KubernetesExecutor: Event log says "Pod failed because of None" for OOMKilled pods #73164

Under which category would you file this issue?

Providers

Apache Airflow version

3.3.1

What happened and how to reproduce it?

When a KubernetesExecutor worker pod is OOM-killed, the executor collects the failure details correctly. The scheduler log shows:

Task dc_cls__...process_area_WE_Golden_Ocean.5 failed in pod airflow-data-cube-cls/dc-cls-...-bvma68of.
Pod phase: Failed, reason: None, message: None, container_type: main, container_name: base,
container_state: terminated, container_reason: OOMKilled, container_message: None, exit_code: 137

But the termination_reason that is passed to the scheduler via event_buffer (and ends up in the task's Event log as Extra info) is built only from pod.status.reason:

# kubernetes_executor.py, _change_state()
termination_reason = f"Pod failed because of {pod_reason}"

Kubernetes leaves pod.status.reason empty for container-level failures such as OOMKilled or a non-zero exit, so the Event log row (event="state mismatch") reads:

... Learn more: https://airflow.apache.org/docs/apache-airflow/stable/troubleshooting.html#task-state-changed-externally Extra info: Pod failed because of None

The task log is empty too (the process was SIGKILLed), so from the UI / REST API an OOM kill is indistinguishable from any other external kill. The only place the real reason exists is the scheduler log.

Introduced by #60800, whose stated goal was exactly to surface OOM / Evicted in the Event log. The same line is still in main and in apache-airflow-providers-cncf-kubernetes==10.22.0.

Reproduce: run any task under KubernetesExecutor with a memory limit lower than the task needs (e.g. executor_config with limits.memory: 64Mi and a task that allocates 200 MiB), wait for the failure, open the task's Event log tab.

What you think should happen instead?

Fall back to the container-level reason when the pod-level one is empty, and include the container name / exit code, e.g.

Extra info: Pod failed because of OOMKilled (container: base, exit code: 137)

Operating System

Debian 12 (official image base), Kubernetes v1.35.5

Deployment

Other 3rd-party Helm chart

Apache Airflow Provider(s)

cncf-kubernetes

Versions of Apache Airflow Providers

apache-airflow-providers-cncf-kubernetes==10.21.1 (same code in 10.22.0 and main)

Kubernetes Version

v1.35.5

Anything else?

Related: #66405 (AIP-97, typed failure causes) is the broader effort; this issue is only about the string that already exists being built from the wrong field.

Are you willing to submit PR?

• Yes I am willing to submit a PR!

Code of Conduct

• I agree to follow this project's Code of Conduct