Question

During a routine rolling deploy of your translation inference service, the error rate spikes to 8% for ~6 minutes each time a batch of pods cycles, then recovers. Symptom repeats on every rollout. Dashboards: newly started pods report Ready and receive traffic immediately, but their first ~90 seconds show very high latency and timeouts; the model artifact is a multi-GB file pulled from object storage at startup, and the first inference triggers lazy GPU kernel/weight loading. p99 on warm pods is fine. Triage and fix.