dag bundle downloads within the configured directory #73756: All 8 checks pass against a real S3 bucket: bundle initialization and repeated refresh with directory prefixes with/without trailing slashes, sibling-prefix objects, and objects whose key equals the configured prefix. Verified downloaded contents, stale-file/directory cleanup, and preservation of excluded remote objects.
   • Handle plain Step Functions execution errors #73881: All 9 checks pass using actual Standard Step Functions executions and DescribeExecution responses. Plain, Unicode, and malformed-JSON error strings are returned correctly; JSON-valued errors and successful output retain their types. In the live empty-string ErrorPath case, AWS omitted the error field and the operator correctly returned None.

   The 17 live-AWS checks all passed. Running the same checks with provider 9.37.0, using the same bucket and executions, reproduced 7 failures (4 S3, 3 Step Functions); reinstalling 9.38.0rc1 restored all 17 passes. Temporary buckets, state machines, and execution roles were deleted and their absence confirmed.

   The earlier 20 local checks using Moto/synthetic HTTP responses also passed (8 failures on 9.37.0), including explicit error: "", malformed success output, and service-error propagation. Those additional cases remain simulated. pip check passed, both merged commits are in the RC tag, and the relevant installed sources match the wheel and tag.

   Scope: the live checks make real AWS calls from the installed bundle/operator/hook. Operator execution is invoked directly, with UI extra-link persistence mocked; this does not cover scheduler/task-runner/XCom end-to-end behavior, an IAM-policy matrix, or the whole Amazon provider.

   👍React with 👍1Reacted by Jarek Potiuk
3. Andrushika commented on Oct 8, 2026

   

   Andrushika

   on Oct 8, 2026

   Contributor

   More actions

   #73652 tested in common.ai 0.11.0rc1, works as expected.
   Thanks for preparing the release, Jarek.

   👍React with 👍1Reacted by Jarek Potiuk

   ❤️React with ❤️1Reacted by Jarek Potiuk
4. ColtenOuO commented on Oct 8, 2026

   

   ColtenOuO

   on Oct 8, 2026

   Contributor

   More actions

   #73384 tested in common-ai 0.11.0rc1, works as expected too!

   Thanks for preparing the release, looking forward to the rapid development of common-ai ahead!

   👍React with 👍1Reacted by Jarek Potiuk
5. pankajkoti commented on Oct 8, 2026

   

   pankajkoti

   on Oct 8, 2026

   Member

   More actions

   #74017 tested that it's included in common-ai 0.11.0rc1 and works as expected.
   Thank you for preparing the release

   👍React with 👍1Reacted by Jarek Potiuk
6. zozo123 commented on Oct 8, 2026

   

   zozo123

   on Oct 8, 2026

   Contributor

   More actions

   #71676 and #74191 tested on rc1, works.

   • Add OpenSandbox backend for sandbox tools #71676 (common.ai 0.11.0rc1): I ran the PyPI wheel on Airflow 3.3.2 / Python 3.12 against a local OpenSandbox server. The PR's system-test Dag and its 64 unit tests pass, and command timeouts and sandbox cleanup behave as expected. One caveat, for a docs follow-up: when the server's egress sidecar runs in its default dns mode, a deny-all SandboxSpec() only blocks DNS, so direct-IP connections still get out. With [egress] mode = "dns+nft" they are blocked too, and everything above still passes. The backend can't tell the two modes apart on policy read-back.
   • Notify downstream Dags when COPY INTO writes a Unity (Databricks) table #74191 (databricks 7.22.0rc1): on a real SQL warehouse, COPY INTO emitted the canonical Unity table asset event and triggered the downstream Dag. Wrong table, wrong workspace and failed SQL emitted no event.

   ---

   Drafted-by: Claude Code (Opus 5.5) (no human review before posting)

   👍React with 👍1Reacted by Jarek Potiuk
7. divyanshus2404 commented on Oct 8, 2026

   

   divyanshus2404

   on Oct 8, 2026

   Contributor

   More actions

   #71711 tested in apache-spark 6.3.3rc1, works as expected.

   Verified against the RC in a clean venv: all masking forms produce the expected output (key=value, key value, quoted values spanning tokens, uppercase keys, tab- and multi-space-separated values, and a sensitive flag with no value), and a multiline value no longer swallows the log lines that follow it.

   The scanner that replaced the regex also closes the complexity gap raised in review on my PR — "secret" * 10000, which still backtracked quadratically in my version (~2.4s), now completes in 0.16ms. 50k single token: 0.66ms.

   Thanks for preparing the release.

   ---

   Drafted-by: Claude Code; reviewed by @divyanshus2404 before posting

   👍React with 👍1Reacted by Jarek Potiuk
8. aaron-y-chen commented on Oct 8, 2026

   

   aaron-y-chen

   on Oct 8, 2026

   Contributor

   More actions

   #73509 tested with apache-airflow-providers-amazon==9.38.0rc1 (Airflow 3.3.2, common-messaging 2.1.1), works as expected :)

   👍React with 👍1Reacted by Jarek Potiuk
9. subhramit commented on Oct 8, 2026

   

   subhramit

   on Oct 8, 2026

   Contributor

   More actions

   Have tested my changes

   • Add dedicated InfluxDB 3 sensor #73522 [influxdb: 2.13.0rc1]
   • Fix Elasticsearch and OpenSearch response wrapper bugs #73725 [elasticsearch: 6.9.2rc1] and [opensearch: 1.14.0rc1]

   using my agent, looks good:

   

   Thanks!

   👍React with 👍1Reacted by Jarek Potiuk
10. yuseok89 commented on Oct 9, 2026

    

    yuseok89

    on Oct 9, 2026

    Contributor

    More actions

    #72892 tested in amazon 9.38.0rc1 and #73880 tested in google 22.7.0rc1, both work as expected.

    👍React with 👍1Reacted by Jarek Potiuk
11. SEPURI-SAI-KRISHNA commented on Oct 9, 2026

    

    SEPURI-SAI-KRISHNA

    on Oct 9, 2026

    Last edited by SEPURI-SAI-KRISHNA

    Contributor

    More actions

    Tested both of the changes I am listed for. Both work as expected.

    Environment: clean virtualenv on Python 3.11, installed from PyPI, apache-airflow 3.3.2. No dbt Cloud account or GCP project involved, since everything below is observable before the first API call. In each case I also checked the previous stable release to confirm the result depends on the change rather than passing either way.

    dbt.cloud 4.11.0rc1, Apply hook_params to deferred dbt Cloud job runs (#74199)

    Walking the installed wheel, the DbtCloudRunJobTrigger construction in operators/dbt.py now passes hook_params; on 4.10.0 it does not. In both releases the trigger already accepted it, kept it across serialize() and built its hook with **self.hook_params, so the operator's call site was the only gap and nothing else had to change.

    	4.10.0	4.11.0rc1
    hook_params passed at the defer site	absent	present

    Checked at runtime with hook_params={"retry_limit": 7, "retry_delay": 3.5}: the operator retains them, its own hook honours them, and a trigger built with them serializes them intact.

    google 22.7.0rc1, Apply impersonation_chain to deferred BigQuery existence checks (#71648)

    Both deferral sites in sensors/bigquery.py now pass impersonation_chain, one for BigQueryTableExistenceTrigger and one for BigQueryTablePartitionExistenceTrigger; on 22.6.0 neither does. As above, the triggers accepted and serialized the argument in both releases, so only the sensors needed the change.

    	22.6.0	22.7.0rc1
    BigQueryTableExistenceTrigger	absent	present
    BigQueryTablePartitionExistenceTrigger	absent	present

    Confirmed the trigger accepts impersonation_chain and serializes it as given.

    Import smoke check

    Importing every module of both providers: all 12 dbt Cloud modules import, and 297 of 333 google modules. The 36 that do not are all missing optional peers in this environment, other providers plus paramiko, kubernetes, apache-beam, cassandra, oracledb and similar, each failing with an explicit message naming what is absent. Nothing failed for a reason internal to either release.

    👀React with 👀1Reacted by Jarek Potiuk
12. FrankYang0529 commented on Oct 9, 2026

    

    FrankYang0529

    on Oct 9, 2026

    Member

    More actions

    Tested apache.kafka 2.1.0rc1 and it works as expected:

    • Add KafkaSharedStreamProducer and KafkaSharedStreamTrigger #68625: I ran two AssetWatchers with KafkaSharedStreamTrigger on the same topic. The two triggers share one Kafka consumer. Both assets get every message. Offsets are committed per partition only after the asset events are stored. After a restart, committed messages are not delivered again. A connection without enable.auto.commit=false is refused. On Airflow 3.1.8 and 2.11.2, importing the shared-stream triggers raises AirflowOptionalProviderFeatureException.

    I found some issues for following providers:

    • google 22.7.0rc1
      • Add GKEPodExecOperator for existing Pods #72577: operators/kubernetes_engine.py now imports KubernetesPodExecOperator at module level. KubernetesPodExecOperator only exists in cncf-kubernetes 10.22.0 and later (https://airflow.apache.org/docs/apache-airflow-providers-cncf-kubernetes/10.22.0/changelog.html). With an older cncf-kubernetes, importing operators/kubernetes_engine.py raises AirflowOptionalProviderFeatureException. So every GKE operator fails to import, not only GKEPodExecOperator. Should we update the GKE module like the EKS module? In Add EKS operator for commands in existing Pods #72542, if eks.py cannot import KubernetesPodExecOperator, it defines a fallback class with the same name. The __init__ of that fallback class raises the same exception that the GKE module raises on import. The other EKS operators still import, and only creating EksPodExecOperator fails. I can create a follow-up PR for this.
    • dbt.cloud 4.11.0rc1
      • Apply hook_params to deferred dbt Cloud job runs #74199: A DbtCloudRunJobOperator with deferrable=True that works on 4.10.0 fails on 4.11.0rc1 when hook_params holds retry_args with tenacity objects. Created a follow-up PR for it Fix deferred dbt Cloud tasks failing with tenacity retry_args #74489.

    👀React with 👀2Reacted by Jarek Potiuk and Alejandro Morgante
13. SEPURI-SAI-KRISHNA commented on Oct 9, 2026

    

    SEPURI-SAI-KRISHNA

    on Oct 9, 2026

    Contributor

    More actions

    Correcting my earlier comment: my verification of #74199 in dbt.cloud 4.11.0rc1 was wrong.

    I checked that hook_params survives serialize(), but I used {"retry_limit": 7, "retry_delay": 3.5}, two JSON serializable scalars. hook_params is a free form dict and the one key whose documented values are Python objects is retry_args, so my input could not exercise the case that matters. With a tenacity object in retry_args the task fails when it defers, and it fails after the dbt Cloud job has already started, which leaves the run orphaned.

    Thanks to @FrankYang0529 for catching it and for the fix in #74489.

    Treating that as a release blocker for dbt.cloud 4.11.0: shipping rc1 as is would regress any deferrable DbtCloudRunJobOperator whose hook_params carries retry_args, which worked on 4.10.0. The google 22.7.0rc1 item I reported is unaffected and my result there stands.

    👍React with 👍1Reacted by PoAn Yang

    👀React with 👀1Reacted by Jarek Potiuk
14. r3wretrhy commented on Oct 9, 2026

    

    r3wretrhy

    on Oct 9, 2026

    Contributor

    More actions

    Checked apache-airflow-providers-apache-beam 6.3.1rc1: #72510 is included and the async hook now launches pipelines via create_subprocess_exec without a shell. Works as expected.

    👍React with 👍1Reacted by Jarek Potiuk
15. 6 remaining items

    Load more

16. AlejandroMorgante commented on Oct 9, 2026

    

    AlejandroMorgante

    on Oct 9, 2026

    Contributor

    More actions

    Validated the following changes and the follow-up requested for this provider release candidate:

    • apache-airflow-providers-google==22.7.0rc1 — Add GKEPodExecOperator for existing Pods #72577: the earlier example_kubernetes_engine.py run passed against real GKE Autopilot (1 passed in 1099.27s), with Kubernetes provider 10.24.0rc1; command execution in an existing Pod, XCom output, and Pod/cluster cleanup were verified.
    • Fix GKE operator imports with older Kubernetes providers #74505 is now merged. A fresh comparison with Kubernetes provider 10.21.0 passed (2 passed in 4.34s): the published Google RC reproduces the module import failure, while the exact merged operator source keeps existing cluster/Pod operators usable and raises the actionable dependency error only when constructing GKEPodExecOperator, preserving the original ImportError as its cause.

    Environment: Breeze, Airflow 3.3.0, Python 3.11.16; Google RC installed from the PyPI wheel. The compatibility fix was checked separately using the merged source; it is not included in 22.7.0rc1 and needs validation in a new RC.

    👍React with 👍1Reacted by Jarek Potiuk
17. abhinav-phi commented on Oct 10, 2026

    

    abhinav-phi

    on Oct 10, 2026

    Contributor

    More actions

    tested #73109 in apache-airflow-providers-amazon 9.38.0rc1, works as expected. environment: clean virtualenv on python 3.11 with apache-airflow 3.3.2, installed from pypi. driving the eks delete trigger against a stubbed client yields the status deleted event, and execute_complete now logs the success line for that and for a success event, while an error event logs nothing. the same run on 9.37.0 misses the success event, so that part is this change doing its job.

    👍React with 👍1Reacted by Jarek Potiuk
18. FrankYang0529 commented on Oct 10, 2026

    

    FrankYang0529

    on Oct 10, 2026

    Member

    More actions

    I found two more issues for following providers:

    • common.ai 0.11.0rc1
      • Rework common.ai Strands support into a plugin and add Google ADK #73898: It added self._create_lock = threading.Lock() to SandboxToolset.__init__. A lock cannot be deep-copied, and Airflow deep-copies tasks and default_args in several places. For example, TaskGroup deep-copies its default_args when it is created. On Airflow 3.3.2, a Dag file like this parses with 0.10.0 but fails with 0.11.0rc1. I create Let SandboxToolset be deep-copied again #74525 for this.
    • amazon 9.38.0rc1
      • Bind the AWS auth manager SAML response to the browser that started the login #73698: The new _awsam_login_state cookie is set with SameSite=Lax. After sign-in, Identity Center sends the SAML response back to /auth/login_callback with an auto-submitted form POST from its own site (the access portal on *.awsapps.com). Browsers do not send SameSite=Lax cookies on cross-site POST requests, so the callback returns 401 "No login in progress for this browser. Start the login from Airflow and try again." I create Fix AWS auth manager login when the IdP is on another site #74529 for this.

    ❤️React with ❤️1Reacted by Jarek Potiuk
19. potiuk commented on Oct 10, 2026

    

    potiuk

    on Oct 10, 2026

    MemberAuthor

    More actions

    Thanks everyone for testing. I ticked #73690, #73109, #74305, #72953 and #73552.

    Four providers are excluded from this wave because of regressions:

    • google 22.7.0rc1: Add GKEPodExecOperator for existing Pods #72577 breaks GKE operator imports with older cncf-kubernetes. Fixed by Fix GKE operator imports with older Kubernetes providers #74505 (merged).
    • dbt.cloud 4.11.0rc1: deferred tasks fail with tenacity retry_args. Fix in Fix deferred dbt Cloud tasks failing with tenacity retry_args #74489.
    • common.ai 0.11.0rc1: Rework common.ai Strands support into a plugin and add Google ADK #73898 made SandboxToolset impossible to deep-copy, so Dags that use it in default_args or a TaskGroup no longer parse. Fix in Let SandboxToolset be deep-copied again #74525.
    • amazon 9.38.0rc1: my Bind the AWS auth manager SAML response to the browser that started the login #73698 sets the login-state cookie with SameSite=Lax. The cross-site SAML POST from IAM Identity Center arrives without it, so every login started from Airflow fails with 401. Fix in Fix AWS auth manager login when the IdP is on another site #74529.

    All four will be re-cut in an ad-hoc wave once the fixes are merged. Results already ticked here for unaffected PRs carry over. Thanks @FrankYang0529 for catching the last two and fixing them, and @AlejandroMorgante for #74505.

    The other providers in this wave stay in the vote.

    ---

    Drafted-by: Claude Code (Opus 5.5); reviewed by @potiuk before posting

    👍React with 👍1Reacted by PoAn Yang
20. SameerMesiah97 commented on Oct 10, 2026

    

    SameerMesiah97

    on Oct 10, 2026

    Contributor

    More actions

    apache-airflow-providers-amazon==9.38.0rc1

    PR #73870: Cannot be tested deterministically. Unit tests are sufficient.

    apache-airflow-providers-snowflake==6.19.0rc1

    PR #70703: Changes are working as expected with no observed regressions.
    PR #73815: Changes are working as expected with no observed regressions.
21. FrankYang0529 commented on Oct 10, 2026

    

    FrankYang0529

    on Oct 10, 2026

    Member

    More actions

    common.ai 0.11.0rc1 has one more regression.

    • Return an approved agent answer as output_type after human review #73904: With enable_hitl_review=True, an AgentOperator task now converts the approved answer with a pydantic TypeAdapter built from output_type. Following cases have different behavior from previous version. I create Fix agent HITL review for list, marker and function output types #74537 to fix it.
      • output_type=[Verdict, Other] (a list of output types): on 0.10.0 the task returns a dict. On 0.11.0rc1 the task fails after the approval with PydanticSchemaGenerationError.
      • output_type=ToolOutput(Verdict) (an output marker): on 0.10.0 the task returns a dict. On 0.11.0rc1 the task fails after the approval with AttributeError.
      • An output function: on 0.10.0 the function runs once. On 0.11.0rc1 it runs again after the approval, so any side effect in the function happens twice.
22. keemgdeok commented on Oct 11, 2026

    

    keemgdeok

    on Oct 11, 2026

    Contributor

    More actions

    Verified #74143 in apache-airflow-providers-cncf-kubernetes==10.24.0rc1 from PyPI, with Airflow 3.3.2 and Python 3.12.14.

    Checked job_poll_interval=0.5, 60, and the default 10 through operator deferral, trigger serialization/reconstruction, and the hook's polling loop, for both successful and failed jobs: 6 passed.

    With provider 10.23.0, the four non-default cases reproduced the bug: the hook requested 10-second sleeps instead. Only the provider version changed; reinstalling the RC restored all six passes. pip check passed in each installation.

    The merge commit is included in the RC tag, and the installed operator, trigger, and hook sources match the PyPI wheel and tag.

    Scope: Kubernetes status responses and asyncio.sleep were mocked. This checks the installed polling path, without a live cluster or scheduler/triggerer deployment.

    ---

    Drafted-by: Codex (GPT-6); reviewed by @keemgdeok before posting
23. filipeaaoliveira commented on Oct 11, 2026

    

    filipeaaoliveira

    on Oct 11, 2026

    Contributor

    More actions

    Verified #71092 in apache-airflow-providers-airbyte==6.2.0rc1 (PyPI) with Airflow 3.3.2, Python 3.12, against a real self-hosted Airbyte OSS 2.4.0 (abctl, OAuth client credentials), Faker → /dev/null connection, via AirbyteTriggerSyncOperator / airflow dags test.

    To simulate a loaded server, a local proxy delayed POST /jobs (and in one case the token request) by 8s:

    • no timeout configured → ReadTimeout at the 5s default, so default behavior is unchanged
    • timeout connection extra (30) → syncs triggered and completed successfully, including when the OAuth token request was also slow
    • AirbyteHook(timeout=...) works and overrides the extra; lower values (2s) are honored
    • invalid extras ("abc", 0, -5) → clear ValueError

    With provider 6.1.0, the same timeout extra is ignored and the call fails at 5s while Airbyte still creates the job, which is the issue this fixes. Also checked directly against Airbyte without the proxy: a full sync succeeded.
24. molcay commented on Oct 12, 2026

    

    molcay

    on Oct 12, 2026

    Last edited by molcay

    Contributor

    More actions

    Hi @potiuk,

    I know we're already on day 4 of the RC and this is a tough time to hit the brakes, but we've unfortunately found a regression in the Google provider. We just need a bit more time to verify if it's tied to the RC changes.

    I'll follow up here by the end of the day with an update. Thanks in advance!

    UPDATE:

    We understand that the regression is not related to the RC changes. We can proceed with the RC
25. yamakazushi commented on Oct 12, 2026

    

    yamakazushi

    on Oct 12, 2026

    Contributor

    More actions

    #73408 works as expected on the amazon 9.38.0rc1 release candidate (Airflow 3.3.2, botocore 1.43.111, stubbed against the real mwaa-serverless service model, no live AWS): the new create_workflow_kwargs / update_workflow_kwargs carry all six previously unreachable CreateWorkflow inputs and all four for UpdateWorkflow onto the request, where 9.37.0 rejected the parameter outright and left NetworkConfiguration unreachable. Also checked Jinja rendering into nested dicts, a Dag serialization round-trip, and the 22 upstream unit tests at the RC tag.
26. fpiped commented on Oct 12, 2026

    

    fpiped

    on Oct 12, 2026

    Contributor

    More actions

    Verified #73211 in apache-airflow-providers-amazon==9.38.0rc1 (PyPI) with Airflow 3.3.2 and Python 3.12:

    • The AwsTaskLogFetcher in the RC matches the merged change: run() waits on the stop event and reads the log stream once more after stop().
    • The fetcher's unit tests from main, run against the installed RC: 33 passed. On 9.37.0 the two new regression tests fail, as expected.
    • A real ECS Fargate task through EcsRunTaskOperator (awslogs_fetch_interval 5 s), whose container writes seven lines, the last one just before exiting: every line reached the task log exactly once. The race itself depends on when stop() lands, so it is covered by the regression test, not by this run.
27. seanghaeli commented on Oct 12, 2026

    

    seanghaeli

    on Oct 12, 2026

    Contributor

    More actions

    Verified #73930 and #73014 in apache-airflow-providers-amazon==9.38.0rc1
28. potiuk commented on Oct 13, 2026

    

    potiuk

    on Oct 13, 2026

    MemberAuthor

    More actions

    Thank you everyone. Providers are released.

    I invite everyone to help improve providers for the next release, a list of open issues can be found here.

    🎉React with 🎉1Reacted by Subhramit Basu
29. potiuk

    closed this as completedon Oct 13, 2026
30. shivaam commented on Oct 14, 2026

    

    shivaam

    on Oct 14, 2026

    Contributor

    More actions

    Verified the reconnect fix in #74416 against apache-airflow-providers-smtp==3.1.1rc1 installed from PyPI on Python 3.11.16 in an isolated Breeze container. The imported SMTP hook was confirmed to come from the installed RC package, not the source checkout.

    • Reusing the same synchronous and asynchronous hook across two context-manager entries successfully sent both emails.
    • Calling the same notifier twice successfully reused its cached hook with a fresh SMTP client.
    • Client state was cleared even when synchronous close() or asynchronous quit() raised; the cleanup errors still propagated.
    • Local Mailpit captured all six expected messages. SMTP provider 3.1.0 reproduced the second-send failures as a negative control.

    The RC tag contains the merged fix and its changelog entry. This is focused verification of the reconnect fix, not a full release-artifact audit or supported-core-version matrix; the core/common dependencies used Breeze's development baseline.

    ---

    Drafted-by: Codex (no human review of this comment text before posting)