GKE deployment pods restart every 2 seconds; app logs to stdout. To diagnose the failing component, which approach should you take?

The most efficient approach is to inspect Stackdriver logs for the specific GKE container. Since the application logs to stdout, these logs are automatically captured by Stackdriver (now Cloud Logging) and associated with the specific container and pod. This provides direct insight into the application's behavior and error messages leading to the restarts. Inspecting logs for each Compute Engine cluster node is too broad and unlikely to pinpoint the application-level issue. Executing into a container to read logs is less efficient and scalable than using Cloud Logging, which aggregates and centralizes logs. Serial port logs are for VM-level issues, not application-level container problems.