A trace shows most request latency inside one downstream service call, while local handler computation is brief. Which investigation follows the evidence?

1. Inspect the slow downstream call, its retries, resource limits, and request parameters before optimizing unrelated local code.
2. Disable tracing permanently because it revealed a slow dependency.
3. Rewrite only the local response JSON serializer despite its negligible measured duration.
4. Assume every latency issue is a cold start regardless of the trace.