Add failure causes and infrastructure retry budgets

[1fanwang]

Airflow can record that a task failed without knowing that the platform removed its worker. The failure then looks like task code and consumes the task's normal retries.

This adds a typed cause contract to workers, the API, and BaseExecutor. Confirmed infrastructure failures can use a separate infra_retries budget; the default is 0, so upgrades do not add attempts. AIP-103 stores the count across scheduler restarts and task clears without a new migration. Ambiguous signals stay unclassified and keep today's behavior.

AIP: https://cwiki.apache.org/confluence/x/0ITMFw

Executor proofs: 1fanwang#28 (ECS Spot), 1fanwang#15 (Celery reason-only). Stateless fallback: 1fanwang#16.

End-to-end evidence

An external kill of a retries=0 task, driven through the real executor and scheduler path:

Cause reported by the executor	Task outcome
none, which is what every executor does today	FAILED
infra with reason Evicted	UP_FOR_RETRY, infrastructure retry 1/1
application with reason OOMKilled	FAILED
reason WorkerLost with no kind	FAILED

Commands and raw output

$ uv run --project airflow-core pytest airflow-core/tests/integration/executors/test_failure_cause.py -q -s... failure_kind=infra failure_reason=Evicted... Using infrastructure retry 1/1 ... reason=Evicted, max_tries=1... Marking task as UP_FOR_RETRY. dag_id=failure_cause_infra... failure_kind=application failure_reason=OOMKilled... Marking task as FAILED. dag_id=failure_cause_application... failure_kind=None failure_reason=WorkerLost... Marking task as FAILED. dag_id=failure_cause_reason_only... failure_kind=None failure_reason=None... Marking task as FAILED. dag_id=failure_cause_unclassified4 passed
$ kind create cluster --name aip97 --config kind-aip97.yaml # 1 control plane, 2 workers
$ kubectl run drain-victim --image=busybox:1.36 \ --overrides='{"spec":{"nodeName":"aip97-worker","terminationGracePeriodSeconds":5}}' \ -- sh -c 'trap "exit 143" TERM; while true; do sleep 1; done'
$ kubectl wait --for=condition=Ready pod/drain-victim --timeout=120s
$ kubectl taint node aip97-worker node.kubernetes.io/unschedulable=aip97:NoExecute
$ kubectl get pod drain-victim -o json | jq -c '[.status.phase, (.status.reason // "-"), ([.status.conditions[]?|select(.type=="DisruptionTarget" and .status=="True")|.reason]|first)]'["Failed","-","DeletionByTaintManager"]
$ uv run --project providers/cncf/kubernetes python -c 'from airflow.providers.cncf.kubernetes.executors.kubernetes_executor_utils import classify_pod_failure as c; print(c(dict(container_reason="Error", disruption_reason="DeletionByTaintManager")))'(<TaskFailureKind.INFRA: 'infra'>, 'DeletionByTaintManager')

---

Was generative AI tooling used to co-author this PR?

• Yes: GitHub Copilot CLI 1.0.80 (GPT-5.6 Sol)

Generated-by: GitHub Copilot CLI 1.0.80 (GPT-5.6 Sol) following the guidelines

---