Question

A read-heavy catalog service on Kubernetes (limit 2Gi) starts getting OOMKilled (exit 137) a few times a day, always 6–10 hours after a pod starts, never right away. The memory graph climbs a slow staircase and never plateaus, but request rate and latency are normal. Heap/RSS grows roughly in proportion to the number of *distinct* product IDs seen since boot, not to current QPS. A release last week added an in-process `map[string]Product` cache 'to cut DB load' — it had a great hit rate in the demo. There is no eviction configured. Triage and mitigate.