Pods OOMKilled in rolling wave

A new version of the feed service rolls out fully at 10:00. No immediate problem. Starting ~10:40 and worsening over the next hour, pods begin getting OOMKilled and entering CrashLoopBackOff in a slow rolling wave; capacity drops, latency climbs, and the autoscaler thrashes. Dashboards: per-pod RSS climbs roughly linearly from pod start until it hits the memory limit, then OOMKill, then the cycle repeats on the fresh pod; GC pause time is flat; request rate is normal. The new version added a per-request in-memory dedup set keyed by item ID, intended to be request-scoped, but it was attached to a module-level singleton. How do you triage and mitigate?