A high-traffic checkout service on GKE handles roughly 50,000 requests per second. Your team enabled distributed tracing with the OpenTelemetry SDK exporting to Cloud Trace, but engineers report that the tracing overhead has increased request latency and Cloud Trace ingestion costs have grown dramatically. You need to retain the ability to diagnose latency issues across representative requests while significantly reducing overhead and cost. What is the most appropriate action?