A developer wants MLflow traces to distinguish retrieval delay from generation delay. How should the workflow be instrumented?

1. Merge all requests into one span without request identifiers.
2. Log only one final success boolean for the whole request.
3. Record only the process startup time once per deployment.
4. Record meaningful child spans around retrieval and model generation within the request trace.