Log enough to debug issues:
- Request and response content (anonymized if needed)
- Latency breakdown
- Model version and configuration
- Error details with stack traces
Use structured logging for easy querying. Connect logs to your monitoring system for correlation.