OOMKilled (exit code 137) occurs when the Linux kernel's OOM killer terminates a container for exceeding its cgroup memory limit — Kubernetes only reports it, not enforces it. Three distinct scenarios produce OOMKilled: container limit breach, kubelet eviction under node memory pressure, and node-level kernel OOM. Diagnosis involves kubectl describe, previous container logs, and Prometheus queries using container_memory_working_set_bytes (not container_memory_usage_bytes, which includes reclaimable page cache). The correct fix is per-container rightsizing: set requests at p95 of working set over 7 days, limits at p99 or 1.5–2x requests. QoS class affects kill priority — Guaranteed pods are last to be killed. JVM containers need -Xmx capped at 75% of the container limit to account for off-heap allocation. At scale, automated rightsizing tools that track PSI metrics can proactively adjust limits before OOM kills occur.
