How do you handle graceful degradation in distributed Python services?

When one downstream dependency degrades (high latency, partial outages), our service tends to cascade rather than degrade gracefully. We've tried timeouts and circuit breakers (pybreaker), but the real problem is backpressure — the calling service keeps retrying aggressively. How does your team structure the fallback path? Do you prefer: - Stale cache returns with a warning header - Reduced feature mode (skip non-critical enrichment steps) - Queue-based async processing for degraded paths Stack: Python 3.12, FastAPI, Redis, PostgreSQL. Would love to hear war stories.